chore: release v0.9.0

Merge feat/roadmap-2026-06-19: pre-release review fixes (2026-06-23)
fix: address pre-release review findings (2026-06-23)
2026-06-23 14:48:37 +03:00 · 2026-06-23 14:21:42 +03:00 · 2026-06-23 14:21:25 +03:00 · 2026-06-23 00:55:49 +03:00 · 2026-06-23 00:50:22 +03:00 · 2026-06-22 23:21:24 +03:00
214 changed files with 39021 additions and 9282 deletions
@@ -354,6 +354,58 @@ jobs:
            docker push "$REGISTRY:latest"
          fi

+      # Best-effort arm64 (Raspberry Pi / arm64 HAOS hosts). Runs AFTER the
+      # amd64 push so the amd64 image always ships even if this fails.
+      # Deliberately avoids `docker buildx` (its docker-container driver needs
+      # nested networking the TrueNAS runners lack — see contexts/ci-cd.md):
+      # instead it cross-builds a single arm64 image via QEMU binfmt and folds
+      # amd64 + arm64 into multi-arch manifest lists under the existing tags.
+      # `continue-on-error` keeps a runner that can't emulate arm64 from
+      # failing the release; the plain amd64 tags pushed above remain valid.
+      - name: Build + publish arm64 (multi-arch manifest, best-effort)
+        if: github.event_name == 'push' && steps.docker-login.outcome == 'success'
+        continue-on-error: true
+        run: |
+          set -e
+          export DOCKER_CLI_EXPERIMENTAL=enabled
+          TAG="${{ gitea.ref_name }}"
+          REGISTRY="${{ steps.meta.outputs.registry }}"
+          VERSION="${{ steps.meta.outputs.version }}"
+
+          # Register arm64 emulation. If the runner forbids privileged
+          # containers this fails and the whole step is skipped.
+          docker run --privileged --rm tonistiigi/binfmt --install arm64
+
+          # Cross-build the arm64 image (QEMU-emulated — slow but uses arm64
+          # manylinux wheels, so no source compilation). Stays in the local
+          # daemon alongside the amd64 image from the previous build step.
+          DOCKER_BUILDKIT=1 docker build \
+            --platform linux/arm64 \
+            --build-arg APP_VERSION="$VERSION" \
+            --label "org.opencontainers.image.version=$VERSION" \
+            --label "org.opencontainers.image.revision=${{ gitea.sha }}" \
+            -t "$REGISTRY:$VERSION-arm64" \
+            ./server
+
+          # Fold amd64 + arm64 into a multi-arch manifest list under each
+          # user-facing tag. The arch-suffixed tags remain pullable directly.
+          publish_manifest() {
+            local t="$1"
+            docker tag "$REGISTRY:$t" "$REGISTRY:$t-amd64"
+            docker push "$REGISTRY:$t-amd64"
+            docker tag "$REGISTRY:$VERSION-arm64" "$REGISTRY:$t-arm64"
+            docker push "$REGISTRY:$t-arm64"
+            docker manifest create --amend "$REGISTRY:$t" \
+              "$REGISTRY:$t-amd64" "$REGISTRY:$t-arm64"
+            docker manifest push "$REGISTRY:$t"
+          }
+
+          publish_manifest "$TAG"
+          publish_manifest "$VERSION"
+          if ! echo "$TAG" | grep -qE '(alpha|beta|rc)'; then
+            publish_manifest "latest"
+          fi
+
  # ── Publish the release (flip draft=false) ─────────────────
  # Runs only after every build job succeeded so users never see a
  # release that's missing artifacts or sha256 sidecars (the in-app
@@ -19,6 +19,7 @@ semantic = true
 # Automatically run `vex update` before search if the index is stale
 auto_update = true

-# Embedder used for semantic indexing. Known IDs: minilm-l6-v2 (default).
+# Embedder used for semantic indexing. IDs: minilm-l6-v2 (default, CPU-fast),
+# jina-code (code-specialized, GPU-worthy), bge-base-en-v1.5, bge-large-en-v1.5.
 # Changing the embedder requires a full reindex.
-# embedder = "minilm-l6-v2"
+embedder = "jina-code"
@@ -2,9 +2,40 @@

 ## Code Search

-**If `ast-index` is available, use it as the PRIMARY code search tool.** It is significantly faster than grep and returns structured, accurate results. Fall back to grep/Glob only when ast-index is not installed, returns empty results, or when searching regex patterns/string literals/comments.
+**Priority order: `vex` (PRIMARY) → `ast-index` (fallback) → Grep/Glob (last resort).** This repo has a fully-featured `.vex.toml` index. Use vex first for any symbol/definition/usage/call-graph lookup. Fall back to ast-index only when vex legitimately can't help, and to Grep/Glob only for regex patterns, string literals, comments, config files, or unparsed languages.

-**IMPORTANT for subagents:** When spawning Agent subagents (Plan, Explore, general-purpose, etc.), always instruct them to use `ast-index` via Bash for code search instead of grep/Glob. Example: include "Use `ast-index search`, `ast-index class`, `ast-index usages` etc. via Bash for code search" in the agent prompt.
+**IMPORTANT — use ALL vex indexing features.** The index is built with every capability enabled, and queries must take advantage of them. Keep them ON and exploit them:
+
+| Capability | Status | Powers |
+| ---------- | ------ | ------ |
+| Semantic embeddings (`jina-code`, 768-dim) | ON | `vex search` (semantic channel), `similar`, `find_similar`, `duplicates` |
+| Call graph | ON | `vex callers`, `callees`, `paths`, `reachable`, `bundle --mode pr-impact` |
+| BM25 | ON | hybrid RRF text channel in `vex search` |
+| Pattern index | ON | `vex pattern` AST-shape matching |
+| C++ includes | ON | include-graph resolution |
+| Body tokens (incremental HNSW) | ON | fast incremental reindex |
+| History | ON | `vex history`, `vex diff <rev>` blame/evolution queries |
+
+**In-session, use the `mcp__vex__*` MCP tools** (`search`, `show`, `usages`, `callers`, `callees`, `bundle`, `outline`, `implementations`, `similar`, `grep`, `status`, etc.) — MCP output is far cheaper in tokens than `Bash("vex …")`. Drop to Bash `vex` only for CLI-only features (`pattern`, `diff`, `paths`, `reachable`, `bundle`, `history`, `--strict`/`--why` flags), for subagent prompts, or for shell composition.
+
+```bash
+vex search "query" --semantic             # Hybrid semantic + BM25 search
+vex show <Symbol>                          # Definition body (prefer over Read)
+vex usages <Symbol> --strict               # Reference sites (AST-precise on T1 langs)
+vex callers <Function>                     # Call sites (function-scoped)
+vex callees <Function>                     # Outgoing calls
+vex paths --from <A> --to <B>              # Multi-hop call-graph path
+vex bundle --mode pr-impact --base master  # Changed symbols + callers + reachable tests
+vex pattern '$X async fn returning Response'  # AST-shape (metavariables)
+vex diff master                            # Symbol-level branch diff
+vex history <Symbol>                        # Commit evolution of a symbol
+```
+
+**Maintenance:** the index has `auto_update = true`, so it refreshes on stale queries. After a `vex self-update`, rerun `vex index --history --semantic --embedder jina-code --device cuda` so newly-added extractors populate and all features stay enabled. Verify with `vex status` — every capability line should read `yes`.
+
+**IMPORTANT for subagents:** Subagents don't inherit MCP. When spawning Agent subagents (Plan, Explore, general-purpose, etc.), instruct them to use `vex` via Bash for code search (e.g. include "Use `vex search`, `vex show`, `vex usages`, `vex callers` via Bash for code search; ast-index is the fallback"). Don't tell them to default to grep/Glob.
+
+**Fallback — `ast-index`** (use only when vex is unavailable):

 ```bash
 ast-index search "Query"                  # Universal search
@@ -82,6 +82,19 @@ LedGrab speaks many protocols, so a single setup can drive everything from a DIY
 - Real-time FPS, latency, and uptime charts
 - Localized in English, Russian, and Chinese

+### Activity Log
+
+The **Activity** tab is a persistent, queryable audit log of everything LedGrab has done — entity changes, auth events, device connections, and system actions.
+
+- Filter by category (auth, device, entity, capture, system), severity, actor, entity type, date range, or free text
+- Live-append of new events as they happen
+- Export as CSV or JSON (authentication required)
+- Entity crosslinks navigate directly to the relevant card
+- **Retention settings** (Settings → Activity Log): configure max age, max entry count, and toggle recording on/off
+- **Clear log** (Settings → Activity Log, requires authentication) — audited: a system entry records who cleared the log and when
+
+> **Note:** The Activity Log is distinct from the **debug Log Viewer** (Settings → General → Open Log Viewer). The Log Viewer is an ephemeral real-time tail of the server's Python log stream (WARNING/ERROR lines, resets on disconnect). The Activity Log is a structured, persistent SQLite-backed record of semantic application events.
+
 ### Home Assistant Integration

 - HACS-compatible custom component (separate repository)
@@ -1,54 +1,91 @@
-## v0.8.1 (2026-05-28)
+## v0.9.0 (2026-06-23)

-### User-facing changes
+A large feature release: a full activity/audit log, two roadmap batches of
+capture and smart-light improvements, per-pixel control for LIFX/Hue/Nanoleaf,
+and new outbound integrations (webhooks + Home Assistant MQTT discovery).

-#### Features
+### Features

-##### Multi-broker MQTT devices
+#### Activity Log
+- Persistent activity/audit log: storage model with migration, recorder with
+  actor context and retention, and event instrumentation across four
+  categories ([1ac4a0f](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/1ac4a0f), [726f39e](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/726f39e), [25c613c](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/25c613c))
+- REST API for list / export / settings / clear ([4a09275](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/4a09275))
+- Activity tab with smart filtering, live updates, and export, plus a
+  dashboard widget and settings panel ([9a0137f](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/9a0137f), [6e1dd21](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6e1dd21))

- The device editor now shows an MQTT **broker picker** for `device_type=mqtt` (in both the add-device and device-settings modals), wired into load / save / validate / dirty-check / clone. An empty selection means "first available broker"
- `mqtt_source_id` is now threaded end-to-end through `DeviceCreate` / `DeviceUpdate` / `DeviceResponse` and the device routes; the referenced broker is validated on create **and** update ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
+#### Per-pixel smart lights
+- LIFX multizone (SetExtendedColorZones) and Tile per-pixel streaming,
+  auto-detected on connect with single-colour fallback ([39b0554](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/39b0554))
+- Philips Hue gradient-lightstrip mapping: Entertainment v2 frames keyed by
+  channel id, with a `hue_gradient_mode` toggle ([39b0554](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/39b0554))
+- Nanoleaf extControl v2 per-panel UDP streaming (`per_panel` mode) ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))

-##### Schema-driven wiring-graph editor
+#### Capture & effects
+- Linear-light blending and spatio-temporal dithering, opt-in per calibration ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))
+- Audio-reactive palette modulation across all 12 procedural effects ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))
+- Color-harmony gradient generator (complementary / analogous / triadic / …) ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))

- The visual graph editor now renders ports and edges generically from a backend-served schema (`GET /api/v1/graph/schema`) instead of hard-coding the connectable-field topology in two places — so client and server can no longer drift
- New `GET /api/v1/graph` returns the full nodes + edges + validation topology, and `GET /api/v1/graph/dependents/{kind}/{id}` reports what references an entity ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
+#### Automations & integrations
+- Solar sunrise/sunset automation trigger (new `utils/solar.py`) ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))
+- Outbound webhook automation action (Discord / IFTTT / Zapier / Node-RED),
+  SSRF-gated at save and fire time ([39b0554](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/39b0554))
+- Home Assistant MQTT auto-discovery: read-only binary sensors per automation,
+  availability via birth/will, with cleanup on disable/delete ([39b0554](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/39b0554))
+- League of Legends poller wired via a `LoLPollManager` + shared runtime state ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))
+- `auth.expose_docs` flag (default off) to view `/docs`, `/redoc`, and
+  `/openapi.json` without a token; all real endpoints stay protected ([126d8f2](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/126d8f2))

-##### Aggregated snapshot endpoint
-
- New `GET /api/v1/snapshot` returns all output targets (with processing state + metrics), devices (with brightness), the source / preset / clock lists, and the system block in a **single response** — collapsing the Home Assistant integration's previous ~2N+M request fan-out into one round trip
- `?include=` fetches only a subset of sections, and an excluded section also skips its server-side work (e.g. cold-cache hardware brightness probes or the blocking NVML performance query) ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
-
-#### Bug Fixes
-
- **Graceful shutdown no longer hangs:** uvicorn's graceful-shutdown wait is now bounded (`GRACEFUL_SHUTDOWN_TIMEOUT`, shared by the desktop, Android, and demo launchers). A lingering events WebSocket (which the browser auto-reconnects) used to keep connections from draining, so the lifespan shutdown never ran — leaving LED targets lit and blocking process exit. Ctrl+C / OS shutdown with the UI open now reliably stops targets and checkpoints the DB ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
- **Device update error codes:** `update_device` no longer masks an intentional 4xx (e.g. an unknown `mqtt_source_id` or failed group validation) as a generic 500 ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
+### Bug Fixes
+- Pre-release review hardening: solar timezone crash, webhook header CRLF,
+  MQTT topic-prefix injection, thread-safe `get_stats` copy, MQTT discovery
+  lock, `reactive_mode` Literal, and calibration-modal accessibility ([0c096db](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/0c096db))
+- Comprehensive review fixes across security, concurrency, performance,
+  Android, and UI ([17dd2e0](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/17dd2e0))
+- Activity Log polish: accessible export menu, i18n placeholders, dashboard
+  section reconciliation, column alignment, ticking time, and no spinner
+  flash on instant filtering ([3dd1ac3](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/3dd1ac3), [ff1ff06](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/ff1ff06), [77284e8](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/77284e8), [ae74cca](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/ae74cca), [077c99c](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/077c99c))

 ---

 ### Development / Internal

-#### Backend
+#### CI/Build
+- Best-effort arm64 multi-arch Docker manifest via QEMU + `docker manifest`
+  (amd64 path untouched) ([6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25))

- **Wiring-graph schema engine** (`api/graph_schema.py`): a pure, unit-tested module that is the single source of truth for which reference fields connect which entity kinds; builds the topology and performs dependency lookup plus cycle / dangling-reference detection without booting the app or any store. The route layer only gathers serialized entities and delegates ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
- **Structured access log:** a new middleware emits one structured line per request, attributing it to the authenticated token's friendly label (the key name, **never** the secret) so traffic can be traced to a client (e.g. `homeassistant` vs `android`). uvicorn's own access log is disabled to avoid duplicate lines ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
- Shared `validate_mqtt_source_exists` (`_mqtt_validation.py`) deduplicates the MQTT-source existence check between the device and output-target routes ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
+#### Chores
+- Activity Log feature plan/subplan scaffold, post-merge cleanup, and context
+  graduated into CLAUDE.md ([1afe7d6](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/1afe7d6), [e584235](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/e584235))

-#### Frontend
-
- Service-worker refresh for the new bundle ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
-
-#### Tests
-
- New suites: graph routes + schema engine, snapshot routes, access-log middleware, `mqtt_source_id` device regressions, and the bounded-shutdown entrypoint. Full suite: **1614 passing** ([a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba))
+> Tests: ~180 new unit tests added across the activity log, roadmap features,
+> and integrations. Release gate green: ruff + tsc + build clean,
+> **pytest 2739 passed / 2 skipped**.

 ---

 <details>
-<summary>All Commits (1)</summary>
+<summary>All Commits</summary>

 | Hash | Message | Author |
-| ---- | ------- | ------ |
-| [a5effba](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/a5effba) | feat: aggregated snapshot + wiring-graph APIs, MQTT device brokers | alexei.dolgolyov |
+|------|---------|--------|
+| [0c096db](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/0c096db) | fix: address pre-release review findings (2026-06-23) | alexei.dolgolyov |
+| [39b0554](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/39b0554) | feat: roadmap round two (2026-06-23) — per-pixel smart-lights + integrations | alexei.dolgolyov |
+| [6745e25](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6745e25) | feat: roadmap batch (2026-06-19) — solar/linear-light/dither/nanoleaf + integrations | alexei.dolgolyov |
+| [126d8f2](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/126d8f2) | feat(auth): add auth.expose_docs flag to view API docs without a token | alexei.dolgolyov |
+| [e584235](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/e584235) | chore(activity-log): post-merge cleanup + graduate context to CLAUDE.md | alexei.dolgolyov |
+| [077c99c](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/077c99c) | fix(activity-log): no spinner flash on instant filtering | alexei.dolgolyov |
+| [ae74cca](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/ae74cca) | fix(activity-log): UI polish - accessible export menu, i18n placeholders, zero-result spinner fix | alexei.dolgolyov |
+| [77284e8](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/77284e8) | fix(activity-log): dashboard section reconciliation + activity column alignment | alexei.dolgolyov |
+| [ff1ff06](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/ff1ff06) | fix(activity-log): post-test polish - localize descriptions, dashboard widget, ticking time | alexei.dolgolyov |
+| [3dd1ac3](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/3dd1ac3) | fix(activity-log): final-review fixes - crosslink keys + sanitize parity | alexei.dolgolyov |
+| [6e1dd21](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/6e1dd21) | feat(activity-log): phase 6 - dashboard widget + settings panel + docs | alexei.dolgolyov |
+| [9a0137f](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/9a0137f) | feat(activity-log): phase 5 - Activity tab (smart filtering, live updates, export) | alexei.dolgolyov |
+| [4a09275](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/4a09275) | feat(activity-log): phase 4 - REST API (list/export/settings/clear) | alexei.dolgolyov |
+| [25c613c](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/25c613c) | feat(activity-log): phase 3 - event instrumentation (4 categories) | alexei.dolgolyov |
+| [726f39e](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/726f39e) | feat(activity-log): phase 2 - recorder, actor context, retention, lifecycle | alexei.dolgolyov |
+| [1ac4a0f](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/1ac4a0f) | feat(activity-log): phase 1 - storage model, migration, repository | alexei.dolgolyov |
+| [1afe7d6](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/1afe7d6) | chore(activity-log): scaffold feature plan and phase subplans | alexei.dolgolyov |
+| [17dd2e0](https://git.dolgolyov-family.by/alexei.dolgolyov/ledgrab/commit/17dd2e0) | fix: resolve comprehensive review findings (security, concurrency, perf, Android, UI) | alexei.dolgolyov |

 </details>
@@ -41,7 +41,7 @@ android {
        // in CI).  See ledgrabVersionCode above.  Was stuck at 1 before —
        // sideload updates silently refused to install.
        versionCode = ledgrabVersionCode
-        versionName = "0.8.1"
+        versionName = "0.9.0"

        // ABI selection. Detect armeabi-v7a wheel presence and opt the
        // ABI in only when the matching pydantic-core wheel is on disk —
@@ -210,6 +210,10 @@ dependencies {
    implementation("androidx.core:core-splashscreen:1.0.1")
    // QR code generation for displaying server URL on TV
    implementation("com.google.zxing:core:3.5.3")
+    // EncryptedSharedPreferences (Android Keystore-backed) for the per-install
+    // server API key (see ApiKeyManager). Falls back to plain SharedPreferences
+    // when the keystore is unavailable.
+    implementation("androidx.security:security-crypto:1.1.0-alpha06")
    // USB-serial drivers (CH340, CP2102, FTDI, Prolific, CDC-ACM) for
    // driving Adalight/AmbiLED controllers plugged into Android TV boxes.
    implementation("com.github.mik3y:usb-serial-for-android:3.8.1")
@@ -14,7 +14,8 @@
    <uses-permission android:name="android.permission.BLUETOOTH_SCAN"
        android:usesPermissionFlags="neverForLocation"
        tools:targetApi="s" />
-    <uses-permission android:name="android.permission.BLUETOOTH_CONNECT" />
+    <uses-permission android:name="android.permission.BLUETOOTH_CONNECT"
+        tools:targetApi="s" />

    <!-- BLE hardware — required=false so non-BT boxes still install. -->
    <uses-feature
@@ -1,7 +1,10 @@
 package com.ledgrab.android

 import android.content.Context
+import android.content.SharedPreferences
 import android.util.Log
+import androidx.security.crypto.EncryptedSharedPreferences
+import androidx.security.crypto.MasterKey
 import java.security.SecureRandom

 /**
@@ -23,8 +26,23 @@ import java.security.SecureRandom
 */
 class ApiKeyManager(context: Context) {

-    private val prefs = context.applicationContext
-        .getSharedPreferences(PREFS_NAME, Context.MODE_PRIVATE)
+    private val appContext = context.applicationContext
+
+    // Prefer Android-Keystore-backed EncryptedSharedPreferences for the API
+    // key. If the keystore is unavailable (some OEM TV-box ROMs ship a broken
+    // or absent keystore, or a key got corrupted), creation throws — fall back
+    // to plain SharedPreferences so a keystore failure NEVER bricks the local
+    // API key (which would 401 every LAN client).
+    private val prefs: SharedPreferences
+
+    init {
+        val (store, isEncrypted) = buildPrefs(appContext)
+        prefs = store
+        // Only run the plain→encrypted migration when the encrypted store is
+        // actually available; on the degraded plain path there is nothing to
+        // migrate INTO (and recoverLegacyKey reads the backup directly).
+        if (isEncrypted) migrateLegacyKeyIfPresent()
+    }

    // Once we've materialised a key in this process, cache it so
    // subsequent reads don't hit prefs and don't risk re-checking
@@ -60,6 +78,20 @@ class ApiKeyManager(context: Context) {
                cached = existing
                return existing
            }
+            // Before minting a fresh key, fall back to any key still in the
+            // legacy plain store (covers a failed/partial encrypted migration:
+            // commit() can return false WITHOUT throwing, so migration may have
+            // left the live key only in the legacy file). Rotating the
+            // per-install key would 401 every already-paired client, so we
+            // generate a brand-new key ONLY when no key exists anywhere.
+            recoverLegacyKey()?.let { recovered ->
+                // Best-effort persist into the encrypted store; cache regardless
+                // so we still return the recovered key if the write keeps failing.
+                runCatching { prefs.edit().putString(KEY_API_KEY, recovered).commit() }
+                cached = recovered
+                Log.i(TAG, "Recovered existing API key from legacy storage")
+                return recovered
+            }
            val generated = generateKey()
            // commit() (synchronous disk write) on the FIRST write so
            // the key is durable before MainActivity encodes it into a
@@ -74,6 +106,115 @@ class ApiKeyManager(context: Context) {
        }
    }

+    /**
+     * Build the backing store, preferring EncryptedSharedPreferences. Returns
+     * (store, isEncrypted). Any keystore failure falls back to the plain prefs
+     * file so the local API key is never lost on a broken-keystore device.
+     */
+    private fun buildPrefs(context: Context): Pair<SharedPreferences, Boolean> {
+        return try {
+            createEncrypted(context) to true
+        } catch (e: Exception) {
+            // The keystore can become invalidated (OS upgrade, device restore,
+            // OEM keystore bug), after which create() throws on EVERY launch and
+            // the corrupt encrypted file is never cleaned up — degrading to plain
+            // prefs forever and (because the live key was only in the encrypted
+            // store) rotating the per-install key on the next mint, 401-ing every
+            // paired client. Self-heal once: delete the corrupt store + master key
+            // alias and retry create() before degrading.
+            Log.w(TAG, "EncryptedSharedPreferences unavailable, attempting one-time reset: ${e.message}")
+            runCatching {
+                context.deleteSharedPreferences(ENCRYPTED_PREFS_NAME)
+                runCatching {
+                    val ks = java.security.KeyStore.getInstance("AndroidKeyStore").apply { load(null) }
+                    if (ks.containsAlias(MasterKey.DEFAULT_MASTER_KEY_ALIAS)) {
+                        ks.deleteEntry(MasterKey.DEFAULT_MASTER_KEY_ALIAS)
+                    }
+                }.onFailure { Log.w(TAG, "Master-key alias cleanup failed: ${it.message}") }
+                createEncrypted(context) to true
+            }.getOrElse {
+                // Still failing after reset — degrade to plain prefs rather than
+                // crashing. Worst case the key is stored unencrypted on a
+                // single-user TV box, which is the pre-existing behaviour.
+                Log.w(TAG, "EncryptedSharedPreferences still unavailable after reset, using plain prefs: ${it.message}")
+                context.getSharedPreferences(PREFS_NAME, Context.MODE_PRIVATE) to false
+            }
+        }
+    }
+
+    private fun createEncrypted(context: Context): SharedPreferences {
+        val masterKey = MasterKey.Builder(context)
+            .setKeyScheme(MasterKey.KeyScheme.AES256_GCM)
+            .build()
+        return EncryptedSharedPreferences.create(
+            context,
+            ENCRYPTED_PREFS_NAME,
+            masterKey,
+            EncryptedSharedPreferences.PrefKeyEncryptionScheme.AES256_SIV,
+            EncryptedSharedPreferences.PrefValueEncryptionScheme.AES256_GCM,
+        )
+    }
+
+    /**
+     * One-time migration: if a key exists in the legacy plain-text prefs file
+     * (from before encrypted storage), copy it into the encrypted store and
+     * remove the plain copy. Preserves the existing key so already-scanned QR
+     * clients keep working — generating a fresh key here would silently 401
+     * every LAN client (see the Data Migration Policy in CLAUDE.md).
+     */
+    private fun migrateLegacyKeyIfPresent() {
+        // Don't migrate if the encrypted store already holds a key.
+        if (!prefs.getString(KEY_API_KEY, null).isNullOrEmpty()) return
+        runCatching {
+            val legacy = appContext.getSharedPreferences(PREFS_NAME, Context.MODE_PRIVATE)
+            val legacyKey = legacy.getString(KEY_API_KEY, null)
+            if (legacyKey != null && legacyKey.length >= MIN_KEY_LENGTH) {
+                // commit() returns false on write failure WITHOUT throwing, so the
+                // runCatching wrapper alone does NOT protect this path. Verify the
+                // encrypted store both committed AND reads back the identical value
+                // before touching the legacy copy — otherwise a silent write
+                // failure could delete the only surviving copy of the key and
+                // rotate it on next launch (401s every paired client — the exact
+                // silent-data-loss the Data Migration Policy forbids).
+                val ok = prefs.edit().putString(KEY_API_KEY, legacyKey).commit()
+                if (ok && prefs.getString(KEY_API_KEY, null) == legacyKey) {
+                    // Keep the value as a .migrated backup (don't hard-delete) per
+                    // the migration policy; remove only the live legacy key so the
+                    // plaintext copy no longer answers reads.
+                    legacy.edit()
+                        .putString(KEY_API_KEY_MIGRATED, legacyKey)
+                        .remove(KEY_API_KEY)
+                        .apply()
+                    Log.i(TAG, "Migrated API key from plain to encrypted storage")
+                } else {
+                    // Leave the legacy key untouched; getOrCreateKey() will recover
+                    // it via recoverLegacyKey() rather than minting a fresh one.
+                    Log.w(TAG, "Encrypted key write unverified — keeping legacy key, not migrating")
+                }
+            }
+        }.onFailure { Log.w(TAG, "Legacy API key migration failed: ${it.message}") }
+    }
+
+    /**
+     * Recover a still-present key from the legacy plain store — either the live
+     * key (failed/never-run migration) or the `.migrated` backup. Returns null
+     * only when no valid key survives.
+     *
+     * This MUST run on the degraded plain-prefs path too (not just the encrypted
+     * path): after a successful migration the live key is moved to the
+     * `.migrated` backup in this same plain file, so when the keystore later
+     * fails and we degrade to plain prefs, the backup is the only surviving
+     * copy. Returning null here (the previous `if (!encrypted) return null`
+     * guard) would mint a fresh key and rotate the per-install key, 401-ing every
+     * paired client.
+     */
+    private fun recoverLegacyKey(): String? {
+        val legacy = appContext.getSharedPreferences(PREFS_NAME, Context.MODE_PRIVATE)
+        val candidate = legacy.getString(KEY_API_KEY, null)
+            ?: legacy.getString(KEY_API_KEY_MIGRATED, null)
+        return candidate?.takeIf { it.length >= MIN_KEY_LENGTH }
+    }
+
    private fun generateKey(): String {
        val bytes = ByteArray(KEY_BYTES)
        SecureRandom().nextBytes(bytes)
@@ -88,7 +229,11 @@ class ApiKeyManager(context: Context) {
    companion object {
        private const val TAG = "ApiKeyManager"
        private const val PREFS_NAME = "ledgrab_auth"
+        private const val ENCRYPTED_PREFS_NAME = "ledgrab_auth_enc"
        private const val KEY_API_KEY = "api_key"
+        // Backup of a migrated legacy key, kept in the plain store per the
+        // Data Migration Policy (never hard-delete user data on rename/move).
+        private const val KEY_API_KEY_MIGRATED = "api_key_migrated"
        private const val KEY_BYTES = 32
        private const val MIN_KEY_LENGTH = 32

@@ -103,12 +103,32 @@ object BleBridge {
        }

        try {
-            bleHandler.post { scanner.startScan(callback) }
+            // startScan runs on the BLE handler thread; a denied
+            // BLUETOOTH_SCAN throws SecurityException there, which would
+            // crash the whole process (an uncaught exception on a handler
+            // thread is fatal). Catch it inside the posted body and report.
+            bleHandler.post {
+                try {
+                    scanner.startScan(callback)
+                } catch (e: SecurityException) {
+                    Log.w(TAG, "BLUETOOTH_SCAN permission denied — scan skipped", e)
+                } catch (e: Exception) {
+                    Log.w(TAG, "BLE startScan failed: ${e.message}")
+                }
+            }
            Thread.sleep(timeoutMs)
        } catch (_: InterruptedException) {
            Thread.currentThread().interrupt()
        } finally {
-            try { bleHandler.post { scanner.stopScan(callback) } } catch (_: SecurityException) {}
+            bleHandler.post {
+                try {
+                    scanner.stopScan(callback)
+                } catch (e: SecurityException) {
+                    Log.w(TAG, "BLUETOOTH_SCAN permission denied — stopScan skipped", e)
+                } catch (e: Exception) {
+                    Log.w(TAG, "BLE stopScan failed: ${e.message}")
+                }
+            }
        }
        return seen.values.toList()
    }
@@ -136,7 +156,18 @@ object BleBridge {
                    newState == BluetoothProfile.STATE_CONNECTED
                            && status == BluetoothGatt.GATT_SUCCESS -> {
                        Log.d(TAG, "GATT connected to $address, discovering services")
-                        gatt.discoverServices()
+                        // Runs on the BLE handler thread; a denied
+                        // BLUETOOTH_CONNECT throws SecurityException here, which
+                        // would crash the process. Catch and fail the connect.
+                        try {
+                            gatt.discoverServices()
+                        } catch (e: SecurityException) {
+                            Log.e(TAG, "BLUETOOTH_CONNECT denied during discoverServices", e)
+                            readyDeferred.complete(false)
+                        } catch (e: Exception) {
+                            Log.w(TAG, "discoverServices failed: ${e.message}")
+                            readyDeferred.complete(false)
+                        }
                    }
                    newState == BluetoothProfile.STATE_DISCONNECTED -> {
                        Log.w(TAG, "GATT disconnected from $address (status=$status)")
@@ -105,12 +105,48 @@ class CaptureService : Service() {
    override fun onStartCommand(intent: Intent?, flags: Int, startId: Int): Int {
        val useRoot = intent?.getBooleanExtra(EXTRA_USE_ROOT, false) ?: false

-        // CRITICAL: startForeground must be called IMMEDIATELY — before
-        // any other work, especially before getMediaProjection(). The
-        // service type must match the work; pass it explicitly via
-        // ServiceCompat so we stay compatible back to API 24.
+        // CRITICAL (Android 14+): for the MediaProjection path, validate the
+        // projection token BEFORE promoting to a foreground service with the
+        // mediaProjection FGS type. On service recreation (system redelivery
+        // or a stale relaunch) the consent token is gone — promoting first and
+        // then discovering the dead token causes a spurious foreground-service
+        // start + immediate stop, which on strict OEMs flickers the
+        // notification or trips a stopSelf loop. Bail out cleanly here, before
+        // startForeground, when the MediaProjection consent data is missing.
        val localIp = NetworkUtils.getLocalIpAddress(this) ?: "—"
        val url = "http://$localIp:$SERVER_PORT"
+
+        val mediaProjectionResultData: Intent? =
+            if (!useRoot) extractProjectionResultData(intent) else null
+        if (!useRoot && (intent == null || mediaProjectionResultData == null)) {
+            // MediaProjection mode can't recover from a redelivery —
+            // the consent token in the original intent is single-use.
+            //
+            // We were launched via startForegroundService(), so the OS REQUIRES
+            // a startForeground() within ~5s even on this immediate-stop path,
+            // or it raises the fatal ForegroundServiceDidNotStartInTimeException.
+            // Promote with a benign SPECIAL_USE type (NOT mediaProjection — we
+            // have no valid consent token, and requesting that type without an
+            // active projection is exactly what we're avoiding) just long enough
+            // to satisfy the contract, then stop.
+            Log.w(TAG, "MediaProjection start without a valid consent token — stopping")
+            runCatching {
+                val bailType = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.UPSIDE_DOWN_CAKE) {
+                    ServiceInfo.FOREGROUND_SERVICE_TYPE_SPECIAL_USE
+                } else {
+                    0
+                }
+                ServiceCompat.startForeground(this, NOTIFICATION_ID, buildNotification(url), bailType)
+            }.onFailure { Log.w(TAG, "Bail-path startForeground failed: ${it.message}") }
+            stopSelf()
+            return START_NOT_STICKY
+        }
+
+        // startForeground must be called IMMEDIATELY after the token check —
+        // before any heavier work like getMediaProjection(). The service type
+        // must match the work; pass it explicitly via ServiceCompat so we stay
+        // compatible back to API 24. The MEDIA_PROJECTION type is only used
+        // here once resultData is confirmed non-null (checked above).
        try {
            val type = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.UPSIDE_DOWN_CAKE) {
                var t = if (useRoot) {
@@ -152,20 +188,13 @@ class CaptureService : Service() {
        // otherwise `isRunning=true` sticks forever when startForeground throws.
        isRunning = true

-        if (intent == null && !useRoot) {
-            // MediaProjection mode can't recover from a redelivery —
-            // the consent token in the original intent is single-use.
-            Log.w(TAG, "Service restarted without intent (MediaProjection mode) — stopping")
-            isRunning = false
-            stopSelf()
-            return START_NOT_STICKY
-        }
-
        try {
            if (useRoot) {
                startRootCapture(url)
            } else {
-                startMediaProjectionCapture(intent!!, url)
+                // mediaProjectionResultData is guaranteed non-null here — the
+                // token was validated before startForeground above.
+                startMediaProjectionCapture(intent!!, mediaProjectionResultData!!, url)
            }
        } catch (e: Exception) {
            Log.e(TAG, "Failed to start capture", e)
@@ -294,20 +323,24 @@ class CaptureService : Service() {
        }
    }

-    private fun startMediaProjectionCapture(intent: Intent, url: String) {
-        val resultCode = intent.getIntExtra(EXTRA_RESULT_CODE, 0)
+    /**
+     * Extract the single-use MediaProjection consent token from the start
+     * intent, or null if the intent is missing/redelivered without it.
+     * Called BEFORE startForeground so the mediaProjection FGS type is only
+     * ever requested when a valid token is present (see onStartCommand).
+     */
+    private fun extractProjectionResultData(intent: Intent?): Intent? {
+        if (intent == null) return null
        @Suppress("DEPRECATION")
-        val resultData = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU) {
+        return if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU) {
            intent.getParcelableExtra(EXTRA_RESULT_DATA, Intent::class.java)
        } else {
            intent.getParcelableExtra(EXTRA_RESULT_DATA)
        }
+    }

-        if (resultData == null) {
-            Log.e(TAG, "No MediaProjection result data")
-            stopSelf()
-            return
-        }
+    private fun startMediaProjectionCapture(intent: Intent, resultData: Intent, url: String) {
+        val resultCode = intent.getIntExtra(EXTRA_RESULT_CODE, 0)

        val projectionManager =
            getSystemService(Context.MEDIA_PROJECTION_SERVICE) as MediaProjectionManager
@@ -6,7 +6,9 @@ import android.service.notification.StatusBarNotification
 import android.util.Log
 import com.chaquo.python.Python
 import java.util.concurrent.ConcurrentHashMap
+import java.util.concurrent.ExecutorService
 import java.util.concurrent.Executors
+import java.util.concurrent.RejectedExecutionException

 /**
 * Captures posted OS notifications and forwards the posting app's display
@@ -25,7 +27,26 @@ class LedGrabNotificationListener : NotificationListenerService() {
    // Serial executor: the Python receiver does a (non-concurrency-safe) history
    // disk write and may play a sound, so pushes must not overlap. Off the main
    // looper to keep the system service responsive.
-    private val pushExecutor = Executors.newSingleThreadExecutor()
+    //
+    // Tied to the listener-connection lifecycle (onListenerConnected /
+    // onListenerDisconnected), NOT onDestroy: this is a system-rebindable
+    // service, so it can be connected/disconnected multiple times across a
+    // single onCreate..onDestroy span. Managing the executor here — combined
+    // with the runCatching guard at the submit site — keeps a notification
+    // that races teardown from triggering RejectedExecutionException on a
+    // shut-down executor. @Volatile so the connect/disconnect callbacks (which
+    // may run on a different thread than onNotificationPosted) publish safely.
+    @Volatile private var pushExecutor: ExecutorService? = null
+
+    // Guards executor creation so the lazy submit-site fallback and
+    // onListenerConnected can't race two executors into existence.
+    private val executorLock = Any()
+
+    // Tracks whether the listener is currently connected. ensureExecutor() only
+    // CREATES a new executor while connected — otherwise a notification racing
+    // onListenerDisconnected (which nulls pushExecutor) would spin up a fresh
+    // executor that nothing reaps until the next disconnect cycle (a thread leak).
+    @Volatile private var connected: Boolean = false

    // packageName -> resolved human-readable label. Matches the app_name the
    // Windows/Linux backends pass, so per-app colors/filters keep working.
@@ -51,17 +72,37 @@ class LedGrabNotificationListener : NotificationListenerService() {

        val label = resolveAppLabel(notification.packageName)

-        pushExecutor.execute {
-            try {
-                Python.getInstance()
-                    .getModule(PY_MODULE)
-                    .callAttr("push_notification", label)
-            } catch (t: Throwable) {
-                // Never crash a system-bound service. Python.getInstance() throws
-                // IllegalStateException if Python.start() hasn't run (e.g. the
-                // service was bound at boot before the app process initialized).
-                // Log at debug — the label is potentially sensitive on a shared TV.
-                Log.d(TAG, "push_notification failed: ${t.message}")
+        // Obtain (creating if needed) the executor. onListenerConnected normally
+        // creates it, but that callback is not reliably invoked on every
+        // OEM/version (re)bind, and a notification can arrive before it fires —
+        // lazily creating here keeps a missing/late onListenerConnected from
+        // permanently disabling notification forwarding. A late submit onto an
+        // executor that onListenerDisconnected is shutting down throws
+        // RejectedExecutionException — guard with runCatching so a notification
+        // racing teardown can never crash this system-bound service.
+        val executor = ensureExecutor() ?: run {
+            Log.d(TAG, "no executor (listener disconnected) — skipping push")
+            return
+        }
+        runCatching {
+            executor.execute {
+                try {
+                    Python.getInstance()
+                        .getModule(PY_MODULE)
+                        .callAttr("push_notification", label)
+                } catch (t: Throwable) {
+                    // Never crash a system-bound service. Python.getInstance() throws
+                    // IllegalStateException if Python.start() hasn't run (e.g. the
+                    // service was bound at boot before the app process initialized).
+                    // Log at debug — the label is potentially sensitive on a shared TV.
+                    Log.d(TAG, "push_notification failed: ${t.message}")
+                }
+            }
+        }.onFailure { e ->
+            if (e is RejectedExecutionException) {
+                Log.d(TAG, "push rejected — listener disconnecting")
+            } else {
+                throw e
            }
        }
    }
@@ -69,24 +110,64 @@ class LedGrabNotificationListener : NotificationListenerService() {
    /** Resolve (and cache) a package's human-readable label; fall back to the package name. */
    private fun resolveAppLabel(pkg: String): String {
        labelCache[pkg]?.let { return it }
+        // Only cache SUCCESSFUL resolutions. Caching the package-name fallback
+        // would permanently pin a wrong label if the PackageManager lookup
+        // failed transiently (e.g. the app was mid-install / still updating).
        val resolved = runCatching {
            val info = packageManager.getApplicationInfo(pkg, 0)
            packageManager.getApplicationLabel(info).toString()
-        }.getOrDefault(pkg)
-        labelCache[pkg] = resolved
-        return resolved
+        }.getOrNull()
+        if (resolved != null) {
+            labelCache[pkg] = resolved
+            return resolved
+        }
+        return pkg
+    }
+
+    /**
+     * Return the push executor, creating it under [executorLock] if absent AND
+     * the listener is connected. Returns null when disconnected so a notification
+     * racing teardown neither submits onto a shutting-down executor nor spins up
+     * a stray one. Safe against a concurrent onListenerConnected/onNotificationPosted
+     * race (single executor) and against a missing onListenerConnected callback.
+     */
+    private fun ensureExecutor(): ExecutorService? {
+        pushExecutor?.let { return it }
+        synchronized(executorLock) {
+            if (!connected) return null
+            return pushExecutor ?: Executors.newSingleThreadExecutor().also { pushExecutor = it }
+        }
    }

    override fun onListenerConnected() {
        Log.i(TAG, "Notification listener connected")
+        // Spin up the push executor on connect. The system can disconnect and
+        // later reconnect this service without destroying it, so own the
+        // executor here rather than in onCreate/onDestroy. onNotificationPosted
+        // also lazily creates it (via ensureExecutor) in case this callback is
+        // late or skipped on some ROMs.
+        connected = true
+        ensureExecutor()
    }

    override fun onListenerDisconnected() {
        Log.i(TAG, "Notification listener disconnected")
+        // Mark disconnected BEFORE nulling the executor so a racing ensureExecutor
+        // sees !connected and skips creating a replacement. Tear the executor
+        // down; a fresh one is created on the next onListenerConnected.
+        connected = false
+        pushExecutor?.let { exec ->
+            pushExecutor = null
+            exec.shutdown()
+        }
    }

    override fun onDestroy() {
-        pushExecutor.shutdown()
+        // Defensive: onListenerDisconnected normally clears this first, but
+        // shut down here too in case onDestroy fires without a prior disconnect.
+        connected = false
+        pushExecutor?.shutdown()
+        pushExecutor = null
        super.onDestroy()
    }

@@ -34,6 +34,7 @@ import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.MainScope
 import kotlinx.coroutines.cancel
 import kotlinx.coroutines.launch
+import kotlinx.coroutines.runInterruptible
 import kotlinx.coroutines.withContext

 /**
@@ -56,6 +57,7 @@ class MainActivity : Activity() {
        private const val REQUEST_POST_NOTIFICATIONS = 1002
        private const val REQUEST_RECORD_AUDIO = 1003
        private const val REQUEST_CAMERA = 1004
+        private const val REQUEST_BLUETOOTH = 1005
        private const val QR_SIZE_PX = 560
        private const val NOTIF_PREFS = "ledgrab_notif"
        private const val KEY_NOTIF_ACCESS_PROMPTED = "notif_access_prompted"
@@ -189,7 +191,13 @@ class MainActivity : Activity() {
        toggleButton.text = getString(R.string.btn_starting)
        statusText.text = getString(R.string.status_checking_root)
        uiScope.launch(Dispatchers.IO) {
-            val rooted = Root.requestGrant()
+            // runInterruptible so a config change (rotation) during the
+            // up-to-10s `su` probe cancels the coroutine AND interrupts the
+            // blocking probe thread — Root.requestGrant honours the interrupt,
+            // destroys the su child, and rethrows, so we don't leak the
+            // process + drain thread. Without this, IO-dispatcher cancellation
+            // would not interrupt the blocking waitFor().
+            val rooted = runInterruptible { Root.requestGrant() }
            withContext(Dispatchers.Main) {
                toggleButton.isEnabled = true
                toggleButton.text = originalText
@@ -214,6 +222,7 @@ class MainActivity : Activity() {
        ensureNotificationPermission()
        ensureNotificationListenerAccess()
        ensureCameraPermission()
+        ensureBluetoothPermissions()
        ContextCompat.startForegroundService(this, CaptureService.createRootIntent(this))
        updateUI()
    }
@@ -236,6 +245,7 @@ class MainActivity : Activity() {
        ensureNotificationListenerAccess()
        ensureAudioPermission()
        ensureCameraPermission()
+        ensureBluetoothPermissions()
        val intent = CaptureService.createIntent(this, resultCode, resultData)
        ContextCompat.startForegroundService(this, intent)
        updateUI()
@@ -536,6 +546,30 @@ class MainActivity : Activity() {
        }
    }

+    /**
+     * Request BLUETOOTH_SCAN + BLUETOOTH_CONNECT (API 31+) so the embedded
+     * server can discover and drive BLE LED controllers (SP110E / Triones /
+     * Zengge). On API < 31 these are install-time legacy permissions
+     * (BLUETOOTH / BLUETOOTH_ADMIN / ACCESS_FINE_LOCATION, maxSdk=30) and
+     * need no runtime grant — so this is a no-op there. Fire-and-forget,
+     * like [ensureAudioPermission]: screen capture works without BLE, and
+     * BleBridge degrades gracefully (empty scan / failed connect) when the
+     * grant is denied, so we don't block on the result. If first granted
+     * here, BLE devices become reachable on the next scan/connect.
+     */
+    private fun ensureBluetoothPermissions() {
+        if (Build.VERSION.SDK_INT < Build.VERSION_CODES.S) return
+        val needed = listOf(
+            Manifest.permission.BLUETOOTH_SCAN,
+            Manifest.permission.BLUETOOTH_CONNECT,
+        ).filter {
+            checkSelfPermission(it) != PackageManager.PERMISSION_GRANTED
+        }
+        if (needed.isEmpty()) return
+        @Suppress("DEPRECATION")
+        requestPermissions(needed.toTypedArray(), REQUEST_BLUETOOTH)
+    }
+
    /** Whether the user has granted notification-listener access to this app. */
    private fun isNotificationAccessGranted(): Boolean =
        NotificationManagerCompat.getEnabledListenerPackages(this).contains(packageName)
@@ -20,6 +20,11 @@ import java.util.concurrent.TimeUnit
 object Root {
    private const val TAG = "Root"

+    // Slice length for the cancellation-aware su probe wait loop. Short
+    // enough that coroutine cancellation is honoured promptly, long enough
+    // to avoid busy-spinning while Magisk's grant dialog is up.
+    private const val POLL_SLICE_MS = 100L
+
    private val SU_PATHS = listOf(
        "/system/bin/su",
        "/system/xbin/su",
@@ -49,17 +54,19 @@ object Root {
            return false
        }

+        var process: Process? = null
        val granted = try {
            // redirectErrorStream merges stderr into stdout so a single
            // drain thread is enough — avoids the classic pipe-buffer
            // deadlock where waitFor() blocks because stderr filled up.
-            val process = ProcessBuilder("su", "-c", "id")
+            val proc = ProcessBuilder("su", "-c", "id")
                .redirectErrorStream(true)
                .start()
+            process = proc
            val outputBuilder = StringBuilder()
            val drain = Thread({
                try {
-                    BufferedReader(InputStreamReader(process.inputStream)).use { r ->
+                    BufferedReader(InputStreamReader(proc.inputStream)).use { r ->
                        val buf = CharArray(512)
                        while (true) {
                            val n = r.read(buf)
@@ -72,17 +79,35 @@ object Root {
                }
            }, "Root-su-drain").apply { isDaemon = true; start() }

-            val finished = process.waitFor(timeoutSeconds, TimeUnit.SECONDS)
+            // Cancellation-aware wait: callers run this on a coroutine
+            // (MainActivity wraps it in runInterruptible), so a config change
+            // mid-probe cancels the coroutine and interrupts this thread.
+            // Poll waitFor() in short slices and honour interruption so we
+            // don't leak the `su` child + its drain thread for up to 10s.
+            // The catch(InterruptedException) below destroys the process; we
+            // re-arm the interrupt and rethrow so coroutine cancellation
+            // propagates cleanly.
+            val deadlineNanos = System.nanoTime() + TimeUnit.SECONDS.toNanos(timeoutSeconds)
+            var finished = false
+            while (System.nanoTime() < deadlineNanos) {
+                if (proc.waitFor(POLL_SLICE_MS, TimeUnit.MILLISECONDS)) {
+                    finished = true
+                    break
+                }
+                // Throws InterruptedException if the thread was interrupted
+                // by coroutine cancellation — handled below to tear down.
+                if (Thread.interrupted()) throw InterruptedException("su probe cancelled")
+            }
            if (!finished) {
-                process.destroyForcibly()
+                proc.destroyForcibly()
                drain.join(500)
                Log.w(TAG, "su -c id timed out after ${timeoutSeconds}s")
                false
            } else {
                drain.join(500)
                val output = synchronized(outputBuilder) { outputBuilder.toString() }
-                if (process.exitValue() != 0) {
-                    Log.w(TAG, "su -c id exited with ${process.exitValue()} output='${output.trim()}'")
+                if (proc.exitValue() != 0) {
+                    Log.w(TAG, "su -c id exited with ${proc.exitValue()} output='${output.trim()}'")
                    false
                } else {
                    val rooted = output.contains("uid=0")
@@ -90,8 +115,17 @@ object Root {
                    rooted
                }
            }
+        } catch (e: InterruptedException) {
+            // Coroutine cancelled mid-probe (e.g. config change). Kill the
+            // su child so it doesn't outlive the cancelled work, re-arm the
+            // interrupt flag, and rethrow so the coroutine cancels cleanly.
+            // Do NOT cache a result — the probe never completed.
+            runCatching { process?.destroyForcibly() }
+            Thread.currentThread().interrupt()
+            throw e
        } catch (e: Exception) {
            Log.w(TAG, "su invocation failed: ${e.message}")
+            runCatching { process?.destroyForcibly() }
            false
        }

@@ -89,14 +89,15 @@ class RootScreenrecord(
        running = true

        try {
-            imageReader = buildImageReader()
-            decoder = buildDecoder(imageReader!!)
-            process = spawnScreenrecord() ?: run {
+            val reader = buildImageReader().also { imageReader = it }
+            val codec = buildDecoder(reader).also { decoder = it }
+            val proc = spawnScreenrecord() ?: run {
                stop()
                return false
            }
-            startInputPump(process!!.inputStream, decoder!!)
-            startOutputDrain(decoder!!)
+            process = proc
+            startInputPump(proc.inputStream, codec)
+            startOutputDrain(codec)
            Log.i(TAG, "Root capture pipeline started (${width}x$height @ ${fps}fps)")
            return true
        } catch (e: Exception) {
@@ -178,6 +179,14 @@ class RootScreenrecord(
                        buffer.get(frameBuffer, row * rowBytes, rowBytes)
                    }
                }
+                // CONTRACT: frameBuffer is REUSED across frames (single-threaded
+                // reader callback — no copy here). Safety depends on the Python
+                // receiver copying the bytes before this callback returns and
+                // overwrites the buffer for the next frame. It does:
+                // PythonBridge.pushRootFrame → root_screenrecord_engine.push_frame
+                // (server/src/ledgrab/core/capture_engines/root_screenrecord_engine.py)
+                // does `rgba[:, :, :3].copy()`, so the queued frame owns its
+                // pixels independently of this buffer. Do NOT remove that copy.
                bridge.pushRootFrame(frameBuffer, width, height)
                framesDeliveredCounter.incrementAndGet()
            } catch (e: Exception) {
@@ -147,6 +147,14 @@ class ScreenCapture(
                    }
                }

+                // CONTRACT: frameBuffer is REUSED across frames (single-threaded
+                // capture handler — no copy here). Safety depends on the Python
+                // receiver copying the bytes before this callback returns and
+                // overwrites the buffer for the next frame. It does:
+                // PythonBridge.pushFrame → mediaprojection_engine.push_frame
+                // (server/src/ledgrab/core/capture_engines/mediaprojection_engine.py)
+                // does `rgba[:, :, :3].copy()`, so the queued frame owns its
+                // pixels independently of this buffer. Do NOT remove that copy.
                bridge.pushFrame(frameBuffer, captureWidth, captureHeight)

                // Advance the pacing accumulator. If we fell badly behind
@@ -31,9 +31,15 @@ Creates the Gitea release with a description table listing all artifacts. **The
 - Produces: **`LedGrab-{tag}-linux-x64.tar.gz`**

 ### 4. `build-docker`
- Plain `docker build` + `docker push` (no Buildx — TrueNAS runners lack nested networking)
+- Plain `docker build` + `docker push` for **amd64** (no Buildx — TrueNAS runners lack nested networking)
 - Registry: `{gitea_host}/{repo}:{tag}`
 - Tags: `v0.x.x`, `0.x.x`, and `latest` (stable only, not alpha/beta/rc)
+- **arm64 is best-effort** (Raspberry Pi / arm64 HAOS): a `continue-on-error` step
+  cross-builds arm64 via **QEMU binfmt** (`tonistiigi/binfmt`) + `docker manifest`
+  (NOT buildx — sidesteps the docker-container-driver networking limit) and folds
+  amd64 + arm64 into multi-arch manifest lists under the same tags, plus
+  `:{tag}-amd64` / `:{tag}-arm64` arch-suffixed tags. If the runner can't run
+  privileged binfmt the step is skipped and the amd64 tags above remain valid.

 ## Build Scripts

@@ -42,6 +42,7 @@ Complete REST + WebSocket API reference for the LedGrab server.
  - [Weather sources](#weather-sources)
  - [Automations](#automations)
  - [Scene presets](#scene-presets)
+  - [Scene playlists](#scene-playlists)
  - [Sync clocks](#sync-clocks)
  - [Webhooks](#webhooks)
  - [HTTP endpoints](#http-endpoints)
@@ -184,7 +185,7 @@ Server configuration: MQTT broker, external URL, shutdown action, log level, ADB

 ## User preferences

-Dashboard layout, notification settings, card display modes, and the global daylight timezone.
+Dashboard layout, notification settings, card display modes, the global daylight timezone, and the first-run onboarding flag.

 | Method | Path | Description |
 | ------ | ---- | ----------- |
@@ -198,6 +199,19 @@ Dashboard layout, notification settings, card display modes, and the global dayl
 | DELETE | `/api/v1/preferences/card-modes` | Delete card-mode preferences; revert to defaults. |
 | GET | `/api/v1/preferences/daylight-timezone` | Read the global IANA timezone for daylight cycles. |
 | PUT | `/api/v1/preferences/daylight-timezone` | Persist the daylight-cycle timezone (empty = server local). |
+| GET | `/api/v1/preferences/onboarding` | Read the first-run onboarding flag (`onboarded: bool`, `completed_at: str\|null`). Defaults to `false`. |
+| PUT | `/api/v1/preferences/onboarding` | Persist the onboarding flag. Server auto-stamps `completed_at` when `onboarded` is set to `true` without a timestamp. |
+
+**Onboarding flag response shape:**
+
+```json
+{
+  "onboarded": true,
+  "completed_at": "2026-06-08T12:00:00.000000+00:00"
+}
+```
+
+Defaults to `{"onboarded": false, "completed_at": null}` when never set.

 ## Backup, restore & server control

@@ -237,7 +251,7 @@ A single aggregated poll endpoint for low-overhead clients.

 | Method | Path | Description |
 | ------ | ---- | ----------- |
-| GET | `/api/v1/snapshot` | Full poll payload (targets, states, metrics, devices, brightness, color/value sources, scene presets, sync clocks, system) in one response. Use `?include=` to request a subset; per-section fault isolation. |
+| GET | `/api/v1/snapshot` | Full poll payload (targets, states, metrics, devices, brightness, color/value sources, scene presets, scene playlists + cycling state, sync clocks, system) in one response. Use `?include=` to request a subset; per-section fault isolation. |

 ## Devices

@@ -518,6 +532,25 @@ Captured snapshots of target state that can be restored.
 | POST | `/api/v1/scene-presets/{preset_id}/recapture` | Re-capture current state into the preset. |
 | POST | `/api/v1/scene-presets/{preset_id}/activate` | Activate the preset (restore captured state). |

+## Scene playlists
+
+Ordered, timed sequences of scene presets that auto-cycle. The engine drives
+**one** playlist at a time — starting a playlist stops any other. Each item
+references a scene preset and holds it for its `duration_seconds` (min 1s)
+before advancing; `loop` repeats from the start and `shuffle` randomises the
+order each cycle.
+
+| Method | Path | Description |
+| ------ | ---- | ----------- |
+| POST | `/api/v1/scene-playlists` | Create a playlist (items reference scene presets). |
+| GET | `/api/v1/scene-playlists` | List all playlists plus the current cycling `state`. |
+| GET | `/api/v1/scene-playlists/state` | Get the current cycling state (idle if nothing runs). |
+| GET | `/api/v1/scene-playlists/{playlist_id}` | Get a playlist by ID. |
+| PUT | `/api/v1/scene-playlists/{playlist_id}` | Update metadata, items, and `loop`/`shuffle`. |
+| DELETE | `/api/v1/scene-playlists/{playlist_id}` | Delete a playlist (stops it first if running). |
+| POST | `/api/v1/scene-playlists/{playlist_id}/start` | Start cycling (stops any other playlist first). |
+| POST | `/api/v1/scene-playlists/stop` | Stop the active playlist (leaves the last scene applied). |
+
 ## Sync clocks

 Shared clocks that drive linked animations with configurable speed.
@@ -629,6 +662,127 @@ The wiring-graph: schema registry, topology, dependents, validation, and subgrap
 | POST | `/api/v1/graph/validate-connection` | Validate a proposed wiring edit (existence, kind, no cycle). |
 | POST | `/api/v1/graph/duplicate` | Deep-clone selected value/color-strip sources with remapped wiring. |

+## Calibration
+
+Guided LED chase and auto-solver for the `CalibrationConfig` stored on a
+color-strip source.  The flow is:
+
+1. **Start** a session (`POST /session`) — stops any running target on the
+   device and remembers it for restore on stop.
+2. **Position** the chase pixel (`POST /session/position`) to walk through
+   each physical corner and record the LED index.
+3. **Solve** (`POST /solve`) — the server computes per-edge LED counts.
+4. **Persist** — call `PUT /api/v1/color-strip-sources/{id}` with the solved
+   `calibration` object to save and hot-reload.
+5. **Stop** (`POST /session/stop`) — clears the device and restores the prior
+   target.
+
+| Method | Path | Description |
+| ------ | ---- | ----------- |
+| POST | `/api/v1/calibration/session` | Start a calibration session on a device (stops the running target, clears to black). |
+| POST | `/api/v1/calibration/session/position` | Advance the chase pixel to LED `index` (± `window` dim neighbours). |
+| POST | `/api/v1/calibration/session/stop` | End the session: clear to black and restore the prior target. |
+| POST | `/api/v1/calibration/session/cancel` | Alias for stop — no calibration is applied. |
+| GET | `/api/v1/calibration/session/state` | Current session state (active, device_id, led_count, last_activity). |
+| POST | `/api/v1/calibration/solve` | Solve per-edge LED counts from 4 corner tap indices. Returns solved config dict (does NOT persist). |
+
+**Session state** response shape:
+
+```json
+{
+  "active": true,
+  "device_id": "dev_abc123",
+  "led_count": 100,
+  "prior_target_id": "ot_xyz456",
+  "last_activity": "2026-06-08T12:34:56.789Z"
+}
+```
+
+**Solve request** (body):
+
+```json
+{
+  "device_id": "dev_abc123",
+  "start_position": "bottom_left",
+  "layout": "clockwise",
+  "corner_indices": [0, 30, 60, 80],
+  "offset": 0
+}
+```
+
+`corner_indices` must be exactly 4 integers, one per screen corner, in the
+strip-walk order defined by `(start_position, layout)`.  Provide either
+`device_id` (preferred — server derives `led_count`) or `led_count` directly.
+
+**Important session behavior:**
+
+- **Stops the running output target** — starting a calibration session immediately
+  stops any output target currently running on that device.  Other clients driving
+  that device will lose their output for the duration of the session.
+- **Single session only** — only one calibration session runs at a time across the
+  whole server.  Starting a new session automatically ends the previous one (clearing
+  and restoring its device first), regardless of which device each session is on.
+- **Idle auto-end** — a session that receives no `position` calls for ~60 seconds is
+  automatically stopped and the prior target restored, so devices are never left dark
+  indefinitely.
+
+**Idle timeout:** a session that receives no `position` calls for 60 seconds
+is automatically stopped and the prior target restored.
+
+## Setup scaffold
+
+One-call first-run helper that creates the full capture-to-output chain and
+returns all entity ids.  The wizard calls this, then starts the output target
+after optional calibration.
+
+| Method | Path | Description |
+| ------ | ---- | ----------- |
+| POST | `/api/v1/setup/scaffold` | Create capture template + picture source + color-strip source + LED output target in one atomic call with rollback on partial failure. Does NOT auto-start the target. |
+
+**Wizard sequence (Phase 4):**
+
+1. Discover or create the device via `POST /api/v1/devices` (full URL
+   normalisation + provider validation runs there).
+2. Call `POST /api/v1/setup/scaffold` with the resulting `device_id`.
+3. Calibrate (Phase 1 endpoints).
+4. Start the output target via `POST /api/v1/output-targets/{id}/start`.
+
+**Request body:**
+
+```json
+{
+  "device_id": "device_abc123",
+  "display_index": 0,
+  "calibration": null
+}
+```
+
+`device_id` is **required** and must reference an existing device (created via
+`POST /api/v1/devices`).  `display_index` selects the monitor to capture
+(0 = primary; range 0–63).  `calibration` is an optional `CalibrationConfig`
+dict; when omitted, `create_default_calibration(led_count)` is used.
+
+**Response (201 Created):**
+
+```json
+{
+  "device_id": "device_abc123",
+  "capture_template_id": "tpl_11223344",
+  "picture_source_id": "ps_aabbccdd",
+  "color_strip_source_id": "css_11223344",
+  "output_target_id": "pt_aabbccdd",
+  "capture_template_reused": true
+}
+```
+
+`capture_template_reused` is `true` when an existing template matched the
+platform engine (no new template was created).
+
+**Rollback:** if any step fails, all entities created within the same call are
+deleted in reverse order so no orphans remain.  The pre-existing device and any
+reused template are never deleted.  Entity "created" events are emitted only
+after the full chain succeeds, so a rollback never produces ghost UI cards.
+
 ## Web UI & PWA

 App-level routes served by FastAPI (not under `/api/v1`).
@@ -54,7 +54,48 @@ When you attach a device, a default calibration is created:
 }
 ```

-## Custom Calibration
+## Automatic Calibration
+
+The easiest way to calibrate your strip is the **Auto-Calibrate** wizard, available directly
+from the calibration modal. No LED counting required — just answer three questions and tap four
+corners.
+
+### Prerequisites
+
+- A **Color Strip Source** (not a device-only target) associated with the strip.
+- A **WLED device** connected and reachable by LedGrab.
+
+### How to Start
+
+1. Open the **Calibration** modal for your strip source (pencil icon → Calibration tab).
+2. Click the **Auto-calibrate** button in the modal footer.
+3. Follow the five-step wizard.
+
+### Wizard Steps
+
+| Step | What you do |
+| ---- | ----------- |
+| 1. Device | Select the WLED device that drives the strip. |
+| 2. Start corner | LED #0 lights up on your device. Tap the corner where you see it. |
+| 3. Direction | Sweep a few LEDs light up in sequence. Tap the direction they move. |
+| 4. Mark corners | Use the step buttons to sweep to each remaining corner, then tap **Mark corner**. Repeat for all 4 corners. |
+| 5. Preview & Save | Review the detected layout (start position, direction, LED counts per edge). Click **Save** to apply. |
+
+### What Happens in the Background
+
+- A calibration session takes exclusive control of the device for the duration of the wizard;
+  any previously running effect is paused and automatically restored when the wizard exits
+  (whether by saving, cancelling, or closing the modal).
+- The solved `CalibrationConfig` is written directly to the Color Strip Source via the existing
+  PUT endpoint and takes effect immediately (no restart needed).
+
+### Tips
+
+- If LED #0 is hard to see, reduce ambient lighting briefly.
+- The wizard works in the browser — desktop and Android TV app both supported.
+- If you make a mistake in step 4, use **Step back** to re-mark the previous corner.
+
+## Manual Calibration

 ### Step 1: Identify Your LED Layout

@@ -28,6 +28,18 @@ API key authentication via Bearer token in the `Authorization` header (`Authoriz
 - When `api_keys` is **set**: a valid Bearer token is required from every client (loopback included).
 - `require_authenticated()` rejects even loopback-anonymous callers on sensitive endpoints (e.g. backup download, secret reveal).

+## Activity / Audit Log
+
+Persistent, queryable audit log of meaningful actions (auth, device, entity CRUD, capture, system), surfaced in the WebUI (Activity tab + Dashboard widget + Settings retention panel).
+
+- **Storage is NOT a `BaseSqliteStore`.** `storage/activity_log_repository.py` is a purpose-built repository over a dedicated indexed `activity_log` table (migration `002_add_activity_log`) — query-on-demand with **keyset pagination** (`seq` cursor), never load-all-into-memory. Don't route it through the entity-store pattern.
+- **Recording.** `core/activity_log/recorder.py` (`ActivityRecorder`) is best-effort (never raises into the audited action) and **thread-safe** (inline on the event loop; `loop.call_soon_threadsafe` from non-loop threads, e.g. zeroconf discovery). It persists the entry **and** fires an `activity_logged` realtime event. Actor comes from the `current_actor` `ContextVar` (set in `verify_api_key`), default `"system"`.
+- **Entity CRUD is auto-audited** via the `fire_entity_event()` choke point in `api/dependencies.py` — every create/update/delete already calls it. **Delete handlers must pass `entity_name`** (the entity is gone by record time). Non-entity events use explicit `recorder.record(...)` (get it via `get_activity_recorder()` DI or `get_module_recorder()` for engine/thread sites).
+- **Never log secrets.** API-key tokens are never recorded. Wrap any untrusted/attacker-controllable string (mDNS names, headers, user-authored names) with `sanitize_display()` (`core/activity_log/sanitize.py`) before it enters a `message`/`metadata` field. Per-IP throttle bounds auth-failure audit writes.
+- **Adding a new audited event:** pick a dotted `action` (e.g. `"thing.created"`), call the recorder; for it to render localized in the UI, add `activity_log.msg.<action>` to all three `static/locales/*.json` (the frontend `localizeMessage()` maps action→template; falls back to the server `message`). Entity-type labels live under `activity_log.entity_type.<type>`.
+- **Adding a new realtime event type** (`pm.fire_event({"type": ...})`): add it to `_ALLOWED_SERVER_EVENT_TYPES` in `static/js/core/events-ws.ts` AND keep `tests/test_events_ws_parity.py` green.
+- **Retention + API.** `core/activity_log/retention.py` prunes by `max_days` + `max_entries` (settings persisted via `db.set_setting("activity_log")`); the recorder's `enabled` flag is rehydrated from those settings on startup. REST in `api/routes/activity_log.py`: `GET /activity-log` (list, `AuthRequired`), `GET /export` (CSV/JSON stream — `require_authenticated`; chunked keyset so it never holds the DB lock across the stream; CSV formula-injection guarded), `GET|PUT /settings` (PUT is `require_authenticated`), `DELETE` (clear — `require_authenticated`, self-audited). The table is covered by the existing whole-DB backup (no `STORE_MAP` change needed).
+
 ## Common Tasks

 ### Adding a new API endpoint
@@ -17,10 +17,19 @@ auth:
  # with a secret you generated yourself (e.g. `openssl rand -hex 32`).
  # Do NOT ship a hard-coded key here — a publicly-known token grants full
  # LAN access to anyone on the network.
-  api_keys: {}
+  api_keys:
+    default: "development-key-change-in-production"
  # api_keys:
  #   my-client: "replace-with-output-of-openssl-rand-hex-32"

+  # Expose the interactive API docs (/docs, /redoc, /openapi.json) WITHOUT a
+  # Bearer token so they can be opened directly in a browser. When true, this
+  # applies to loopback AND LAN clients. Only the API *surface* (route paths +
+  # parameter schemas) is exposed — calling an endpoint from Swagger still
+  # requires the token via its "Authorize" button, and every other route stays
+  # protected. Leave false unless you want browsable docs on your network.
+  expose_docs: false
+
 # Storage paths default to ./data relative to the server's working directory.
 # Set LEDGRAB_DATA_DIR in the environment to point at a different data root
 # (the whole dir — both the database and assets), or uncomment the block
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"

 [project]
 name = "ledgrab"
-version = "0.8.1"
+version = "0.9.0"
 description = "Ambient lighting system that captures screen content and drives LED strips in real time"
 authors = [
    {name = "Alexei Dolgolyov", email = "dolgolyov.alexei@gmail.com"}
@@ -18,6 +18,7 @@ from .routes.audio_templates import router as audio_templates_router
 from .routes.value_sources import router as value_sources_router
 from .routes.automations import router as automations_router
 from .routes.scene_presets import router as scene_presets_router
+from .routes.scene_playlists import router as scene_playlists_router
 from .routes.webhooks import router as webhooks_router
 from .routes.sync_clocks import router as sync_clocks_router
 from .routes.color_strip_processing import router as cspt_router
@@ -35,6 +36,9 @@ from .routes.pattern_templates import router as pattern_templates_router
 from .routes.preferences import router as preferences_router
 from .routes.snapshot import router as snapshot_router
 from .routes.graph import router as graph_router
+from .routes.calibration import router as calibration_router
+from .routes.setup import router as setup_router
+from .routes.activity_log import router as activity_log_router

 router = APIRouter()
 router.include_router(system_router)
@@ -53,6 +57,7 @@ router.include_router(output_targets_router)
 router.include_router(output_targets_control_router)
 router.include_router(automations_router)
 router.include_router(scene_presets_router)
+router.include_router(scene_playlists_router)
 router.include_router(webhooks_router)
 router.include_router(sync_clocks_router)
 router.include_router(cspt_router)
@@ -70,5 +75,8 @@ router.include_router(pattern_templates_router)
 router.include_router(preferences_router)
 router.include_router(snapshot_router)
 router.include_router(graph_router)
+router.include_router(calibration_router)
+router.include_router(setup_router)
+router.include_router(activity_log_router)

 __all__ = ["router"]
@@ -3,18 +3,137 @@
 import asyncio
 import json
 import secrets
+import threading
+import time
+from collections import OrderedDict
 from typing import Annotated
+from urllib.parse import urlparse

 from fastapi import Depends, HTTPException, Request, Security, status
 from fastapi.security import HTTPAuthorizationCredentials, HTTPBearer
 from starlette.websockets import WebSocket, WebSocketDisconnect

 from ledgrab.config import get_config
+from ledgrab.core.activity_log.context import current_actor
+from ledgrab.core.activity_log.sanitize import sanitize_display
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.utils import get_logger
 from ledgrab.utils.net_classify import is_loopback as _classify_is_loopback

 logger = get_logger(__name__)

+# ── Auth-failure audit throttle (H3) ───────────────────────────────────────
+#
+# Unauthenticated callers can hammer any auth path; without a recording
+# throttle each attempt would write one SQLite row AND broadcast one WS event,
+# providing a cheap disk/broadcast amplification vector.
+#
+# Mitigation: record at most one ``auth.rejected`` audit entry per client IP
+# per _AUTH_RECORD_WINDOW seconds.  The auth decision (401) is NEVER
+# suppressed — only the *audit recording* is de-duplicated.
+#
+# Memory safety: the throttle dict is capped at _AUTH_THROTTLE_HARD_CAP
+# entries.  When the cap is exceeded the oldest-inserted IP is evicted in O(1)
+# so the dict stays bounded regardless of the number of distinct source IPs an
+# attacker can forge.
+#
+# Thread safety: the throttle dict is guarded by ``_auth_record_lock`` (mirrors
+# ``_auth_fail_lock`` in routes/game_integration) so the compound
+# read/evict/insert is atomic. The HTTP auth dependency runs on the event loop
+# (``verify_api_key`` is async), but ``_record_auth_failure`` is reached from
+# both the HTTP and WebSocket auth paths and must remain safe if ever called
+# from a background thread — the lock is uncontended on the loop, so it costs
+# nothing while preventing a KeyError / "dict changed size" from ever turning
+# an intended 401 into a 500.
+
+_AUTH_RECORD_WINDOW: float = 10.0  # seconds — one record per IP per window
+_AUTH_THROTTLE_HARD_CAP: int = 512  # max IPs tracked simultaneously
+
+# ip -> monotonic timestamp of last *recorded* auth.rejected entry.
+# OrderedDict so the oldest insertion can be evicted in O(1) via popitem.
+_auth_record_last: "OrderedDict[str, float]" = OrderedDict()
+_auth_record_lock = threading.Lock()
+
+
+def _should_record_auth_failure(client_ip: str) -> bool:
+    """Return True when an ``auth.rejected`` record should be written for *client_ip*.
+
+    Suppresses duplicates within _AUTH_RECORD_WINDOW seconds.  Evicts the
+    oldest entry when the dict exceeds _AUTH_THROTTLE_HARD_CAP to prevent
+    unbounded memory growth under IP-spray attacks.
+
+    Thread-safe: the entire read/evict/insert is performed under
+    ``_auth_record_lock`` so concurrent threadpool workers cannot corrupt the
+    dict or raise mid-mutation.
+    """
+    now = time.monotonic()
+    with _auth_record_lock:
+        last = _auth_record_last.get(client_ip)
+        if last is not None and (now - last) < _AUTH_RECORD_WINDOW:
+            return False  # suppress: within the de-dup window
+
+        # Enforce hard cap before inserting: evict the oldest entry in O(1).
+        if len(_auth_record_last) >= _AUTH_THROTTLE_HARD_CAP:
+            _auth_record_last.popitem(last=False)
+
+        # Refresh recency: move/insert this IP to the most-recent end so the
+        # popitem(last=False) above always drops a genuinely old entry.
+        _auth_record_last[client_ip] = now
+        _auth_record_last.move_to_end(client_ip)
+        return True
+
+
+def _record_auth_failure(reason: str, client_host: str | None) -> None:
+    """Best-effort: record an auth failure audit entry (never raises).
+
+    SECURITY: the attempted token is NEVER passed here; only the reason and
+    the caller's IP/label are recorded.
+
+    THROTTLE: at most one ``auth.rejected`` record is written per client IP
+    per _AUTH_RECORD_WINDOW seconds to prevent disk/WS-broadcast amplification
+    DoS.  The 401 response is always returned regardless.
+
+    The whole body is wrapped so an audit-path failure can never convert an
+    intended 401 into a 500 (honors the "never raises" contract).
+    """
+    try:
+        if not _should_record_auth_failure(client_host or "unknown"):
+            return  # throttled — drop duplicate recording for this IP/window
+
+        from ledgrab.core.activity_log.recorder import get_module_recorder
+
+        rec = get_module_recorder()
+        if rec is None:
+            return
+        rec.record(
+            category=ActivityCategory.AUTH,
+            action="auth.rejected",
+            severity=ActivitySeverity.WARNING,
+            actor="anonymous",
+            message=f"Authentication failed: {reason}",
+            metadata={"reason": reason, "client": client_host or "unknown"},
+        )
+    except Exception as exc:  # never raise into the auth path
+        logger.warning("auth-failure audit recording failed: %s", exc)
+
+
+def _record_ws_auth_success(label: str, client_host: str | None) -> None:
+    """Best-effort: record a successful WebSocket session establishment."""
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is None:
+        return
+    rec.record(
+        category=ActivityCategory.AUTH,
+        action="auth.ws_connected",
+        severity=ActivitySeverity.INFO,
+        actor=label,
+        message=f"WebSocket session established by '{label}'",
+        metadata={"client": client_host or "unknown"},
+    )
+
+
 # Security scheme for Bearer token
 security = HTTPBearer(auto_error=False)

@@ -49,7 +168,7 @@ def _is_loopback(host: str | None) -> bool:
    return _classify_is_loopback(host)


-def verify_api_key(
+async def verify_api_key(
    request: Request,
    credentials: Annotated[HTTPAuthorizationCredentials | None, Security(security)],
 ) -> str:
@@ -63,6 +182,13 @@ def verify_api_key(
      LAN access requires an API key).
    - When API keys ARE configured, valid Bearer credentials are required.

+    This is an ``async`` dependency on purpose: token comparison is CPU-trivial,
+    and an async dependency runs in the SAME task/context as the route handler,
+    so ``current_actor.set(...)`` below is visible to ``ActivityRecorder`` when
+    the handler later records an entity event. A sync dependency would run in a
+    throwaway threadpool context and the actor mutation would be discarded,
+    attributing every audited action to "system".
+
    Args:
        request: incoming request (used to read client host)
        credentials: HTTP authorization credentials
@@ -81,10 +207,12 @@ def verify_api_key(
        # No keys configured — allow loopback only.
        if _is_loopback(client_host):
            request.state.auth_label = "anonymous"
+            current_actor.set("anonymous")
            return "anonymous"
        # Allow caller to authenticate explicitly even without configured keys?
        # No — there are no keys to compare against. Reject.
        logger.warning("Rejected LAN request from %s: no API key configured", client_host)
+        _record_auth_failure("LAN access rejected: no API key configured", client_host)
        raise HTTPException(
            status_code=status.HTTP_401_UNAUTHORIZED,
            detail=(
@@ -97,24 +225,32 @@ def verify_api_key(
    # Check if credentials are provided
    if not credentials:
        logger.warning("Request missing Authorization header")
+        _record_auth_failure("missing Bearer token", client_host)
        raise HTTPException(
            status_code=status.HTTP_401_UNAUTHORIZED,
            detail="Missing API key - authentication is required",
            headers={"WWW-Authenticate": "Bearer"},
        )

-    # Extract token
+    # Extract token — NEVER log or record the token value itself.
    token = credentials.credentials

-    # Find matching key and return its label using constant-time comparison
+    # Find matching key and return its label using constant-time comparison.
+    # Compare UTF-8 byte encodings: secrets.compare_digest raises TypeError on
+    # non-ASCII str (an attacker can put 0x80-0xFF in the Authorization header,
+    # which Starlette latin-1-decodes to a non-ASCII str). Byte comparison is
+    # well-defined for any input and preserves constant-time behavior, so a
+    # bad/non-ASCII token cleanly falls through to the 401 below instead of 500.
+    token_b = (token or "").encode("utf-8")
    authenticated_as = None
    for label, api_key in config.auth.api_keys.items():
-        if secrets.compare_digest(token, api_key):
+        if secrets.compare_digest(token_b, api_key.encode("utf-8")):
            authenticated_as = label
            break

    if not authenticated_as:
        logger.warning("Invalid API key attempt")
+        _record_auth_failure("invalid Bearer token", client_host)
        raise HTTPException(
            status_code=status.HTTP_401_UNAUTHORIZED,
            detail="Invalid API key",
@@ -127,6 +263,9 @@ def verify_api_key(
    # Stash the friendly label so the access-log middleware can attribute the
    # request to a client without re-running the token comparison.
    request.state.auth_label = authenticated_as
+    # Set the actor ContextVar so ActivityRecorder can resolve it without
+    # threading it through every call site.
+    current_actor.set(authenticated_as)
    return authenticated_as


@@ -135,6 +274,31 @@ def verify_api_key(
 AuthRequired = Annotated[str, Depends(verify_api_key)]


+async def verify_docs_access(
+    request: Request,
+    credentials: Annotated[HTTPAuthorizationCredentials | None, Security(security)],
+) -> str:
+    """Auth gate for the OpenAPI docs routes (/docs, /redoc, /openapi.json).
+
+    When ``auth.expose_docs`` is True, the docs pages load anonymously from any
+    client (loopback and LAN) so they can be viewed in a browser without a
+    Bearer token. Only the API *surface* is exposed this way — every other
+    endpoint still goes through :func:`verify_api_key`.
+
+    When ``auth.expose_docs`` is False (default), this delegates to
+    :func:`verify_api_key`, so docs require a token exactly like the rest of
+    the API.
+    """
+    if get_config().auth.expose_docs:
+        request.state.auth_label = "anonymous-docs"
+        return "anonymous-docs"
+    return await verify_api_key(request, credentials)
+
+
+# Dependency for the OpenAPI docs routes — relaxed when auth.expose_docs is set
+DocsAccess = Annotated[str, Depends(verify_docs_access)]
+
+
 def require_authenticated(label: str) -> None:
    """Reject the anonymous (loopback) auth label.

@@ -190,12 +354,30 @@ async def accept_and_authenticate_ws(websocket: WebSocket, timeout: float = 3.0)
    # a strong signal even before the token check. Non-browser clients
    # legitimately omit Origin; those fall through to the auth handshake.
    config = get_config()
+    client_host = websocket.client.host if websocket.client else None
    origin = websocket.headers.get("origin")
    if not _is_origin_allowed(origin, config.server.cors_origins):
        logger.warning(
            "Rejected WebSocket from origin %r (not in cors_origins)",
            origin,
        )
+        # Sanitize first so urlparse does not choke on control chars / ANSI / NUL
+        # embedded by an attacker in the Origin header (e.g. \n triggers IPv6 parse
+        # error in Python's urlsplit on malformed netloc).
+        _safe_origin_raw = sanitize_display(origin) if origin else ""
+        try:
+            _netloc = urlparse(_safe_origin_raw).netloc if _safe_origin_raw else ""
+        except ValueError:
+            # Malformed IPv6 addresses (e.g. "http://[::1" without closing "]")
+            # cause urlparse to raise ValueError. Fall back to "unknown" — do NOT
+            # fall back to the raw origin string, which could carry query params
+            # or path components containing secrets.
+            _netloc = ""
+        _safe_origin = sanitize_display(_netloc or "unknown")
+        _record_auth_failure(
+            f"WebSocket origin rejected: {_safe_origin!r}",
+            client_host,
+        )
        try:
            await websocket.close(code=WS_ORIGIN_CLOSE_CODE)
        except _WS_SEND_BENIGN_EXC:
@@ -210,6 +392,7 @@ async def accept_and_authenticate_ws(websocket: WebSocket, timeout: float = 3.0)
        except _WS_SEND_BENIGN_EXC:
            pass
        return None
+    _record_ws_auth_success(label, client_host)
    return label


@@ -217,12 +400,19 @@ async def accept_and_authenticate_ws(websocket: WebSocket, timeout: float = 3.0)


 def _match_api_key(token: str) -> str | None:
-    """Return the label matching *token* using constant-time comparison, or None."""
+    """Return the label matching *token* using constant-time comparison, or None.
+
+    Compares UTF-8 byte encodings so a non-ASCII token (a JSON string in the WS
+    auth message trivially carries non-ASCII) cannot raise TypeError out of
+    ``secrets.compare_digest`` — it simply fails to match and yields a clean
+    ``auth_error`` instead of crashing the handler.
+    """
    config = get_config()
    if not token:
        return None
+    token_b = token.encode("utf-8")
    for label, api_key in config.auth.api_keys.items():
-        if secrets.compare_digest(token, api_key):
+        if secrets.compare_digest(token_b, api_key.encode("utf-8")):
            return label
    return None

@@ -275,6 +465,7 @@ async def verify_ws_auth(
                return None
            return "anonymous"
        logger.warning("WebSocket auth timeout after %.1fs from %s", timeout, client_host)
+        _record_auth_failure("WebSocket auth timeout", client_host)
        try:
            await websocket.send_json({"type": "auth_error", "reason": "auth timeout"})
        except _WS_SEND_BENIGN_EXC:
@@ -332,6 +523,7 @@ async def verify_ws_auth(
            await websocket.send_json({"type": "auth_ok"})
            return "anonymous"
        logger.warning("Rejected LAN WebSocket from %s: no API key configured", client_host)
+        _record_auth_failure("LAN WebSocket rejected: no API key configured", client_host)
        try:
            await websocket.send_json(
                {
@@ -343,10 +535,11 @@ async def verify_ws_auth(
            pass
        return None

-    # Keys configured: require a matching token.
+    # Keys configured: require a matching token.  NEVER log the token value.
    label = _match_api_key(token or "")
    if not label:
        logger.warning("Invalid WebSocket auth attempt from %s", client_host)
+        _record_auth_failure("invalid WebSocket token", client_host)
        try:
            await websocket.send_json({"type": "auth_error", "reason": "invalid token"})
        except _WS_SEND_BENIGN_EXC:
@@ -19,6 +19,7 @@ from ledgrab.storage.audio_template_store import AudioTemplateStore
 from ledgrab.storage.value_source_store import ValueSourceStore
 from ledgrab.storage.automation_store import AutomationStore
 from ledgrab.storage.scene_preset_store import ScenePresetStore
+from ledgrab.storage.scene_playlist_store import ScenePlaylistStore
 from ledgrab.storage.sync_clock_store import SyncClockStore
 from ledgrab.storage.color_strip_processing_template_store import (
    ColorStripProcessingTemplateStore,
@@ -27,6 +28,7 @@ from ledgrab.storage.gradient_store import GradientStore
 from ledgrab.storage.weather_source_store import WeatherSourceStore
 from ledgrab.storage.asset_store import AssetStore
 from ledgrab.core.automations.automation_engine import AutomationEngine
+from ledgrab.core.scenes.playlist_engine import PlaylistEngine
 from ledgrab.core.weather.weather_manager import WeatherManager
 from ledgrab.core.backup.auto_backup import AutoBackupEngine
 from ledgrab.core.processing.sync_clock_manager import SyncClockManager
@@ -35,11 +37,17 @@ from ledgrab.storage.home_assistant_store import HomeAssistantStore
 from ledgrab.core.home_assistant.ha_manager import HomeAssistantManager
 from ledgrab.storage.game_integration_store import GameIntegrationStore
 from ledgrab.core.game_integration.event_bus import GameEventBus
+from ledgrab.core.game_integration.lol_poll_manager import LoLPollManager
 from ledgrab.storage.mqtt_source_store import MQTTSourceStore
 from ledgrab.core.mqtt.mqtt_manager import MQTTManager
 from ledgrab.storage.http_endpoint_store import HTTPEndpointStore
 from ledgrab.storage.audio_processing_template_store import AudioProcessingTemplateStore
 from ledgrab.storage.pattern_template_store import PatternTemplateStore
+from ledgrab.core.activity_log.recorder import ActivityRecorder, get_module_recorder
+from ledgrab.core.activity_log.retention import ActivityLogRetentionEngine
+from ledgrab.core.activity_log.sanitize import sanitize_display
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+from ledgrab.storage.activity_log_repository import ActivityLogRepository

 T = TypeVar("T")

@@ -110,6 +118,14 @@ def get_automation_engine() -> AutomationEngine:
    return _get("automation_engine", "Automation engine")


+def get_scene_playlist_store() -> ScenePlaylistStore:
+    return _get("scene_playlist_store", "Scene playlist store")
+
+
+def get_playlist_engine() -> PlaylistEngine:
+    return _get("playlist_engine", "Playlist engine")
+
+
 def get_auto_backup_engine() -> AutoBackupEngine:
    return _get("auto_backup_engine", "Auto-backup engine")

@@ -158,6 +174,15 @@ def get_game_event_bus() -> GameEventBus:
    return _get("game_event_bus", "Game event bus")


+def get_lol_poll_manager() -> LoLPollManager | None:
+    """LoL poll manager, or None if not wired (e.g. minimal test harnesses).
+
+    Polling is a best-effort background feature, so callers guard on None
+    rather than 500-ing a CRUD request when the manager is absent.
+    """
+    return _deps.get("lol_poll_manager")
+
+
 def get_mqtt_store() -> MQTTSourceStore:
    return _get("mqtt_store", "MQTT source store")

@@ -186,16 +211,87 @@ def get_update_service() -> UpdateService:
    return _get("update_service", "Update service")


+def get_activity_recorder() -> ActivityRecorder:
+    return _get("activity_recorder", "Activity recorder")
+
+
+def get_activity_log_repo() -> ActivityLogRepository:
+    return _get("activity_log_repo", "Activity log repository")
+
+
+def get_activity_log_retention_engine() -> ActivityLogRetentionEngine:
+    return _get("activity_log_retention_engine", "Activity log retention engine")
+
+
 # ── Event helper ────────────────────────────────────────────────────────


-def fire_entity_event(entity_type: str, action: str, entity_id: str) -> None:
-    """Fire an entity_changed event via the ProcessorManager event bus.
+# entity_type → (_deps key, store method name) for human-name resolution.
+# Module-level constant: built once at import rather than per audited mutation
+# (``_resolve_entity_name`` is the create/update audit choke point).
+_STORE_LOOKUP: dict[str, tuple[str, str]] = {
+    "output_target": ("output_target_store", "get_target"),
+    "device": ("device_store", "get_device"),
+    "picture_source": ("picture_source_store", "get_source"),
+    "audio_source": ("audio_source_store", "get_source"),
+    "color_strip_source": ("color_strip_store", "get_source"),
+    "template": ("template_store", "get_template"),
+    "capture_template": ("template_store", "get_template"),
+    "pp_template": ("pp_template_store", "get_template"),
+    "automation": ("automation_store", "get_automation"),
+    "scene_preset": ("scene_preset_store", "get_preset"),
+    "scene_playlist": ("scene_playlist_store", "get_playlist"),
+    "sync_clock": ("sync_clock_store", "get_clock"),
+    "gradient": ("gradient_store", "get_gradient"),
+    "audio_template": ("audio_template_store", "get_template"),
+    "value_source": ("value_source_store", "get_source"),
+    "cspt": ("cspt_store", "get_template"),
+    "audio_processing_template": ("audio_processing_template_store", "get_template"),
+    "pattern_template": ("pattern_template_store", "get_template"),
+    "home_assistant_source": ("ha_store", "get_source"),
+    "mqtt_source": ("mqtt_store", "get_source"),
+    "http_endpoint": ("http_endpoint_store", "get_endpoint"),
+}
+
+
+def _resolve_entity_name(entity_type: str, entity_id: str) -> str | None:
+    """Best-effort: look up a human name for *entity_id* from the matching store.
+
+    Returns ``None`` when the store is missing, the entity is gone, or any
+    exception occurs (e.g. during delete the entity may have just been removed).
+    """
+    entry = _STORE_LOOKUP.get(entity_type)
+    if entry is None:
+        return None
+    store_key, method_name = entry
+    store = _deps.get(store_key)
+    if store is None:
+        return None
+    try:
+        obj = getattr(store, method_name)(entity_id)
+        if obj is not None:
+            return getattr(obj, "name", None)
+    except Exception:
+        pass
+    return None
+
+
+def fire_entity_event(
+    entity_type: str,
+    action: str,
+    entity_id: str,
+    entity_name: str | None = None,
+) -> None:
+    """Fire an entity_changed event via the ProcessorManager event bus and
+    record an audit entry.

    Args:
        entity_type: e.g. "device", "output_target", "color_strip_source"
        action: "created", "updated", or "deleted"
        entity_id: The entity's unique ID
+        entity_name: Human-readable name.  For deletes: **must** be passed
+            explicitly (entity is already gone when we get here).
+            For create/update: resolved from the store when not supplied.
    """
    pm = _deps.get("processor_manager")
    if pm is not None:
@@ -208,6 +304,38 @@ def fire_entity_event(entity_type: str, action: str, entity_id: str) -> None:
            }
        )

+    # ── Audit record (best-effort) ──────────────────────────────────────────
+    rec = get_module_recorder()
+    if rec is None:
+        return
+
+    # Resolve name when not explicitly provided (create / update paths).
+    # For deleted: entity already gone — rely on the explicitly passed name.
+    resolved_name = entity_name
+    if resolved_name is None and action != "deleted":
+        resolved_name = _resolve_entity_name(entity_type, entity_id)
+
+    # Build a concise human message.
+    # Sanitize the display name before interpolating into the free-text message
+    # (user-authored names hit the CSV/export trust surface).
+    safe_display_name = sanitize_display(resolved_name) if resolved_name else None
+    display_name = f"'{safe_display_name}'" if safe_display_name else entity_id
+    action_word = {"created": "created", "updated": "updated", "deleted": "deleted"}.get(
+        action, action
+    )
+    entity_label = entity_type.replace("_", " ")
+    message = f"{entity_label.capitalize()} {display_name} {action_word}"
+
+    rec.record(
+        category=ActivityCategory.ENTITY,
+        action=f"entity.{action}",
+        severity=ActivitySeverity.INFO,
+        entity_type=entity_type,
+        entity_id=entity_id,
+        entity_name=sanitize_display(resolved_name) if resolved_name else None,
+        message=message,
+    )
+

 # ── Initialization ──────────────────────────────────────────────────────

@@ -226,7 +354,9 @@ def init_dependencies(
    value_source_store: ValueSourceStore | None = None,
    automation_store: AutomationStore | None = None,
    scene_preset_store: ScenePresetStore | None = None,
+    scene_playlist_store: ScenePlaylistStore | None = None,
    automation_engine: AutomationEngine | None = None,
+    playlist_engine: PlaylistEngine | None = None,
    auto_backup_engine: AutoBackupEngine | None = None,
    sync_clock_store: SyncClockStore | None = None,
    sync_clock_manager: SyncClockManager | None = None,
@@ -240,11 +370,15 @@ def init_dependencies(
    ha_manager: HomeAssistantManager | None = None,
    game_integration_store: GameIntegrationStore | None = None,
    game_event_bus: GameEventBus | None = None,
+    lol_poll_manager: LoLPollManager | None = None,
    mqtt_store: MQTTSourceStore | None = None,
    mqtt_manager: MQTTManager | None = None,
    http_endpoint_store: HTTPEndpointStore | None = None,
    audio_processing_template_store: AudioProcessingTemplateStore | None = None,
    pattern_template_store: PatternTemplateStore | None = None,
+    activity_recorder: ActivityRecorder | None = None,
+    activity_log_repo: ActivityLogRepository | None = None,
+    activity_log_retention_engine: ActivityLogRetentionEngine | None = None,
 ):
    """Initialize global dependencies."""
    _deps.update(
@@ -262,7 +396,9 @@ def init_dependencies(
            "value_source_store": value_source_store,
            "automation_store": automation_store,
            "scene_preset_store": scene_preset_store,
+            "scene_playlist_store": scene_playlist_store,
            "automation_engine": automation_engine,
+            "playlist_engine": playlist_engine,
            "auto_backup_engine": auto_backup_engine,
            "sync_clock_store": sync_clock_store,
            "sync_clock_manager": sync_clock_manager,
@@ -276,10 +412,14 @@ def init_dependencies(
            "ha_manager": ha_manager,
            "game_integration_store": game_integration_store,
            "game_event_bus": game_event_bus,
+            "lol_poll_manager": lol_poll_manager,
            "mqtt_store": mqtt_store,
            "mqtt_manager": mqtt_manager,
            "http_endpoint_store": http_endpoint_store,
            "audio_processing_template_store": audio_processing_template_store,
            "pattern_template_store": pattern_template_store,
+            "activity_recorder": activity_recorder,
+            "activity_log_repo": activity_log_repo,
+            "activity_log_retention_engine": activity_log_retention_engine,
        }
    )
@@ -0,0 +1,468 @@
+"""Activity-log REST API — query / filter / export / settings / clear.
+
+Endpoints
+---------
+GET    /api/v1/activity-log           List (filterable, keyset-paginated)
+GET    /api/v1/activity-log/export    Streaming CSV or JSON export
+GET    /api/v1/activity-log/settings  Retention settings
+PUT    /api/v1/activity-log/settings  Update retention settings (requires non-anonymous auth)
+DELETE /api/v1/activity-log           Clear all entries (requires non-anonymous auth)
+
+Auth posture
+------------
+- List + read settings (``GET``): ``AuthRequired`` (loopback-anonymous is fine).
+- Export, update settings (``PUT``), and clear: ``require_authenticated()``
+  (loopback-anonymous is rejected; mirrors the backup download / secret-reveal
+  pattern from ``backup.py``).  Updating settings can disable auditing or prune
+  the trail, so it is gated like the destructive clear.
+
+CSV injection
+-------------
+Cells that begin with =, +, -, @, TAB, or CR can trigger formula execution in
+spreadsheet apps (OWASP Formula Injection).  ``_csv_safe`` prefixes any such cell
+with a single quote so formulas are inert.  Fields already go through
+``sanitize_display`` in Phase 3 instrumentation, but the CSV writer applies its
+own guard as defence-in-depth.
+
+Export generator + lock
+-----------------------
+``repo.iter_export()`` fetches rows in bounded batches, holding the DB ``_lock``
+only around each batch fetch and releasing it before yielding — so a slow or
+stalled client never blocks other DB operations.  The ``StreamingResponse``
+generator is wrapped in a ``try/finally`` block so the batch generator is closed
+even when the client disconnects mid-stream.
+"""
+
+from __future__ import annotations
+
+import csv
+import io
+import json
+from datetime import datetime, timezone
+from typing import Annotated, Iterator
+
+from fastapi import APIRouter, Depends, Query
+from fastapi.responses import StreamingResponse
+
+from ledgrab.api.auth import AuthRequired, require_authenticated
+from ledgrab.api.dependencies import (
+    get_activity_log_repo,
+    get_activity_log_retention_engine,
+    get_activity_recorder,
+)
+from ledgrab.api.schemas.activity_log import (
+    ActivityLogPageResponse,
+    ActivityLogSettingsResponse,
+    UpdateActivityLogSettingsRequest,
+)
+from ledgrab.core.activity_log.recorder import ActivityRecorder, entry_to_dict
+from ledgrab.core.activity_log.retention import ActivityLogRetentionEngine
+from ledgrab.core.activity_log.sanitize import sanitize_display
+from ledgrab.storage.activity_log import ActivityCategory, ActivityLogFilters, ActivitySeverity
+from ledgrab.storage.activity_log_repository import ActivityLogRepository
+
+router = APIRouter(prefix="/api/v1/activity-log", tags=["Activity Log"])
+
+# Hard cap on the per-request limit to prevent runaway queries.
+_MAX_LIMIT = 200
+_DEFAULT_LIMIT = 50
+
+# Bounds on the text filter params so a multi-KB ``q`` / actor / entity filter
+# can't enlarge the LIKE pattern and bound params per page (FastAPI returns 422
+# on overflow). The free-text ``q`` gets a larger budget than the id filters.
+_MAX_TEXT_FILTER = 256
+_MAX_ID_FILTER = 128
+
+# CSV export columns (matches entry_to_dict key order)
+_CSV_COLUMNS = [
+    "id",
+    "ts",
+    "category",
+    "action",
+    "severity",
+    "actor",
+    "entity_type",
+    "entity_id",
+    "entity_name",
+    "message",
+    "metadata",
+]
+
+# Characters that trigger formula injection in spreadsheet apps (OWASP).
+# Leading TAB and CR are also recognised triggers by Excel / Google Sheets.
+_FORMULA_PREFIXES = ("=", "+", "-", "@", "\t", "\r")
+
+# Cap for export-cell sanitization. Effectively no truncation (a single audit
+# field never approaches this) — we reuse sanitize_display only to strip
+# NUL/control/ANSI from CSV cells, not to shorten them.
+_EXPORT_CELL_MAXLEN = 1_000_000
+
+
+def _csv_safe(value: str) -> str:
+    """Prefix formula-injection triggers with a literal single-quote.
+
+    A cell starting with =, +, -, or @ can execute as a formula in Excel /
+    Google Sheets.  OWASP recommends prepending a single quote to neutralise it.
+    """
+    if value and value[0] in _FORMULA_PREFIXES:
+        return "'" + value
+    return value
+
+
+def _redact_for_anon(entry_dict: dict, auth_label: str) -> dict:
+    """Redact the source-IP metadata for anonymous (loopback) callers.
+
+    The streaming export is gated by ``require_authenticated`` precisely because
+    the log can contain client IPs (e.g. ``auth.rejected`` / ``auth.ws_connected``
+    store ``metadata.client``). The list endpoint allows loopback-anonymous
+    callers, so to keep the posture consistent we mask that one field for the
+    ``"anonymous"`` label rather than handing it back what export withholds.
+    """
+    if auth_label != "anonymous":
+        return entry_dict
+    meta = entry_dict.get("metadata")
+    if isinstance(meta, dict) and "client" in meta:
+        return {**entry_dict, "metadata": {**meta, "client": "[redacted]"}}
+    return entry_dict
+
+
+def _build_filters(
+    categories: list[str] | None,
+    severities: list[str] | None,
+    actor: str | None,
+    entity_type: str | None,
+    entity_id: str | None,
+    since: datetime | None,
+    until: datetime | None,
+    q: str | None,
+) -> ActivityLogFilters:
+    """Assemble an ``ActivityLogFilters`` dataclass from query parameters."""
+    return ActivityLogFilters(
+        categories=categories or None,
+        severities=severities or None,
+        actor=actor or None,
+        entity_type=entity_type or None,
+        entity_id=entity_id or None,
+        since=since,
+        until=until,
+        message_like=q or None,
+    )
+
+
+# ---------------------------------------------------------------------------
+# GET /api/v1/activity-log  — list
+# ---------------------------------------------------------------------------
+
+
+@router.get("", response_model=ActivityLogPageResponse, summary="List activity-log entries")
+def list_activity_log(
+    auth: AuthRequired,
+    repo: ActivityLogRepository = Depends(get_activity_log_repo),
+    # ── Filters ────────────────────────────────────────────────────────────
+    categories: Annotated[
+        list[str] | None,
+        Query(
+            description=(
+                "Filter by category (repeatable or comma-separated). "
+                "Values: auth, device, entity, capture, system"
+            )
+        ),
+    ] = None,
+    severities: Annotated[
+        list[str] | None,
+        Query(description="Filter by severity (repeatable). Values: info, warning, error"),
+    ] = None,
+    actor: Annotated[
+        str | None,
+        Query(max_length=_MAX_ID_FILTER, description="Filter by actor label (exact match)"),
+    ] = None,
+    entity_type: Annotated[
+        str | None,
+        Query(max_length=_MAX_ID_FILTER, description="Filter by entity type (exact match)"),
+    ] = None,
+    entity_id: Annotated[
+        str | None,
+        Query(max_length=_MAX_ID_FILTER, description="Filter by entity id (exact match)"),
+    ] = None,
+    since: Annotated[
+        datetime | None,
+        Query(description="Return entries at or after this ISO-8601 datetime"),
+    ] = None,
+    until: Annotated[
+        datetime | None,
+        Query(description="Return entries at or before this ISO-8601 datetime"),
+    ] = None,
+    q: Annotated[
+        str | None,
+        Query(
+            max_length=_MAX_TEXT_FILTER,
+            description="Free-text search in the message field (substring)",
+        ),
+    ] = None,
+    # ── Pagination ─────────────────────────────────────────────────────────
+    before_seq: Annotated[
+        int | None,
+        Query(
+            description=(
+                "Keyset cursor: pass the 'next_before_seq' from the previous page "
+                "to get the following (older) page. Omit for the first (newest) page."
+            )
+        ),
+    ] = None,
+    limit: Annotated[
+        int,
+        Query(
+            ge=1,
+            le=_MAX_LIMIT,
+            description=f"Max entries per page (default {_DEFAULT_LIMIT}, max {_MAX_LIMIT})",
+        ),
+    ] = _DEFAULT_LIMIT,
+) -> ActivityLogPageResponse:
+    """Return the newest matching entries, oldest-first within the page.
+
+    Keyset pagination: the response includes ``next_before_seq`` — pass it
+    as ``before_seq`` in the next request to get the next (older) page.
+    The ``total`` field is the count of all entries matching the current
+    filters across all pages.
+    """
+    filters = _build_filters(categories, severities, actor, entity_type, entity_id, since, until, q)
+
+    # Fetch limit+1 rows to detect whether an older page exists.
+    #
+    # query() fetches DESC internally (newest-first) then reverses to ascending.
+    # With limit+1, the result is ascending: [oldest_probe, ..., newest].
+    # When we got exactly limit+1 rows, has_more is True and the probe row
+    # (index 0 — the oldest) is the extra one.  We keep the newest `limit` rows
+    # by slicing [1:], which is the actual page content for the client.
+    # When we got <= limit rows, this is the last page and all rows are included.
+    effective_limit = min(limit, _MAX_LIMIT)
+    # query_with_seq returns (seq, entry) ascending (oldest-first within page),
+    # so the seq is already in hand — no extra get_seq_for_id round-trip.
+    rows_plus = repo.query_with_seq(filters, before_seq=before_seq, limit=effective_limit + 1)
+    has_more = len(rows_plus) > effective_limit
+    # When over-fetched, drop the oldest probe row (index 0) and keep the newest.
+    rows = rows_plus[1:] if has_more else rows_plus
+
+    total = repo.count(filters)
+
+    # next_before_seq: the seq of the oldest entry on this page (rows[0]).
+    # The next request passes before_seq=X to get entries with seq < X.
+    next_before_seq: int | None = rows[0][0] if (has_more and rows) else None
+
+    return ActivityLogPageResponse(
+        entries=[_redact_for_anon(entry_to_dict(e), auth) for _seq, e in rows],  # type: ignore[arg-type]
+        next_before_seq=next_before_seq,
+        has_more=has_more,
+        total=total,
+    )
+
+
+# ---------------------------------------------------------------------------
+# GET /api/v1/activity-log/export  — streaming export (CSV or JSON)
+# ---------------------------------------------------------------------------
+
+
+def _export_csv_generator(
+    repo: ActivityLogRepository,
+    filters: ActivityLogFilters,
+) -> Iterator[bytes]:
+    """Yield UTF-8-encoded CSV chunks one row at a time.
+
+    The generator wraps ``repo.iter_export()`` in a ``try/finally`` so the DB
+    lock is released even on early client disconnect (which triggers
+    ``GeneratorExit``).
+    """
+    gen = repo.iter_export(filters)
+    try:
+        # Header
+        buf = io.StringIO()
+        writer = csv.writer(buf)
+        writer.writerow(_CSV_COLUMNS)
+        yield buf.getvalue().encode("utf-8")
+
+        for entry in gen:
+            d = entry_to_dict(entry)
+            row = []
+            for col in _CSV_COLUMNS:
+                if col == "metadata":
+                    # json.dumps escapes control chars (<0x20) as \uXXXX, so the
+                    # metadata cell can't carry raw NUL/CR/ANSI into the file.
+                    cell = json.dumps(d.get(col) or {})
+                else:
+                    # Defense-in-depth: strip NUL/control/ANSI from string cells
+                    # at the export boundary so a (current or future) un-sanitized
+                    # call site can't leak control chars into the CSV. csv.writer
+                    # quotes embedded newlines but does not strip control chars.
+                    cell = sanitize_display(str(d.get(col, "") or ""), maxlen=_EXPORT_CELL_MAXLEN)
+                row.append(_csv_safe(cell))
+            buf = io.StringIO()
+            writer = csv.writer(buf)
+            writer.writerow(row)
+            yield buf.getvalue().encode("utf-8")
+    finally:
+        gen.close()
+
+
+def _export_json_generator(
+    repo: ActivityLogRepository,
+    filters: ActivityLogFilters,
+) -> Iterator[bytes]:
+    """Yield a streamed JSON array, one entry per chunk.
+
+    Format: ``[\\n{entry},\\n{entry},\\n...]\\n``
+    The generator wraps ``repo.iter_export()`` in a ``try/finally`` so the DB
+    lock is released even on early client disconnect.
+    """
+    gen = repo.iter_export(filters)
+    try:
+        first = True
+        yield b"[\n"
+        for entry in gen:
+            d = entry_to_dict(entry)
+            chunk = json.dumps(d, ensure_ascii=False, default=str)
+            if first:
+                yield chunk.encode("utf-8")
+                first = False
+            else:
+                yield b",\n" + chunk.encode("utf-8")
+        yield b"\n]\n"
+    finally:
+        gen.close()
+
+
+@router.get("/export", summary="Export activity-log entries (streaming CSV or JSON)")
+def export_activity_log(
+    auth: AuthRequired,
+    repo: ActivityLogRepository = Depends(get_activity_log_repo),
+    # ── Format ────────────────────────────────────────────────────────────
+    format: Annotated[
+        str,
+        Query(description="Export format: 'csv' or 'json'"),
+    ] = "csv",
+    # ── Same filters as list ───────────────────────────────────────────────
+    categories: Annotated[list[str] | None, Query()] = None,
+    severities: Annotated[list[str] | None, Query()] = None,
+    actor: Annotated[str | None, Query(max_length=_MAX_ID_FILTER)] = None,
+    entity_type: Annotated[str | None, Query(max_length=_MAX_ID_FILTER)] = None,
+    entity_id: Annotated[str | None, Query(max_length=_MAX_ID_FILTER)] = None,
+    since: Annotated[datetime | None, Query()] = None,
+    until: Annotated[datetime | None, Query()] = None,
+    q: Annotated[str | None, Query(max_length=_MAX_TEXT_FILTER)] = None,
+) -> StreamingResponse:
+    """Stream all matching entries as CSV or JSON.
+
+    Requires a non-anonymous API key (loopback-anonymous access is rejected
+    because the log may contain IP addresses and entity names).
+    """
+    require_authenticated(auth)
+
+    if format not in ("csv", "json"):
+        from fastapi import HTTPException
+
+        raise HTTPException(
+            status_code=422,
+            detail="'format' must be 'csv' or 'json'",
+        )
+
+    filters = _build_filters(categories, severities, actor, entity_type, entity_id, since, until, q)
+
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H%M%S")
+
+    if format == "csv":
+        filename = f"activity-log-{timestamp}.csv"
+        media_type = "text/csv; charset=utf-8"
+        generator = _export_csv_generator(repo, filters)
+    else:
+        filename = f"activity-log-{timestamp}.json"
+        media_type = "application/json"
+        generator = _export_json_generator(repo, filters)
+
+    return StreamingResponse(
+        generator,
+        media_type=media_type,
+        headers={"Content-Disposition": f'attachment; filename="{filename}"'},
+    )
+
+
+# ---------------------------------------------------------------------------
+# GET /api/v1/activity-log/settings
+# PUT /api/v1/activity-log/settings
+# ---------------------------------------------------------------------------
+
+
+@router.get(
+    "/settings",
+    response_model=ActivityLogSettingsResponse,
+    summary="Get activity-log retention settings",
+)
+def get_activity_log_settings(
+    _: AuthRequired,
+    engine: ActivityLogRetentionEngine = Depends(get_activity_log_retention_engine),
+) -> ActivityLogSettingsResponse:
+    """Return the current activity-log retention settings."""
+    return ActivityLogSettingsResponse(**engine.get_settings())
+
+
+@router.put(
+    "/settings",
+    response_model=ActivityLogSettingsResponse,
+    summary="Update activity-log retention settings",
+)
+async def update_activity_log_settings(
+    auth: AuthRequired,
+    body: UpdateActivityLogSettingsRequest,
+    engine: ActivityLogRetentionEngine = Depends(get_activity_log_retention_engine),
+) -> ActivityLogSettingsResponse:
+    """Update the activity-log retention settings (applied immediately).
+
+    Requires a non-anonymous API key (loopback-anonymous access is rejected)
+    because disabling the log or pruning retention is equivalent in impact to
+    clearing the audit trail.
+
+    Setting ``enabled=false`` records an audit entry BEFORE the flag takes
+    effect so the last entry in the log shows who disabled recording.
+    """
+    require_authenticated(auth)
+    result = await engine.update_settings(
+        enabled=body.enabled,
+        max_days=body.max_days,
+        max_entries=body.max_entries,
+    )
+    return ActivityLogSettingsResponse(**result)
+
+
+# ---------------------------------------------------------------------------
+# DELETE /api/v1/activity-log  — clear
+# ---------------------------------------------------------------------------
+
+
+@router.delete("", summary="Clear all activity-log entries")
+def clear_activity_log(
+    auth: AuthRequired,
+    repo: ActivityLogRepository = Depends(get_activity_log_repo),
+    recorder: ActivityRecorder = Depends(get_activity_recorder),
+) -> dict:
+    """Delete all activity-log entries.
+
+    Requires a non-anonymous API key (loopback-anonymous access is rejected).
+    The clear operation itself is audited — a ``system/activity_log_cleared``
+    entry is recorded AFTER the wipe, so the log shows who cleared it and how
+    many rows were removed.
+
+    Returns ``{"deleted": <count>}``.
+    """
+    require_authenticated(auth)
+
+    deleted = repo.clear()
+
+    # Record the clear action (best-effort — recorder never raises).
+    recorder.record(
+        category=ActivityCategory.SYSTEM,
+        action="activity_log.cleared",
+        severity=ActivitySeverity.INFO,
+        actor=auth,
+        message=f"Activity log cleared ({deleted} entries removed)",
+        metadata={"deleted_count": deleted},
+    )
+
+    return {"deleted": deleted}
@@ -182,6 +182,12 @@ async def delete_audio_source(
    css_store: ColorStripStore = Depends(get_color_strip_store),
 ):
    """Delete an audio source."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_source(source_id).name
+    except Exception:
+        pass
+
    try:
        # Check if any CSS entities reference this audio source
        from ledgrab.storage.color_strip_source import AudioColorStripSource
@@ -194,7 +200,7 @@ async def delete_audio_source(
                raise ValueError(f"Cannot delete: referenced by color strip source '{css.name}'")

        store.delete_source(source_id)
-        fire_entity_event("audio_source", "deleted", source_id)
+        fire_entity_event("audio_source", "deleted", source_id, entity_name=_entity_name)
    except EntityNotFoundError as e:
        raise HTTPException(status_code=404, detail=str(e))

@@ -12,28 +12,35 @@ from ledgrab.api.dependencies import (
    get_scene_preset_store,
 )
 from ledgrab.api.schemas.automations import (
+    ActionSchema,
    AutomationCreate,
    AutomationListResponse,
    AutomationResponse,
+    AutomationTriggerResponse,
    AutomationUpdate,
    RuleSchema,
 )
 from ledgrab.core.automations.automation_engine import AutomationEngine
 from ledgrab.storage.automation import (
+    Action,
    ApplicationRule,
    DisplayStateRule,
    HomeAssistantRule,
    HTTPPollRule,
+    ManualTriggerRule,
    MQTTRule,
    Rule,
+    SolarRule,
    StartupRule,
    SystemIdleRule,
    TimeOfDayRule,
+    WebhookAction,
    WebhookRule,
 )
 from ledgrab.storage.automation_store import AutomationStore
 from ledgrab.storage.scene_preset_store import ScenePresetStore
 from ledgrab.utils import get_logger
+from ledgrab.utils.safe_source import validate_polling_url
 from ledgrab.storage.base_store import EntityNotFoundError

 logger = get_logger(__name__)
@@ -55,6 +62,20 @@ def _rule_from_schema(s: RuleSchema) -> Rule:
            days_of_week=s.days_of_week or [],
            timezone=s.timezone or "",
        ),
+        # SolarRule.from_dict validates events, clamps offsets/coords, and
+        # filters weekdays — route the raw schema values through it.
+        "solar": lambda: SolarRule.from_dict(
+            {
+                "start_event": s.start_event,
+                "start_offset_minutes": s.start_offset_minutes,
+                "end_event": s.end_event,
+                "end_offset_minutes": s.end_offset_minutes,
+                "latitude": s.latitude,
+                "longitude": s.longitude,
+                "days_of_week": s.days_of_week or [],
+                "timezone": s.timezone or "",
+            }
+        ),
        "system_idle": lambda: SystemIdleRule(
            idle_minutes=s.idle_minutes if s.idle_minutes is not None else 5,
            when_idle=s.when_idle if s.when_idle is not None else True,
@@ -72,6 +93,7 @@ def _rule_from_schema(s: RuleSchema) -> Rule:
            token=s.token or secrets.token_hex(16),
        ),
        "startup": lambda: StartupRule(),
+        "manual_trigger": lambda: ManualTriggerRule(),
        "home_assistant": lambda: HomeAssistantRule(
            ha_source_id=s.ha_source_id or "",
            entity_id=s.entity_id or "",
@@ -95,6 +117,43 @@ def _rule_to_schema(r: Rule) -> RuleSchema:
    return RuleSchema(**d)


+def _action_from_schema(s: ActionSchema) -> Action:
+    """Build a domain Action from its request schema, validating the webhook URL.
+
+    The SSRF gate runs here (save time) AND again at fire time, closing the
+    DNS-rebinding window. A bad/blocked URL rejects the whole save with 400.
+    """
+    if s.action_type != "webhook":
+        raise ValueError(f"Unknown action type: {s.action_type}")
+    url = (s.webhook_url or "").strip()
+    if not url:
+        raise ValueError("webhook action requires a webhook_url")
+    method = (s.method or "POST").upper()
+    if method not in ("POST", "PUT", "GET"):
+        raise ValueError(f"Invalid webhook method: {method}. Must be POST, PUT or GET.")
+    fire_on = s.fire_on or "activate"
+    if fire_on not in ("activate", "deactivate", "both"):
+        raise ValueError(f"Invalid fire_on: {fire_on}. Must be activate, deactivate or both.")
+    # content_type is emitted verbatim as the outbound Content-Type header — reject
+    # control chars (CR/LF) so it can't be used to inject additional HTTP headers.
+    content_type = (s.content_type or "application/json").strip()
+    if len(content_type) > 128 or any(ord(c) < 0x20 or ord(c) > 0x7E for c in content_type):
+        raise ValueError("Invalid content_type: control or non-ASCII characters are not allowed.")
+    # Raises HTTPException(400) on a blocked/loopback/metadata target.
+    validate_polling_url(url)
+    return WebhookAction(
+        webhook_url=url,
+        method=method,
+        body_template=s.body_template or "",
+        content_type=content_type,
+        fire_on=fire_on,
+    )
+
+
+def _action_to_schema(a: Action) -> ActionSchema:
+    return ActionSchema(**a.to_dict())
+
+
 def _automation_to_response(
    automation, engine: AutomationEngine, request: Request = None
 ) -> AutomationResponse:
@@ -130,6 +189,7 @@ def _automation_to_response(
        last_activated_at=state.get("last_activated_at"),
        last_deactivated_at=state.get("last_deactivated_at"),
        tags=automation.tags,
+        actions=[_action_to_schema(a) for a in getattr(automation, "actions", [])],
        icon=getattr(automation, "icon", "") or "",
        icon_color=getattr(automation, "icon_color", "") or "",
        created_at=automation.created_at,
@@ -186,6 +246,7 @@ async def create_automation(

    try:
        rules = [_rule_from_schema(r) for r in data.rules]
+        actions = [_action_from_schema(a) for a in data.actions]
    except EntityNotFoundError as e:
        raise HTTPException(status_code=404, detail=str(e))

@@ -201,6 +262,7 @@ async def create_automation(
        deactivation_mode=data.deactivation_mode,
        deactivation_scene_preset_id=data.deactivation_scene_preset_id,
        tags=data.tags,
+        actions=actions,
        icon=data.icon,
        icon_color=data.icon_color,
    )
@@ -283,6 +345,13 @@ async def update_automation(
        except ValueError as e:
            raise HTTPException(status_code=400, detail=str(e))

+    actions = None
+    if data.actions is not None:
+        try:
+            actions = [_action_from_schema(a) for a in data.actions]
+        except ValueError as e:
+            raise HTTPException(status_code=400, detail=str(e))
+
    try:
        # If disabling, deactivate first
        if data.enabled is False:
@@ -297,6 +366,7 @@ async def update_automation(
            rules=rules,
            deactivation_mode=data.deactivation_mode,
            tags=data.tags,
+            actions=actions,
            icon=data.icon,
            icon_color=data.icon_color,
        )
@@ -329,6 +399,12 @@ async def delete_automation(
    engine: AutomationEngine = Depends(get_automation_engine),
 ):
    """Delete an automation."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_automation(automation_id).name
+    except Exception:
+        pass
+
    # Deactivate first
    await engine.deactivate_if_active(automation_id)

@@ -337,7 +413,7 @@ async def delete_automation(
    except ValueError as e:
        raise HTTPException(status_code=404, detail=str(e))

-    fire_entity_event("automation", "deleted", automation_id)
+    fire_entity_event("automation", "deleted", automation_id, entity_name=_entity_name)


 # ===== Enable/Disable =====
@@ -388,3 +464,37 @@ async def disable_automation(
        raise HTTPException(status_code=404, detail=str(e))

    return _automation_to_response(automation, engine, request)
+
+
+# ===== Manual trigger =====
+
+
+@router.post(
+    "/api/v1/automations/{automation_id}/trigger",
+    response_model=AutomationTriggerResponse,
+    tags=["Automations"],
+)
+async def trigger_automation(
+    automation_id: str,
+    _auth: AuthRequired,
+    store: AutomationStore = Depends(get_automation_store),
+    engine: AutomationEngine = Depends(get_automation_engine),
+):
+    """Manually fire an automation.
+
+    Evaluates the automation's rules with its manual trigger satisfied — so it
+    "still checks all of the rules" under the automation's ``rule_logic`` — and,
+    if it should activate, applies its scene once. Independent of the ``enabled``
+    flag (that gates only the background evaluation loop).
+    """
+    try:
+        automation = store.get_automation(automation_id)
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))
+
+    try:
+        status, errors = await engine.fire_manual_trigger(automation)
+    except Exception as e:  # noqa: BLE001 — surface a structured error, never a bare 500
+        logger.error("Manual trigger failed for automation %s: %s", automation_id, e)
+        return AutomationTriggerResponse(status="error", errors=[str(e)])
+    return AutomationTriggerResponse(status=status, errors=errors)
@@ -27,6 +27,7 @@ from ledgrab.api.schemas.system import (
 )
 from ledgrab.config import get_config
 from ledgrab.core.backup.auto_backup import AutoBackupEngine
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.storage.asset_store import AssetStore
 from ledgrab.storage.database import Database, freeze_writes
 from ledgrab.utils import get_logger, read_upload_capped
@@ -35,6 +36,22 @@ logger = get_logger(__name__)

 router = APIRouter()

+
+def _record_system(action: str, message: str, metadata: dict | None = None) -> None:
+    """Best-effort audit record for a system-level event."""
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action=action,
+            severity=ActivitySeverity.INFO,
+            message=message,
+            metadata=metadata or {},
+        )
+
+
 _SERVER_DIR = Path(__file__).resolve().parents[4]


@@ -143,6 +160,8 @@ def backup_config(
    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%dT%H%M%S")
    filename = f"ledgrab-backup-{timestamp}.zip"

+    _record_system("backup.created", f"Backup downloaded: {filename}", {"filename": filename})
+
    return StreamingResponse(
        zip_buffer,
        media_type="application/zip",
@@ -243,6 +262,7 @@ async def restore_config(

    freeze_writes()
    logger.info("Database restored from uploaded backup. Scheduling restart...")
+    _record_system("backup.restored", "Database restored from uploaded backup")
    _schedule_restart()

    return RestoreResponse(
@@ -257,6 +277,7 @@ def restart_server(_: AuthRequired):
    """Schedule a server restart and return immediately."""
    from ledgrab.server_ref import _broadcast_restarting

+    _record_system("server.restarting", "Server restart requested by user")
    _broadcast_restarting()
    _schedule_restart()
    return {"status": "restarting"}
@@ -267,6 +288,7 @@ def shutdown_server(_: AuthRequired):
    """Gracefully shut down the server."""
    from ledgrab.server_ref import request_shutdown

+    _record_system("server.shutdown_requested", "Server shutdown requested by user")
    request_shutdown()
    return {"status": "shutting_down"}

@@ -300,11 +322,17 @@ async def update_auto_backup_settings(
    engine: AutoBackupEngine = Depends(get_auto_backup_engine),
 ):
    """Update auto-backup settings (enable/disable, interval, max backups)."""
-    return await engine.update_settings(
+    result = await engine.update_settings(
        enabled=body.enabled,
        interval_hours=body.interval_hours,
        max_backups=body.max_backups,
    )
+    _record_system(
+        "settings.changed",
+        f"Auto-backup settings updated (enabled={body.enabled})",
+        {"setting_key": "auto_backup", "enabled": body.enabled},
+    )
+    return result


@router.post("/api/v1/system/auto-backup/trigger", tags=["System"])
@@ -365,4 +393,5 @@ async def delete_saved_backup(
        engine.delete_backup(filename)
    except (ValueError, FileNotFoundError) as e:
        raise HTTPException(status_code=404, detail=str(e))
+    _record_system("backup.deleted", f"Saved backup deleted: {filename}", {"filename": filename})
    return {"status": "deleted", "filename": filename}
@@ -0,0 +1,272 @@
+"""Calibration session and solver API routes.
+
+Endpoints
+---------
+POST /api/v1/calibration/session
+    Start a calibration session on a device (stops any running target on that
+    device and remembers it for restore on stop).
+
+POST /api/v1/calibration/session/position
+    Advance the chase pixel to a specific LED index on the active device.
+
+POST /api/v1/calibration/session/stop
+    End the session: clear the device to black and restore the prior target.
+
+POST /api/v1/calibration/session/cancel
+    Alias for stop (does not apply any solved calibration).
+
+GET  /api/v1/calibration/session/state
+    Return the current session state (active, device, last_activity, …).
+
+POST /api/v1/calibration/solve
+    Pure-logic: solve a CalibrationConfig from 4 corner tap indices.
+    Does NOT persist — the caller must follow up with
+    ``PUT /api/v1/color-strip-sources/{id}`` to persist.
+
+Persist path
+------------
+The existing ``PUT /api/v1/color-strip-sources/{id}`` already accepts a
+``calibration`` field on ``PictureCSSUpdate`` / ``PictureAdvancedCSSUpdate``
+and hot-reloads running streams automatically (see
+``api/routes/color_strip_sources/crud.py``).  There is NO duplicate endpoint
+here.  Phase 3 UI calls the existing PUT to persist.
+"""
+
+from fastapi import APIRouter, Depends, HTTPException
+
+from ledgrab.api.auth import AuthRequired
+from ledgrab.api.dependencies import get_processor_manager
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+from ledgrab.api.schemas.calibration import (
+    CalibrationSessionPositionRequest,
+    CalibrationSessionStartRequest,
+    CalibrationSessionStateResponse,
+    CalibrationSolveRequest,
+    CalibrationSolvedResponse,
+)
+from ledgrab.core.capture.calibration import solve_calibration
+from ledgrab.core.capture.calibration_session import get_calibration_session
+from ledgrab.core.processing.processor_manager import ProcessorManager
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+router = APIRouter()
+
+
+# ── Session endpoints ─────────────────────────────────────────────────────────
+
+
+@router.post(
+    "/api/v1/calibration/session",
+    response_model=CalibrationSessionStateResponse,
+    tags=["Calibration"],
+    status_code=201,
+)
+async def start_calibration_session(
+    body: CalibrationSessionStartRequest,
+    _auth: AuthRequired,
+    manager: ProcessorManager = Depends(get_processor_manager),
+) -> CalibrationSessionStateResponse:
+    """Start a calibration session on a device.
+
+    Stops any target currently processing on that device (it will be restored
+    when the session ends).  Only one session can be active at a time; starting
+    a new one terminates the previous one first.
+    """
+    session = get_calibration_session()
+    try:
+        await session.start(body.device_id, manager)
+    except ValueError as exc:
+        raise HTTPException(status_code=404, detail=str(exc))
+    except Exception as exc:
+        logger.error("Failed to start calibration session: %s", exc, exc_info=True)
+        raise HTTPException(status_code=500, detail="Internal server error")
+
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action="calibration.started",
+            severity=ActivitySeverity.INFO,
+            entity_type="device",
+            entity_id=body.device_id,
+            message=f"Calibration session started for device '{body.device_id}'",
+        )
+
+    return CalibrationSessionStateResponse(**session.get_state())
+
+
+@router.post(
+    "/api/v1/calibration/session/position",
+    response_model=CalibrationSessionStateResponse,
+    tags=["Calibration"],
+)
+async def calibration_session_position(
+    body: CalibrationSessionPositionRequest,
+    _auth: AuthRequired,
+    manager: ProcessorManager = Depends(get_processor_manager),  # noqa: ARG001
+) -> CalibrationSessionStateResponse:
+    """Advance the chase pixel to a specific LED index on the active device.
+
+    ``index`` must be 0-based and < ``led_count``.  Returns 422 when out of
+    range (Pydantic ``ge=0``) or 400 if the session is not active / index
+    exceeds led_count.
+    """
+    session = get_calibration_session()
+    try:
+        await session.position(body.index, body.window)
+    except RuntimeError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+    except Exception as exc:
+        logger.error("Failed to set calibration pixel index=%d: %s", body.index, exc, exc_info=True)
+        raise HTTPException(status_code=500, detail="Internal server error")
+
+    return CalibrationSessionStateResponse(**session.get_state())
+
+
+@router.post(
+    "/api/v1/calibration/session/stop",
+    response_model=CalibrationSessionStateResponse,
+    tags=["Calibration"],
+)
+async def stop_calibration_session(
+    _auth: AuthRequired,
+    manager: ProcessorManager = Depends(get_processor_manager),  # noqa: ARG001
+) -> CalibrationSessionStateResponse:
+    """End the calibration session.
+
+    Clears the device to black and restores the previously-running target (if
+    any).  Safe to call even when no session is active (returns inactive state).
+    """
+    session = get_calibration_session()
+    try:
+        await session.stop()
+    except Exception as exc:
+        logger.error("Failed to stop calibration session: %s", exc, exc_info=True)
+        raise HTTPException(status_code=500, detail="Internal server error")
+
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action="calibration.stopped",
+            severity=ActivitySeverity.INFO,
+            message="Calibration session stopped",
+        )
+
+    return CalibrationSessionStateResponse(**session.get_state())
+
+
+@router.post(
+    "/api/v1/calibration/session/cancel",
+    response_model=CalibrationSessionStateResponse,
+    tags=["Calibration"],
+)
+async def cancel_calibration_session(
+    _auth: AuthRequired,
+    manager: ProcessorManager = Depends(get_processor_manager),  # noqa: ARG001
+) -> CalibrationSessionStateResponse:
+    """Cancel the calibration session (alias for stop — no calibration is applied)."""
+    session = get_calibration_session()
+    try:
+        await session.cancel()
+    except Exception as exc:
+        logger.error("Failed to cancel calibration session: %s", exc, exc_info=True)
+        raise HTTPException(status_code=500, detail="Internal server error")
+
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action="calibration.cancelled",
+            severity=ActivitySeverity.INFO,
+            message="Calibration session cancelled",
+        )
+
+    return CalibrationSessionStateResponse(**session.get_state())
+
+
+@router.get(
+    "/api/v1/calibration/session/state",
+    response_model=CalibrationSessionStateResponse,
+    tags=["Calibration"],
+)
+async def get_calibration_session_state(
+    _auth: AuthRequired,
+) -> CalibrationSessionStateResponse:
+    """Return the current calibration session state."""
+    return CalibrationSessionStateResponse(**get_calibration_session().get_state())
+
+
+# ── Solver endpoint ───────────────────────────────────────────────────────────
+
+
+@router.post(
+    "/api/v1/calibration/solve",
+    response_model=CalibrationSolvedResponse,
+    tags=["Calibration"],
+)
+async def solve_calibration_endpoint(
+    body: CalibrationSolveRequest,
+    _auth: AuthRequired,
+    manager: ProcessorManager = Depends(get_processor_manager),
+) -> CalibrationSolvedResponse:
+    """Solve a CalibrationConfig from 4 corner tap indices.
+
+    Returns the computed per-edge LED counts.  Does NOT persist — call
+    ``PUT /api/v1/color-strip-sources/{id}`` with ``calibration`` in the body
+    to save.
+
+    Provide either *device_id* (preferred, server derives led_count) or
+    *led_count* directly.  Returns 404 if *device_id* is not found, 422 on
+    invalid enum values, 400 on logical errors (e.g. corner_indices length).
+    """
+    # Resolve led_count
+    led_count = body.led_count
+    if body.device_id is not None:
+        if body.device_id not in manager._devices:
+            raise HTTPException(
+                status_code=404,
+                detail=f"Device {body.device_id!r} not found",
+            )
+        ds = manager._devices[body.device_id]
+        led_count = ds.led_count
+
+    if led_count is None or led_count <= 0:
+        raise HTTPException(
+            status_code=400,
+            detail="led_count must be a positive integer",
+        )
+
+    try:
+        cfg = solve_calibration(
+            led_count=led_count,
+            start_position=body.start_position,
+            layout=body.layout,
+            corner_indices=body.corner_indices,
+            offset=body.offset,
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+    except Exception as exc:
+        logger.error("Failed to solve calibration: %s", exc, exc_info=True)
+        raise HTTPException(status_code=500, detail="Internal server error")
+
+    return CalibrationSolvedResponse(
+        mode="simple",
+        layout=cfg.layout,
+        start_position=cfg.start_position,
+        leds_top=cfg.leds_top,
+        leds_right=cfg.leds_right,
+        leds_bottom=cfg.leds_bottom,
+        leds_left=cfg.leds_left,
+        offset=cfg.offset,
+    )
@@ -167,6 +167,12 @@ async def delete_color_strip_source(
    target_store: OutputTargetStore = Depends(get_output_target_store),
 ):
    """Delete a color strip source. Returns 409 if referenced by any LED target."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_source(source_id).name
+    except Exception:
+        pass
+
    try:
        target_names = target_store.get_targets_referencing_css(source_id)
        if target_names:
@@ -201,7 +207,7 @@ async def delete_color_strip_source(
                "Delete or reassign the processed source(s) first.",
            )
        store.delete_source(source_id)
-        fire_entity_event("color_strip_source", "deleted", source_id)
+        fire_entity_event("color_strip_source", "deleted", source_id, entity_name=_entity_name)
    except HTTPException:
        raise
    except ValueError as e:
@@ -96,13 +96,16 @@ def _device_to_response(device) -> DeviceResponse:
        espnow_channel=device.espnow_channel,
        hue_paired=bool(device.hue_username and device.hue_client_key),
        hue_entertainment_group_id=device.hue_entertainment_group_id,
+        hue_gradient_mode=device.hue_gradient_mode,
        yeelight_min_interval_ms=device.yeelight_min_interval_ms,
        wiz_min_interval_ms=device.wiz_min_interval_ms,
        lifx_min_interval_ms=device.lifx_min_interval_ms,
+        lifx_per_zone=device.lifx_per_zone,
        govee_min_interval_ms=device.govee_min_interval_ms,
        opc_channel=device.opc_channel,
        nanoleaf_paired=bool(device.nanoleaf_token),
        nanoleaf_min_interval_ms=device.nanoleaf_min_interval_ms,
+        nanoleaf_per_panel=device.nanoleaf_per_panel,
        spi_speed_hz=device.spi_speed_hz,
        spi_led_type=device.spi_led_type,
        chroma_device_type=device.chroma_device_type,
@@ -261,6 +264,9 @@ async def create_device(
            hue_username=device_data.hue_username or "",
            hue_client_key=device_data.hue_client_key or "",
            hue_entertainment_group_id=device_data.hue_entertainment_group_id or "",
+            hue_gradient_mode=(
+                device_data.hue_gradient_mode if device_data.hue_gradient_mode is not None else True
+            ),
            yeelight_min_interval_ms=(
                device_data.yeelight_min_interval_ms
                if device_data.yeelight_min_interval_ms is not None
@@ -276,6 +282,7 @@ async def create_device(
                if device_data.lifx_min_interval_ms is not None
                else 50
            ),
+            lifx_per_zone=bool(device_data.lifx_per_zone),
            govee_min_interval_ms=(
                device_data.govee_min_interval_ms
                if device_data.govee_min_interval_ms is not None
@@ -288,6 +295,7 @@ async def create_device(
                if device_data.nanoleaf_min_interval_ms is not None
                else 100
            ),
+            nanoleaf_per_panel=bool(device_data.nanoleaf_per_panel),
            spi_speed_hz=device_data.spi_speed_hz or 800000,
            spi_led_type=device_data.spi_led_type or "WS2812B",
            chroma_device_type=device_data.chroma_device_type or "chromalink",
@@ -631,13 +639,16 @@ async def update_device(
            hue_username=update_data.hue_username,
            hue_client_key=update_data.hue_client_key,
            hue_entertainment_group_id=update_data.hue_entertainment_group_id,
+            hue_gradient_mode=update_data.hue_gradient_mode,
            yeelight_min_interval_ms=update_data.yeelight_min_interval_ms,
            wiz_min_interval_ms=update_data.wiz_min_interval_ms,
            lifx_min_interval_ms=update_data.lifx_min_interval_ms,
+            lifx_per_zone=update_data.lifx_per_zone,
            govee_min_interval_ms=update_data.govee_min_interval_ms,
            opc_channel=update_data.opc_channel,
            nanoleaf_token=update_data.nanoleaf_token,
            nanoleaf_min_interval_ms=update_data.nanoleaf_min_interval_ms,
+            nanoleaf_per_panel=update_data.nanoleaf_per_panel,
            spi_speed_hz=update_data.spi_speed_hz,
            spi_led_type=update_data.spi_led_type,
            chroma_device_type=update_data.chroma_device_type,
@@ -701,6 +712,13 @@ async def delete_device(
 ):
    """Delete/detach a device. Returns 409 if referenced by a target."""
    try:
+        # Resolve name before deletion for the audit record.
+        _entity_name: str | None = None
+        try:
+            _entity_name = store.get_device(device_id).name
+        except Exception:
+            pass
+
        # Check if any target references this device
        refs = target_store.get_targets_for_device(device_id)
        if refs:
@@ -728,7 +746,7 @@ async def delete_device(
        # Delete from storage
        store.delete_device(device_id)

-        fire_entity_event("device", "deleted", device_id)
+        fire_entity_event("device", "deleted", device_id, entity_name=_entity_name)
        logger.info(f"Deleted device {device_id}")

    except HTTPException:
@@ -6,6 +6,7 @@ adapter metadata, and diagnostics.

 import threading
 import time
+from collections import defaultdict
 from typing import Any

 from fastapi import APIRouter, Depends, HTTPException, Request
@@ -16,6 +17,7 @@ from ledgrab.api.dependencies import (
    get_database,
    get_game_integration_store,
    get_game_event_bus,
+    get_lol_poll_manager,
 )
 from ledgrab.api.schemas.game_integration import (
    AdapterInfoResponse,
@@ -36,9 +38,16 @@ from ledgrab.api.schemas.game_integration import (
 )
 from ledgrab.core.game_integration.adapter_registry import AdapterRegistry
 from ledgrab.core.game_integration.event_bus import GameEventBus
-from ledgrab.core.game_integration.events import GameEvent
+from ledgrab.core.game_integration.lol_poll_manager import LoLPollManager
+from ledgrab.core.game_integration.runtime_state import (
+    cleanup_state as _cleanup_state,
+    get_prev_state as _get_prev_state,
+    get_stats as _get_stats,
+    record_events as _record_events,
+    set_prev_state as _set_prev_state,
+)
 from ledgrab.storage.base_store import EntityNotFoundError
-from ledgrab.storage.game_integration import EventMapping
+from ledgrab.storage.game_integration import _SECRET_CONFIG_KEYS, EventMapping
 from ledgrab.storage.game_integration_store import GameIntegrationStore
 from ledgrab.utils import get_logger

@@ -46,15 +55,77 @@ logger = get_logger(__name__)

 router = APIRouter()

-# ── Per-integration runtime state (in-memory, not persisted) ──────────────
+# Per-integration runtime state (prev-state + stats + payload processing) lives
+# in ``core/game_integration/runtime_state.py`` and is imported above under the
+# legacy ``_get_prev_state`` / ``_record_events`` / … names so both this route
+# and the LoL poll manager share one set of counters.

-_integration_state_lock = threading.Lock()

-# integration_id -> prev_state dict for diff-based trigger detection
-_prev_states: dict[str, dict[str, Any]] = {}
+# ── Failed-auth rate limiter (brute-force defence on the ingest route) ─────
+#
+# The ingest route is high-frequency (games push at 16-64 Hz), so we do NOT
+# rate-limit every event — that would throttle legitimate gameplay traffic.
+# Instead we throttle only FAILED-auth attempts per source IP (the only thing
+# an attacker without the token can produce). This mirrors the IP-based
+# limiter in routes/webhooks.py (~30/min) but scopes it to failures so a
+# brute-forcer is locked out after _AUTH_FAIL_LIMIT bad tokens per minute
+# while authenticated high-rate ingestion is completely unaffected.
+_AUTH_FAIL_LIMIT = 30
+_AUTH_FAIL_WINDOW = 60.0  # seconds
+_AUTH_FAIL_HITS_HARD_CAP = 1024
+_LOOPBACK_HOSTS = frozenset({"127.0.0.1", "::1", "localhost"})
+_auth_fail_hits: dict[str, list[float]] = defaultdict(list)
+_auth_fail_lock = threading.Lock()

-# integration_id -> runtime stats
-_integration_stats: dict[str, dict[str, Any]] = {}
+
+def _rate_limit_key(request: Request) -> str:
+    """Pick a stable client identifier for rate-limiting.
+
+    When the immediate peer is loopback (assumed reverse-proxy), use the
+    first ``X-Forwarded-For`` entry; otherwise use the peer's IP.
+    """
+    peer = request.client.host if request.client else "unknown"
+    if peer in _LOOPBACK_HOSTS:
+        xff = request.headers.get("x-forwarded-for", "")
+        if xff:
+            return xff.split(",", 1)[0].strip() or peer
+    return peer
+
+
+def _check_auth_fail_rate_limit(client_ip: str) -> None:
+    """Raise 429 if *client_ip* exceeded the failed-auth attempt limit."""
+    now = time.time()
+    window_start = now - _AUTH_FAIL_WINDOW
+    with _auth_fail_lock:
+        timestamps = [t for t in _auth_fail_hits[client_ip] if t > window_start]
+        _auth_fail_hits[client_ip] = timestamps
+        if len(timestamps) >= _AUTH_FAIL_LIMIT:
+            raise HTTPException(
+                status_code=429,
+                detail="Too many failed authentication attempts. Try again later.",
+            )
+
+
+def _record_auth_failure(client_ip: str) -> None:
+    """Record a failed-auth attempt for *client_ip* (bounded memory)."""
+    now = time.time()
+    window_start = now - _AUTH_FAIL_WINDOW
+    with _auth_fail_lock:
+        _auth_fail_hits[client_ip].append(now)
+        # Periodic cleanup of stale IPs to prevent unbounded growth.
+        if len(_auth_fail_hits) > 100:
+            stale = [ip for ip, ts in _auth_fail_hits.items() if not ts or ts[-1] < window_start]
+            for ip in stale:
+                del _auth_fail_hits[ip]
+        # Hard cap against an attacker spraying many distinct X-Forwarded-For
+        # values; drop the oldest-touched IPs.
+        if len(_auth_fail_hits) > _AUTH_FAIL_HITS_HARD_CAP:
+            ordered = sorted(
+                _auth_fail_hits.items(),
+                key=lambda kv: kv[1][-1] if kv[1] else 0.0,
+            )
+            for ip, _ in ordered[: len(ordered) - _AUTH_FAIL_HITS_HARD_CAP]:
+                _auth_fail_hits.pop(ip, None)


 def _schema_to_fields(schema: dict[str, Any]) -> list[dict[str, Any]]:
@@ -82,59 +153,47 @@ def _schema_to_fields(schema: dict[str, Any]) -> list[dict[str, Any]]:
    return fields


-def _get_prev_state(integration_id: str) -> dict[str, Any]:
-    """Get or create the prev_state dict for an integration."""
-    with _integration_state_lock:
-        if integration_id not in _prev_states:
-            _prev_states[integration_id] = {}
-        return _prev_states[integration_id]
-
-
-def _set_prev_state(integration_id: str, state: dict[str, Any]) -> None:
-    """Update the prev_state dict for an integration."""
-    with _integration_state_lock:
-        _prev_states[integration_id] = state
-
-
-def _record_events(integration_id: str, events: list[GameEvent]) -> None:
-    """Record event stats for an integration."""
-    with _integration_state_lock:
-        if integration_id not in _integration_stats:
-            _integration_stats[integration_id] = {
-                "event_count": 0,
-                "event_counts_by_type": {},
-                "last_event_time": None,
-            }
-        stats = _integration_stats[integration_id]
-        for event in events:
-            stats["event_count"] += 1
-            stats["event_counts_by_type"][event.event_type] = (
-                stats["event_counts_by_type"].get(event.event_type, 0) + 1
-            )
-            stats["last_event_time"] = event.timestamp
-
-
-def _get_stats(integration_id: str) -> dict[str, Any]:
-    """Get runtime stats for an integration."""
-    with _integration_state_lock:
-        return _integration_stats.get(
-            integration_id,
-            {"event_count": 0, "event_counts_by_type": {}, "last_event_time": None},
-        )
-
-
-def _cleanup_state(integration_id: str) -> None:
-    """Remove runtime state for a deleted integration."""
-    with _integration_state_lock:
-        _prev_states.pop(integration_id, None)
-        _integration_stats.pop(integration_id, None)
-
-
 # ── Helper: convert config to response ────────────────────────────────────


+def _redact_secrets(adapter_config: dict[str, Any]) -> dict[str, Any]:
+    """Return a copy of *adapter_config* with secret values masked.
+
+    The adapter ``auth_token`` is a live shared secret (it authenticates the
+    ingest endpoint). It is encrypted at rest, but the response builder echoes
+    the in-memory *decrypted* config, so without masking any API caller
+    (loopback-anonymous by default) could read the cleartext token. We never
+    return the secret over the API — the edit form submits a blank value to
+    keep the existing secret (see ``_merge_preserved_secrets``).
+    """
+    cfg = dict(adapter_config)
+    for key in _SECRET_CONFIG_KEYS:
+        if cfg.get(key):
+            cfg[key] = ""  # mask — never echo the secret to the client
+    return cfg
+
+
+def _merge_preserved_secrets(
+    incoming: dict[str, Any] | None, existing: Any
+) -> dict[str, Any] | None:
+    """Preserve a stored secret when an update submits a blank/absent one.
+
+    Because the API masks secrets in responses, the edit form re-submits a
+    blank value for an unchanged secret. Without this merge that blank would
+    overwrite (and destroy) the stored token. A non-empty incoming value is a
+    deliberate change and is kept as-is.
+    """
+    if incoming is None:
+        return None
+    merged = dict(incoming)
+    for key in _SECRET_CONFIG_KEYS:
+        if not merged.get(key) and existing.adapter_config.get(key):
+            merged[key] = existing.adapter_config[key]
+    return merged
+
+
 def _config_to_response(config: Any) -> GameIntegrationResponse:
-    """Convert a GameIntegrationConfig to its API response."""
+    """Convert a GameIntegrationConfig to its API response (secrets redacted)."""
    from ledgrab.api.schemas.game_integration import EventMappingSchema

    return GameIntegrationResponse(
@@ -142,7 +201,7 @@ def _config_to_response(config: Any) -> GameIntegrationResponse:
        name=config.name,
        adapter_type=config.adapter_type,
        enabled=config.enabled,
-        adapter_config=config.adapter_config,
+        adapter_config=_redact_secrets(config.adapter_config),
        event_mappings=[
            EventMappingSchema(
                event_type=m.event_type,
@@ -234,6 +293,7 @@ async def create_integration(
    data: GameIntegrationCreate,
    _auth: AuthRequired,
    store: GameIntegrationStore = Depends(get_game_integration_store),
+    lol_mgr: LoLPollManager | None = Depends(get_lol_poll_manager),
 ):
    """Create a new game integration config."""
    try:
@@ -262,6 +322,8 @@ async def create_integration(
        )

        fire_entity_event("game_integration", "created", config.id)
+        if lol_mgr is not None:
+            lol_mgr.sync(store.get_all_integrations())
        return _config_to_response(config)

    except EntityNotFoundError as e:
@@ -301,6 +363,7 @@ async def update_integration(
    data: GameIntegrationUpdate,
    _auth: AuthRequired,
    store: GameIntegrationStore = Depends(get_game_integration_store),
+    lol_mgr: LoLPollManager | None = Depends(get_lol_poll_manager),
 ):
    """Update a game integration config."""
    try:
@@ -318,12 +381,20 @@ async def update_integration(
                for m in data.event_mappings
            ]

+        # Preserve a stored secret when the update submits a blank token
+        # (the API masks secrets, so the edit form re-sends a blank value
+        # for an unchanged secret — see _merge_preserved_secrets).
+        adapter_config = data.adapter_config
+        if adapter_config is not None:
+            existing = store.get_integration(integration_id)
+            adapter_config = _merge_preserved_secrets(adapter_config, existing)
+
        config = store.update_integration(
            integration_id=integration_id,
            name=data.name,
            adapter_type=data.adapter_type,
            enabled=data.enabled,
-            adapter_config=data.adapter_config,
+            adapter_config=adapter_config,
            event_mappings=mappings,
            description=data.description,
            tags=data.tags,
@@ -332,6 +403,8 @@ async def update_integration(
        )

        fire_entity_event("game_integration", "updated", integration_id)
+        if lol_mgr is not None:
+            lol_mgr.sync(store.get_all_integrations())
        return _config_to_response(config)

    except EntityNotFoundError as e:
@@ -352,11 +425,14 @@ async def delete_integration(
    integration_id: str,
    _auth: AuthRequired,
    store: GameIntegrationStore = Depends(get_game_integration_store),
+    lol_mgr: LoLPollManager | None = Depends(get_lol_poll_manager),
 ):
    """Delete a game integration config."""
    try:
        store.delete_integration(integration_id)
        _cleanup_state(integration_id)
+        if lol_mgr is not None:
+            lol_mgr.sync(store.get_all_integrations())
        fire_entity_event("game_integration", "deleted", integration_id)
    except EntityNotFoundError as e:
        raise HTTPException(status_code=404, detail=str(e))
@@ -387,7 +463,16 @@ async def ingest_event(
    called before standard API auth.

    No AuthRequired dependency — adapter-level auth is used instead.
+
+    Rate limiting is scoped to FAILED-auth attempts per source IP (see
+    ``_check_auth_fail_rate_limit``) so legitimate high-rate ingestion is
+    never throttled, but a brute-forcer is locked out after the threshold.
    """
+    client_ip = _rate_limit_key(request)
+    # Block IPs that have already burned through the failed-auth budget,
+    # before doing any work (cheap brute-force lockout).
+    _check_auth_fail_rate_limit(client_ip)
+
    try:
        config = store.get_integration(integration_id)
    except EntityNotFoundError:
@@ -402,9 +487,18 @@ async def ingest_event(
    except ValueError as e:
        raise HTTPException(status_code=400, detail=str(e))

-    # Adapter-level auth check
+    # Adapter-level auth check. Treat ANY exception from validate_auth as an
+    # auth failure (rate-limited + 403), never a 500 — a malformed/attacker-
+    # controlled token must not crash the handler nor bypass the brute-force
+    # lockout counter.
    headers = dict(request.headers)
-    if not adapter_cls.validate_auth(headers, payload.data, config.adapter_config):
+    try:
+        authed = adapter_cls.validate_auth(headers, payload.data, config.adapter_config)
+    except Exception as exc:
+        logger.warning("validate_auth raised for %s: %s", integration_id, exc)
+        authed = False
+    if not authed:
+        _record_auth_failure(client_ip)
        raise HTTPException(status_code=403, detail="Adapter authentication failed")

    # Parse payload through adapter
@@ -152,13 +152,19 @@ async def delete_gradient(
    css_store: ColorStripStore = Depends(get_color_strip_store),
 ):
    """Delete a gradient (fails if built-in or referenced by sources)."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_gradient(gradient_id).name
+    except Exception:
+        pass
+
    try:
        # Check references
        for source in css_store.get_all_sources():
            if getattr(source, "gradient_id", None) == gradient_id:
                raise ValueError(f"Cannot delete: referenced by color strip source '{source.name}'")
        store.delete_gradient(gradient_id)
-        fire_entity_event("gradient", "deleted", gradient_id)
+        fire_entity_event("gradient", "deleted", gradient_id, entity_name=_entity_name)
    except (ValueError, EntityNotFoundError) as e:
        status = 404 if "not found" in str(e).lower() else 400
        raise HTTPException(status_code=status, detail=str(e))
@@ -2,6 +2,7 @@

 import asyncio
 import json
+from urllib.parse import urlparse

 from fastapi import APIRouter, Depends, HTTPException, Query

@@ -28,6 +29,7 @@ from ledgrab.storage.base_store import EntityNotFoundError
 from ledgrab.storage.home_assistant_source import HomeAssistantSource
 from ledgrab.storage.home_assistant_store import HomeAssistantStore
 from ledgrab.utils import get_logger
+from ledgrab.utils.net_classify import validate_lan_host

 logger = get_logger(__name__)

@@ -37,6 +39,23 @@ router = APIRouter()
 _REDACTED_TOKEN = "***"


+def _validate_ha_host(host: str | None) -> None:
+    """Reject literal public/link-local/metadata IPs for a HA source host.
+
+    HA sources are LAN-by-design (loopback + private ranges allowed), so we
+    gate the user-supplied ``host`` with the same shared classifier the LED
+    device providers use (``validate_lan_host``). The HA host is stored as
+    ``host:port`` (e.g. ``192.168.1.100:8123``), so strip the port first via
+    ``urlparse`` — which also handles bracketed IPv6 literals. Hostnames /
+    mDNS labels pass through (classified UNPARSEABLE). Raises ``ValueError``
+    on a literal public IP, which the callers translate to HTTP 400.
+    """
+    if not host:
+        return
+    bare_host = urlparse(f"//{host.strip()}").hostname or host.strip()
+    validate_lan_host(bare_host)
+
+
 def _to_response(
    source: HomeAssistantSource,
    manager: HomeAssistantManager,
@@ -99,6 +118,7 @@ async def create_ha_source(
    manager: HomeAssistantManager = Depends(get_ha_manager),
 ):
    try:
+        _validate_ha_host(data.host)
        source = store.create_source(
            name=data.name,
            host=data.host,
@@ -153,6 +173,7 @@ async def update_ha_source(
    manager: HomeAssistantManager = Depends(get_ha_manager),
 ):
    try:
+        _validate_ha_host(data.host)
        source = store.update_source(
            source_id,
            name=data.name,
@@ -42,6 +42,8 @@ def _to_response(source: MQTTSource, manager: MQTTManager) -> MQTTSourceResponse
        password_set=bool(source.password),
        client_id=source.client_id,
        base_topic=source.base_topic,
+        publish_ha_discovery=getattr(source, "publish_ha_discovery", False),
+        discovery_prefix=getattr(source, "discovery_prefix", "homeassistant"),
        connected=runtime.is_connected if runtime else False,
        description=source.description,
        tags=source.tags,
@@ -90,6 +92,8 @@ async def create_mqtt_source(
            password=data.password,
            client_id=data.client_id,
            base_topic=data.base_topic,
+            publish_ha_discovery=data.publish_ha_discovery,
+            discovery_prefix=data.discovery_prefix,
            description=data.description,
            tags=data.tags,
            icon=data.icon,
@@ -97,6 +101,8 @@ async def create_mqtt_source(
        )
    except ValueError as e:
        raise HTTPException(status_code=400, detail=str(e))
+    # Publish HA discovery if the new source opted in.
+    await manager.sync_discovery(source.id)
    fire_entity_event("mqtt_source", "created", source.id)
    return _to_response(source, manager)

@@ -141,6 +147,8 @@ async def update_mqtt_source(
            password=data.password,
            client_id=data.client_id,
            base_topic=data.base_topic,
+            publish_ha_discovery=data.publish_ha_discovery,
+            discovery_prefix=data.discovery_prefix,
            description=data.description,
            tags=data.tags,
            icon=data.icon,
@@ -151,6 +159,8 @@ async def update_mqtt_source(
    except ValueError as e:
        raise HTTPException(status_code=400, detail=str(e))
    await manager.update_source(source_id)
+    # Reconcile HA discovery (publish if enabled, clear if turned off).
+    await manager.sync_discovery(source_id)
    fire_entity_event("mqtt_source", "updated", source.id)
    return _to_response(source, manager)

@@ -162,6 +172,9 @@ async def delete_mqtt_source(
    store: MQTTSourceStore = Depends(get_mqtt_store),
    manager: MQTTManager = Depends(get_mqtt_manager),
 ):
+    # Clear any HA discovery configs (needs the source still present to build
+    # the exact retained topics) before deleting the row.
+    await manager.disable_discovery(source_id)
    try:
        store.delete_source(source_id)
    except EntityNotFoundError:
@@ -624,6 +624,13 @@ async def delete_target(
 ):
    """Delete a output target. Stops processing first if active."""
    try:
+        # Resolve name before deletion for the audit record.
+        _entity_name: str | None = None
+        try:
+            _entity_name = target_store.get_target(target_id).name
+        except Exception:
+            pass
+
        # Stop processing if running
        try:
            await manager.stop_processing(target_id)
@@ -641,7 +648,7 @@ async def delete_target(
        # Delete from store
        target_store.delete_target(target_id)

-        fire_entity_event("output_target", "deleted", target_id)
+        fire_entity_event("output_target", "deleted", target_id, entity_name=_entity_name)
        logger.info(f"Deleted target {target_id}")

    except ValueError as e:
@@ -12,6 +12,7 @@ from ledgrab.api.dependencies import (
    get_picture_source_store,
    get_processor_manager,
 )
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.api.schemas.output_targets import (
    BulkTargetRequest,
    BulkTargetResponse,
@@ -28,6 +29,7 @@ from ledgrab.storage.color_strip_source import (
 from ledgrab.storage.picture_source_store import PictureSourceStore
 from ledgrab.storage.wled_output_target import WledOutputTarget
 from ledgrab.storage.output_target_store import OutputTargetStore
+from ledgrab.core.activity_log.sanitize import sanitize_display
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)
@@ -35,6 +37,23 @@ logger = get_logger(__name__)
 router = APIRouter()


+def _record_capture(action: str, target_id: str, target_name: str | None, message: str) -> None:
+    """Best-effort audit record for a capture start/stop action."""
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.CAPTURE,
+            action=action,
+            severity=ActivitySeverity.INFO,
+            entity_type="output_target",
+            entity_id=target_id,
+            entity_name=sanitize_display(target_name) if target_name else None,
+            message=message,
+        )
+
+
 # ===== BULK PROCESSING CONTROL ENDPOINTS =====


@@ -53,10 +72,18 @@ async def bulk_start_processing(

    for target_id in body.ids:
        try:
-            target_store.get_target(target_id)
+            _tgt = target_store.get_target(target_id)
            await manager.start_processing(target_id)
            started.append(target_id)
            logger.info(f"Bulk start: started processing for target {target_id}")
+            _tgt_name_raw = getattr(_tgt, "name", None)
+            _tgt_safe = sanitize_display(_tgt_name_raw) if _tgt_name_raw else None
+            _record_capture(
+                "capture.started",
+                target_id,
+                _tgt_safe,
+                f"Capture started for target '{_tgt_safe or target_id}' (bulk)",
+            )
        except ValueError as e:
            errors[target_id] = str(e)
        except RuntimeError as e:
@@ -78,6 +105,7 @@ async def bulk_start_processing(
 async def bulk_stop_processing(
    body: BulkTargetRequest,
    _auth: AuthRequired,
+    target_store: OutputTargetStore = Depends(get_output_target_store),
    manager: ProcessorManager = Depends(get_processor_manager),
 ):
    """Stop processing for multiple output targets. Returns lists of stopped IDs and per-ID errors."""
@@ -89,6 +117,18 @@ async def bulk_stop_processing(
            await manager.stop_processing(target_id)
            stopped.append(target_id)
            logger.info(f"Bulk stop: stopped processing for target {target_id}")
+            _tgt_name: str | None = None
+            try:
+                _tgt_name = target_store.get_target(target_id).name
+            except Exception:
+                pass
+            _tgt_name_safe = sanitize_display(_tgt_name) if _tgt_name else None
+            _record_capture(
+                "capture.stopped",
+                target_id,
+                _tgt_name_safe,
+                f"Capture stopped for target '{_tgt_name_safe or target_id}' (bulk)",
+            )
        except ValueError as e:
            errors[target_id] = str(e)
        except Exception as e:
@@ -112,11 +152,19 @@ async def start_processing(
    logger.info("Start processing requested for target %s", target_id)
    try:
        # Verify target exists in store
-        target_store.get_target(target_id)
+        target = target_store.get_target(target_id)

        await manager.start_processing(target_id)

        logger.info(f"Started processing for target {target_id}")
+        _tgt_name_raw2 = getattr(target, "name", None)
+        _tgt_safe2 = sanitize_display(_tgt_name_raw2) if _tgt_name_raw2 else None
+        _record_capture(
+            "capture.started",
+            target_id,
+            _tgt_safe2,
+            f"Capture started for target '{_tgt_safe2 or target_id}'",
+        )
        return {"status": "started", "target_id": target_id}

    except ValueError as e:
@@ -137,6 +185,7 @@ async def start_processing(
 async def stop_processing(
    target_id: str,
    _auth: AuthRequired,
+    target_store: OutputTargetStore = Depends(get_output_target_store),
    manager: ProcessorManager = Depends(get_processor_manager),
 ):
    """Stop processing for a output target."""
@@ -144,6 +193,18 @@ async def stop_processing(
        await manager.stop_processing(target_id)

        logger.info(f"Stopped processing for target {target_id}")
+        _target_name: str | None = None
+        try:
+            _target_name = target_store.get_target(target_id).name
+        except Exception:
+            pass
+        _target_name_safe = sanitize_display(_target_name) if _target_name else None
+        _record_capture(
+            "capture.stopped",
+            target_id,
+            _target_name_safe,
+            f"Capture stopped for target '{_target_name_safe or target_id}'",
+        )
        return {"status": "stopped", "target_id": target_id}

    except ValueError as e:
@@ -374,6 +374,12 @@ async def delete_picture_source(
    css_store: ColorStripStore = Depends(get_color_strip_store),
 ):
    """Delete a picture source."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_stream(stream_id).name
+    except Exception:
+        pass
+
    try:
        # Check if any target transitively references this stream via a CSS
        target_names = store.get_targets_referencing(stream_id, target_store, css_store)
@@ -395,7 +401,7 @@ async def delete_picture_source(
                f"{css_names}. Please reassign or delete those first.",
            )
        store.delete_stream(stream_id)
-        fire_entity_event("picture_source", "deleted", stream_id)
+        fire_entity_event("picture_source", "deleted", stream_id, entity_name=_entity_name)
    except HTTPException:
        raise
    except EntityNotFoundError as e:
@@ -15,6 +15,7 @@ daylight value-source / color-strip-source. Stored as
 empty/missing meaning "use system local time".
 """

+from datetime import datetime, timezone
 from typing import Any

 from fastapi import APIRouter, Body, Depends, HTTPException
@@ -38,6 +39,7 @@ router = APIRouter()
 _DASHBOARD_LAYOUT_KEY = "dashboard_layout"
 _NOTIFICATION_PREFS_KEY = "notification_preferences"
 _CARD_MODES_KEY = "card_modes"
+_ONBOARDING_KEY = "onboarded"


 class DaylightTimezonePreference(BaseModel):
@@ -285,4 +287,75 @@ async def put_daylight_timezone_preference(
    return DaylightTimezonePreference(timezone=saved)


+# ---------------------------------------------------------------------------
+# Onboarding flag
+# ---------------------------------------------------------------------------
+
+
+class OnboardingPreference(BaseModel):
+    """Persistent first-run onboarding flag."""
+
+    onboarded: bool = Field(
+        False,
+        description="True once the user has completed the first-run wizard.",
+    )
+    completed_at: str | None = Field(
+        None,
+        description="ISO timestamp of when onboarding was first marked complete; null otherwise.",
+    )
+
+
+@router.get(
+    "/api/v1/preferences/onboarding",
+    response_model=OnboardingPreference,
+    tags=["Preferences"],
+)
+async def get_onboarding(
+    _: AuthRequired,
+    db: Database = Depends(get_database),
+) -> OnboardingPreference:
+    """Return the first-run onboarding status.
+
+    Defaults to ``{onboarded: false, completed_at: null}`` when the flag has
+    never been set.
+    """
+    raw = db.get_setting(_ONBOARDING_KEY)
+    if not raw:
+        return OnboardingPreference()
+    try:
+        return OnboardingPreference.model_validate(raw)
+    except Exception as exc:
+        logger.warning("Stored onboarding preference invalid (%s); using default", exc)
+        return OnboardingPreference()
+
+
+@router.put(
+    "/api/v1/preferences/onboarding",
+    response_model=OnboardingPreference,
+    tags=["Preferences"],
+)
+async def put_onboarding(
+    _: AuthRequired,
+    body: OnboardingPreference,
+    db: Database = Depends(get_database),
+) -> OnboardingPreference:
+    """Persist the onboarding flag.
+
+    When ``onboarded`` is set to ``true`` and ``completed_at`` is not provided,
+    the server stamps the current UTC time automatically.
+    When ``onboarded`` is ``false``, ``completed_at`` is cleared.
+    """
+    if body.onboarded and body.completed_at is None:
+        body = OnboardingPreference(
+            onboarded=True,
+            completed_at=datetime.now(timezone.utc).isoformat(),
+        )
+    elif not body.onboarded:
+        body = OnboardingPreference(onboarded=False, completed_at=None)
+
+    db.set_setting(_ONBOARDING_KEY, body.model_dump())
+    logger.info("Onboarding flag updated: onboarded=%s", body.onboarded)
+    return body
+
+
 __all__ = ["router", "DAYLIGHT_TIMEZONE_KEY"]
@@ -0,0 +1,328 @@
+"""Scene playlist API routes — CRUD plus start/stop/state cycling control."""
+
+import uuid
+from datetime import datetime, timezone
+
+from fastapi import APIRouter, Depends, HTTPException
+
+from ledgrab.api.auth import AuthRequired
+from ledgrab.core.activity_log.sanitize import sanitize_display
+from ledgrab.api.dependencies import (
+    fire_entity_event,
+    get_playlist_engine,
+    get_scene_playlist_store,
+    get_scene_preset_store,
+)
+from ledgrab.api.schemas.scene_playlists import (
+    PlaylistRuntimeStateSchema,
+    ScenePlaylistCreate,
+    ScenePlaylistListResponse,
+    ScenePlaylistResponse,
+    ScenePlaylistUpdate,
+)
+from ledgrab.core.scenes.playlist_engine import PlaylistEngine, PlaylistError
+from ledgrab.storage.base_store import EntityNotFoundError
+from ledgrab.storage.scene_playlist import PlaylistItem, ScenePlaylist
+from ledgrab.storage.scene_playlist_store import ScenePlaylistStore
+from ledgrab.storage.scene_preset_store import ScenePresetStore
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+router = APIRouter()
+
+
+def _playlist_to_response(playlist: ScenePlaylist, engine: PlaylistEngine) -> ScenePlaylistResponse:
+    return ScenePlaylistResponse(
+        id=playlist.id,
+        name=playlist.name,
+        description=playlist.description,
+        items=[
+            {"scene_preset_id": i.scene_preset_id, "duration_seconds": i.duration_seconds}
+            for i in playlist.items
+        ],
+        loop=playlist.loop,
+        shuffle=playlist.shuffle,
+        order=playlist.order,
+        tags=playlist.tags,
+        icon=getattr(playlist, "icon", "") or "",
+        icon_color=getattr(playlist, "icon_color", "") or "",
+        is_running=engine.get_running_playlist_id() == playlist.id,
+        created_at=playlist.created_at,
+        updated_at=playlist.updated_at,
+    )
+
+
+def _items_from_schema(items) -> list[PlaylistItem]:
+    return [
+        PlaylistItem(scene_preset_id=i.scene_preset_id, duration_seconds=i.duration_seconds)
+        for i in items
+    ]
+
+
+def _validate_preset_refs(items, preset_store: ScenePresetStore) -> None:
+    """Reject playlist items that reference a non-existent scene preset."""
+    for item in items:
+        try:
+            preset_store.get_preset(item.scene_preset_id)
+        except (ValueError, EntityNotFoundError):
+            raise HTTPException(
+                status_code=400,
+                detail=f"Scene preset not found: {item.scene_preset_id}",
+            )
+
+
+# ===== CRUD =====
+
+
+@router.post(
+    "/api/v1/scene-playlists",
+    response_model=ScenePlaylistResponse,
+    tags=["Scene Playlists"],
+    status_code=201,
+)
+async def create_scene_playlist(
+    data: ScenePlaylistCreate,
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    preset_store: ScenePresetStore = Depends(get_scene_preset_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Create a new scene playlist."""
+    _validate_preset_refs(data.items, preset_store)
+
+    now = datetime.now(timezone.utc)
+    playlist = ScenePlaylist(
+        id=f"playlist_{uuid.uuid4().hex[:8]}",
+        name=data.name,
+        description=data.description,
+        items=_items_from_schema(data.items),
+        loop=data.loop,
+        shuffle=data.shuffle,
+        order=store.count(),
+        tags=data.tags if data.tags is not None else [],
+        icon=data.icon or "",
+        icon_color=data.icon_color or "",
+        created_at=now,
+        updated_at=now,
+    )
+
+    try:
+        playlist = store.create_playlist(playlist)
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+
+    fire_entity_event("scene_playlist", "created", playlist.id)
+    return _playlist_to_response(playlist, engine)
+
+
+@router.get(
+    "/api/v1/scene-playlists",
+    response_model=ScenePlaylistListResponse,
+    tags=["Scene Playlists"],
+)
+async def list_scene_playlists(
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """List all scene playlists plus the current cycling state."""
+    playlists = store.get_all_playlists()
+    return ScenePlaylistListResponse(
+        playlists=[_playlist_to_response(p, engine) for p in playlists],
+        count=len(playlists),
+        state=PlaylistRuntimeStateSchema(**engine.get_state()),
+    )
+
+
+# NOTE: the static ``/state`` path is declared before ``/{playlist_id}`` so it
+# is matched first and not swallowed by the path parameter.
+@router.get(
+    "/api/v1/scene-playlists/state",
+    response_model=PlaylistRuntimeStateSchema,
+    tags=["Scene Playlists"],
+)
+async def get_playlist_state(
+    _auth: AuthRequired,
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Get the current playlist cycling state (idle if nothing is running)."""
+    return PlaylistRuntimeStateSchema(**engine.get_state())
+
+
+@router.get(
+    "/api/v1/scene-playlists/{playlist_id}",
+    response_model=ScenePlaylistResponse,
+    tags=["Scene Playlists"],
+)
+async def get_scene_playlist(
+    playlist_id: str,
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Get a single scene playlist."""
+    try:
+        playlist = store.get_playlist(playlist_id)
+    except (ValueError, EntityNotFoundError) as e:
+        raise HTTPException(status_code=404, detail=str(e))
+    return _playlist_to_response(playlist, engine)
+
+
+@router.put(
+    "/api/v1/scene-playlists/{playlist_id}",
+    response_model=ScenePlaylistResponse,
+    tags=["Scene Playlists"],
+)
+async def update_scene_playlist(
+    playlist_id: str,
+    data: ScenePlaylistUpdate,
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    preset_store: ScenePresetStore = Depends(get_scene_preset_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Update a scene playlist's metadata, items, and playback flags."""
+    new_items = None
+    if data.items is not None:
+        _validate_preset_refs(data.items, preset_store)
+        new_items = _items_from_schema(data.items)
+
+    try:
+        playlist = store.update_playlist(
+            playlist_id,
+            name=data.name,
+            description=data.description,
+            items=new_items,
+            loop=data.loop,
+            shuffle=data.shuffle,
+            order=data.order,
+            tags=data.tags,
+            icon=data.icon,
+            icon_color=data.icon_color,
+        )
+    except (ValueError, EntityNotFoundError) as e:
+        raise HTTPException(
+            status_code=404 if "not found" in str(e).lower() else 400, detail=str(e)
+        )
+
+    fire_entity_event("scene_playlist", "updated", playlist_id)
+    return _playlist_to_response(playlist, engine)
+
+
+@router.delete(
+    "/api/v1/scene-playlists/{playlist_id}",
+    status_code=204,
+    tags=["Scene Playlists"],
+)
+async def delete_scene_playlist(
+    playlist_id: str,
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Delete a scene playlist (stops it first if it is currently cycling)."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_playlist(playlist_id).name
+    except Exception:
+        pass
+
+    try:
+        store.delete_playlist(playlist_id)
+    except (ValueError, EntityNotFoundError) as e:
+        raise HTTPException(status_code=404, detail=str(e))
+
+    await engine.stop_if_running(playlist_id)
+    fire_entity_event("scene_playlist", "deleted", playlist_id, entity_name=_entity_name)
+
+
+# ===== Cycling control =====
+
+
+@router.post(
+    "/api/v1/scene-playlists/{playlist_id}/start",
+    response_model=PlaylistRuntimeStateSchema,
+    tags=["Scene Playlists"],
+)
+async def start_scene_playlist(
+    playlist_id: str,
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Start cycling a playlist (stops any currently-running playlist first)."""
+    try:
+        store.get_playlist(playlist_id)
+    except (ValueError, EntityNotFoundError) as e:
+        raise HTTPException(status_code=404, detail=str(e))
+
+    try:
+        await engine.start_playlist(playlist_id)
+    except PlaylistError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+
+    fire_entity_event("scene_playlist", "updated", playlist_id)
+
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+    from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+    rec = get_module_recorder()
+    if rec is not None:
+        _pl_name: str | None = None
+        try:
+            _pl_name = store.get_playlist(playlist_id).name
+        except Exception:
+            pass
+        _safe_pl_name = sanitize_display(_pl_name) if _pl_name else None
+        rec.record(
+            category=ActivityCategory.CAPTURE,
+            action="playlist.started",
+            severity=ActivitySeverity.INFO,
+            entity_type="scene_playlist",
+            entity_id=playlist_id,
+            entity_name=_safe_pl_name,
+            message=f"Playlist '{_safe_pl_name or playlist_id}' started",
+        )
+
+    return PlaylistRuntimeStateSchema(**engine.get_state())
+
+
+@router.post(
+    "/api/v1/scene-playlists/stop",
+    response_model=PlaylistRuntimeStateSchema,
+    tags=["Scene Playlists"],
+)
+async def stop_scene_playlist(
+    _auth: AuthRequired,
+    store: ScenePlaylistStore = Depends(get_scene_playlist_store),
+    engine: PlaylistEngine = Depends(get_playlist_engine),
+):
+    """Stop the active playlist (leaves the last applied scene in place)."""
+    stopped_id = engine.get_running_playlist_id()
+    _stopped_name: str | None = None
+    if stopped_id:
+        try:
+            _stopped_name = store.get_playlist(stopped_id).name
+        except Exception:
+            pass
+    await engine.stop()
+    if stopped_id:
+        fire_entity_event("scene_playlist", "updated", stopped_id)
+
+        from ledgrab.core.activity_log.recorder import get_module_recorder
+        from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+        rec = get_module_recorder()
+        if rec is not None:
+            _safe_stopped_name = sanitize_display(_stopped_name) if _stopped_name else None
+            rec.record(
+                category=ActivityCategory.CAPTURE,
+                action="playlist.stopped",
+                severity=ActivitySeverity.INFO,
+                entity_type="scene_playlist",
+                entity_id=stopped_id,
+                entity_name=_safe_stopped_name,
+                message=f"Playlist '{_safe_stopped_name or stopped_id}' stopped",
+            )
+
+    return PlaylistRuntimeStateSchema(**engine.get_state())
@@ -6,6 +6,7 @@ from datetime import datetime, timezone
 from fastapi import APIRouter, Depends, HTTPException

 from ledgrab.api.auth import AuthRequired
+from ledgrab.core.activity_log.sanitize import sanitize_display
 from ledgrab.api.dependencies import (
    fire_entity_event,
    get_output_target_store,
@@ -208,12 +209,18 @@ async def delete_scene_preset(
    store: ScenePresetStore = Depends(get_scene_preset_store),
 ):
    """Delete a scene preset."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_preset(preset_id).name
+    except Exception:
+        pass
+
    try:
        store.delete_preset(preset_id)
    except ValueError as e:
        raise HTTPException(status_code=404, detail=str(e))

-    fire_entity_event("scene_preset", "deleted", preset_id)
+    fire_entity_event("scene_preset", "deleted", preset_id, entity_name=_entity_name)


 # ===== Recapture =====
@@ -282,4 +289,21 @@ async def activate_scene_preset(
        logger.info(f"Scene preset '{preset.name}' activated successfully")

    fire_entity_event("scene_preset", "updated", preset_id)
+
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+    from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+    rec = get_module_recorder()
+    if rec is not None:
+        _safe_preset_name = sanitize_display(preset.name) if preset.name else None
+        rec.record(
+            category=ActivityCategory.CAPTURE,
+            action="scene.activated",
+            severity=ActivitySeverity.INFO,
+            entity_type="scene_preset",
+            entity_id=preset_id,
+            entity_name=_safe_preset_name,
+            message=f"Scene preset '{_safe_preset_name or preset_id}' activated",
+        )
+
    return ActivateResponse(status=status, errors=errors)
@@ -0,0 +1,330 @@
+"""Setup scaffold endpoint.
+
+Wires a complete capture → color-strip → output chain in one call, with
+automatic rollback if any step fails so no orphan entities are left behind.
+
+POST /api/v1/setup/scaffold
+    Body: ScaffoldRequest — device_id (required, must already exist),
+          display_index, optional calibration dict.
+    Returns: ScaffoldResponse — ids of every created/reused entity.
+    Fires ``entity_changed`` events for every entity created in this call,
+    but ONLY after the full chain succeeds (no mid-chain events).
+    Does NOT auto-start the target (the frontend starts it after calibration).
+
+Rollback contract
+-----------------
+Entities created during THIS request are tracked in a local list.  If any
+step raises, they are deleted in reverse-creation order before re-raising.
+Because "created" events are deferred until after the chain completes, a
+rollback never produces ghost UI cards — no event for a rolled-back entity
+is ever emitted.
+
+The device is never part of the rollback set: scaffold requires an existing
+device (created via ``POST /api/v1/devices`` which runs full validation).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from fastapi import APIRouter, Depends, HTTPException
+
+from ledgrab.api.auth import AuthRequired
+from ledgrab.api.dependencies import (
+    fire_entity_event,
+    get_color_strip_store,
+    get_device_store,
+    get_output_target_store,
+    get_picture_source_store,
+    get_processor_manager,
+    get_template_store,
+)
+from ledgrab.api.schemas.setup import ScaffoldRequest, ScaffoldResponse
+from ledgrab.core.capture.calibration import calibration_from_dict, create_default_calibration
+from ledgrab.core.capture_engines.factory import EngineRegistry
+from ledgrab.core.processing.processor_manager import ProcessorManager
+from ledgrab.storage.base_store import EntityNotFoundError
+from ledgrab.storage.color_strip_store import ColorStripStore
+from ledgrab.storage.output_target_store import OutputTargetStore
+from ledgrab.storage.picture_source_store import PictureSourceStore
+from ledgrab.storage import DeviceStore
+from ledgrab.storage.template_store import TemplateStore
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+router = APIRouter()
+
+_DEFAULT_TARGET_FPS = 30
+
+
+# ---------------------------------------------------------------------------
+# Helper: capture template
+# ---------------------------------------------------------------------------
+
+
+def _get_or_create_capture_template(
+    template_store: TemplateStore,
+    created_ids: list[tuple[str, str]],
+) -> tuple[str, bool]:
+    """Return (template_id, reused).
+
+    Tries to find an existing template whose engine_type matches the platform's
+    best available engine.  Falls back to creating a fresh one.
+    """
+    best_engine = EngineRegistry.get_best_available_engine()
+    if not best_engine:
+        raise HTTPException(
+            status_code=503,
+            detail="No capture engine available on this platform; cannot scaffold.",
+        )
+
+    # Try to reuse an existing template with the same engine
+    for tpl in template_store.get_all_templates():
+        if tpl.engine_type == best_engine:
+            logger.info(
+                "Scaffold: reusing existing capture template %s (engine=%s)",
+                tpl.id,
+                best_engine,
+            )
+            return tpl.id, True
+
+    # None found — create a fresh one
+    engine_class = EngineRegistry.get_engine(best_engine)
+    default_config = engine_class.get_default_config()
+    try:
+        tpl = template_store.create_template(
+            name=f"Scaffold capture ({best_engine})",
+            engine_type=best_engine,
+            engine_config=default_config,
+            description="Auto-created by first-run scaffold",
+        )
+    except ValueError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+
+    created_ids.append(("capture_template", tpl.id))
+    logger.info("Scaffold: created capture template %s (engine=%s)", tpl.id, best_engine)
+    return tpl.id, False
+
+
+# ---------------------------------------------------------------------------
+# Helper: rollback
+# ---------------------------------------------------------------------------
+
+
+def _rollback(
+    created_ids: list[tuple[str, str]],
+    *,
+    template_store: TemplateStore,
+    picture_source_store: PictureSourceStore,
+    css_store: ColorStripStore,
+    output_target_store: OutputTargetStore,
+    manager: ProcessorManager | None = None,
+) -> None:
+    """Delete entities created during this call, in reverse order.
+
+    Only entities listed in ``created_ids`` are deleted; reused/pre-existing
+    entities (including the device) are never touched.
+
+    If *manager* is provided, any ``output_target`` entity in the rollback set
+    is also unregistered from the ProcessorManager before store deletion, so no
+    half-registered target is left behind.
+    """
+    store_map: dict[str, Any] = {
+        "capture_template": template_store,
+        "picture_source": picture_source_store,
+        "color_strip_source": css_store,
+        "output_target": output_target_store,
+    }
+    for entity_type, entity_id in reversed(created_ids):
+        # Unregister output targets from the processor manager first
+        if entity_type == "output_target" and manager is not None:
+            try:
+                manager.remove_target(entity_id)
+                logger.info("Scaffold rollback: unregistered target %s from manager", entity_id)
+            except (ValueError, RuntimeError) as exc:
+                logger.debug(
+                    "Scaffold rollback: manager unregister skipped for %s — %s",
+                    entity_id,
+                    exc,
+                )
+
+        store = store_map.get(entity_type)
+        if store is None:
+            logger.warning("Scaffold rollback: unknown entity type %r — skipping", entity_type)
+            continue
+        try:
+            store.delete(entity_id)
+            logger.info("Scaffold rollback: deleted %s %s", entity_type, entity_id)
+        except Exception as exc:
+            logger.error(
+                "Scaffold rollback: failed to delete %s %s — %s",
+                entity_type,
+                entity_id,
+                exc,
+            )
+
+
+# ---------------------------------------------------------------------------
+# Route
+# ---------------------------------------------------------------------------
+
+
+@router.post(
+    "/api/v1/setup/scaffold",
+    response_model=ScaffoldResponse,
+    status_code=201,
+    tags=["Setup"],
+)
+async def scaffold_setup(
+    data: ScaffoldRequest,
+    _auth: AuthRequired,
+    device_store: DeviceStore = Depends(get_device_store),
+    template_store: TemplateStore = Depends(get_template_store),
+    picture_source_store: PictureSourceStore = Depends(get_picture_source_store),
+    css_store: ColorStripStore = Depends(get_color_strip_store),
+    output_target_store: OutputTargetStore = Depends(get_output_target_store),
+    manager: ProcessorManager = Depends(get_processor_manager),
+) -> ScaffoldResponse:
+    """Create a ready-to-start LED capture chain.
+
+    Steps (each uses the real store create method for validation and ID gen):
+
+    1. Look up the existing device (404 if not found).
+    2. Find or create a capture template for the platform-best engine.
+    3. Create a raw picture source (``display_index`` + ``capture_template_id``).
+    4. Create a picture color-strip source with either the provided calibration
+       or ``create_default_calibration(led_count)``.
+    5. Create a LED output target linking the device to the CSS.
+
+    All created entities emit ``entity_changed`` events, but ONLY after the
+    full chain succeeds — events are collected and fired at the very end.
+    On any error the entities created so far are deleted in reverse order
+    (rollback), and no "created" events are emitted (no ghost UI cards).
+    The output target is NOT started — the frontend starts it after the
+    optional calibration step.
+    """
+    created_ids: list[tuple[str, str]] = []
+    # Deferred "created" events: (entity_type, entity_id) — fired only on success.
+    pending_events: list[tuple[str, str]] = []
+
+    rollback_stores = dict(
+        template_store=template_store,
+        picture_source_store=picture_source_store,
+        css_store=css_store,
+        output_target_store=output_target_store,
+        manager=manager,
+    )
+
+    try:
+        # ── Step 1: resolve existing device ─────────────────────────────────
+        try:
+            device = device_store.get(data.device_id)
+        except EntityNotFoundError:
+            raise HTTPException(status_code=404, detail=f"Device not found: {data.device_id}")
+        device_id = device.id
+        led_count = device.led_count
+
+        # ── Step 2: capture template ─────────────────────────────────────────
+        capture_template_id, template_reused = _get_or_create_capture_template(
+            template_store, created_ids
+        )
+        if not template_reused:
+            pending_events.append(("capture_template", capture_template_id))
+
+        # ── Step 3: picture source ───────────────────────────────────────────
+        ps_name = f"Screen {data.display_index} (scaffold)"
+        try:
+            picture_source = picture_source_store.create_stream(
+                name=ps_name,
+                stream_type="raw",
+                display_index=data.display_index,
+                capture_template_id=capture_template_id,
+                target_fps=_DEFAULT_TARGET_FPS,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc))
+        created_ids.append(("picture_source", picture_source.id))
+        pending_events.append(("picture_source", picture_source.id))
+        logger.info("Scaffold: created picture source %s", picture_source.id)
+
+        # ── Step 4: color-strip source ───────────────────────────────────────
+        if data.calibration is not None:
+            try:
+                calibration = calibration_from_dict(data.calibration)
+            except Exception as exc:
+                raise HTTPException(
+                    status_code=422,
+                    detail=f"Invalid calibration dict: {exc}",
+                )
+        else:
+            try:
+                calibration = create_default_calibration(led_count)
+            except ValueError as exc:
+                raise HTTPException(status_code=400, detail=str(exc))
+
+        css_name = "Screen capture (scaffold)"
+        try:
+            css = css_store.create_source(
+                name=css_name,
+                source_type="picture",
+                picture_source_id=picture_source.id,
+                calibration=calibration,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc))
+        created_ids.append(("color_strip_source", css.id))
+        pending_events.append(("color_strip_source", css.id))
+        logger.info("Scaffold: created color-strip source %s", css.id)
+
+        # ── Step 5: LED output target ────────────────────────────────────────
+        target_name = "LED output (scaffold)"
+        try:
+            target = output_target_store.create_wled_target(
+                name=target_name,
+                device_id=device_id,
+                color_strip_source_id=css.id,
+            )
+        except ValueError as exc:
+            raise HTTPException(status_code=400, detail=str(exc))
+        created_ids.append(("output_target", target.id))
+        pending_events.append(("output_target", target.id))
+        logger.info("Scaffold: created output target %s", target.id)
+
+        # ── Step 5b: register target with ProcessorManager ───────────────────
+        try:
+            target.register_with_manager(manager)
+        except ValueError as exc:
+            logger.warning(
+                "Scaffold: could not register target %s in processor manager: %s",
+                target.id,
+                exc,
+            )
+
+    except HTTPException:
+        _rollback(created_ids, **rollback_stores)
+        raise
+    except Exception as exc:
+        logger.error("Scaffold: unexpected error — rolling back: %s", exc, exc_info=True)
+        _rollback(created_ids, **rollback_stores)
+        raise HTTPException(status_code=500, detail="Internal server error during scaffold")
+
+    # ── Full chain succeeded — fire all deferred "created" events ───────────
+    for entity_type, entity_id in pending_events:
+        fire_entity_event(entity_type, "created", entity_id)
+
+    logger.info(
+        "Scaffold complete: device=%s tpl=%s ps=%s css=%s target=%s",
+        device_id,
+        capture_template_id,
+        picture_source.id,
+        css.id,
+        target.id,
+    )
+    return ScaffoldResponse(
+        device_id=device_id,
+        capture_template_id=capture_template_id,
+        picture_source_id=picture_source.id,
+        color_strip_source_id=css.id,
+        output_target_id=target.id,
+        capture_template_reused=template_reused,
+    )
@@ -30,7 +30,9 @@ from ledgrab.api.dependencies import (
    get_color_strip_store,
    get_device_store,
    get_output_target_store,
+    get_playlist_engine,
    get_processor_manager,
+    get_scene_playlist_store,
    get_scene_preset_store,
    get_sync_clock_manager,
    get_sync_clock_store,
@@ -43,6 +45,7 @@ from ledgrab.utils import get_logger
 from .color_strip_sources.crud import list_color_strip_sources
 from .devices import list_devices, resolve_device_brightness
 from .output_targets import batch_target_metrics, batch_target_states, list_targets
+from .scene_playlists import list_scene_playlists
 from .scene_presets import list_scene_presets
 from .sync_clocks import list_sync_clocks
 from .system import get_system_performance, health_check
@@ -53,7 +56,9 @@ logger = get_logger(__name__)

 router = APIRouter()

-# Selectable snapshot sections — these are exactly the response top-level keys.
+# Selectable snapshot sections — these are exactly the response top-level keys,
+# except ``scene_playlists`` which also emits a companion ``playlist_state`` key
+# (the single global cycling state; see the handler).
 SNAPSHOT_SECTIONS = (
    "targets",
    "target_states",
@@ -63,6 +68,7 @@ SNAPSHOT_SECTIONS = (
    "css_sources",
    "value_sources",
    "scene_presets",
+    "scene_playlists",
    "sync_clocks",
    "system",
 )
@@ -135,6 +141,8 @@ async def get_snapshot(
    css_store=Depends(get_color_strip_store),
    value_store=Depends(get_value_source_store),
    preset_store=Depends(get_scene_preset_store),
+    playlist_store=Depends(get_scene_playlist_store),
+    playlist_engine=Depends(get_playlist_engine),
    clock_store=Depends(get_sync_clock_store),
    clock_manager=Depends(get_sync_clock_manager),
    update_service=Depends(get_update_service),
@@ -152,6 +160,8 @@ async def get_snapshot(
          "css_sources": [...],
          "value_sources": [...],
          "scene_presets": [...],
+          "scene_playlists": [...],
+          "playlist_state": {...},   # companion to scene_playlists
          "sync_clocks": [...],
          "system": {"performance": {...}, "health": {...}, "update": {...}}
        }
@@ -184,6 +194,14 @@ async def get_snapshot(
        result["value_sources"] = (await list_value_sources(_auth, None, value_store)).sources
    if "scene_presets" in sections:
        result["scene_presets"] = (await list_scene_presets(_auth, preset_store)).presets
+    if "scene_playlists" in sections:
+        # One call returns both the playlist list (each with ``is_running``) and
+        # the single global cycling state (current index / preset / dwell). The
+        # state is emitted as a companion top-level key because it describes the
+        # one running playlist, not any individual list entry.
+        playlists = await list_scene_playlists(_auth, playlist_store, playlist_engine)
+        result["scene_playlists"] = playlists.playlists
+        result["playlist_state"] = playlists.state
    if "sync_clocks" in sections:
        clocks = await list_sync_clocks(_auth, clock_store, clock_manager)
        result["sync_clocks"] = clocks.clocks
@@ -149,6 +149,12 @@ async def delete_sync_clock(
    manager: SyncClockManager = Depends(get_sync_clock_manager),
 ):
    """Delete a synchronization clock (fails if referenced by CSS or value sources)."""
+    _entity_name: str | None = None
+    try:
+        _entity_name = store.get_clock(clock_id).name
+    except Exception:
+        pass
+
    try:
        # Check references
        for source in css_store.get_all_sources():
@@ -159,7 +165,7 @@ async def delete_sync_clock(
                raise ValueError(f"Cannot delete: referenced by value source '{vs.name}'")
        manager.release_all_for(clock_id)
        store.delete_clock(clock_id)
-        fire_entity_event("sync_clock", "deleted", clock_id)
+        fire_entity_event("sync_clock", "deleted", clock_id, entity_name=_entity_name)
    except EntityNotFoundError as e:
        raise HTTPException(status_code=404, detail=str(e))

@@ -21,11 +21,29 @@ from ledgrab.api.schemas.system import (
    ShutdownActionRequest,
    ShutdownActionResponse,
 )
+from ledgrab.core.activity_log.sanitize import sanitize_display
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.storage.database import Database
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)

+
+def _record_setting(action: str, key: str, message: str) -> None:
+    """Best-effort audit record for a high-value settings change."""
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action=action,
+            severity=ActivitySeverity.INFO,
+            message=message,
+            metadata={"setting_key": key},
+        )
+
+
 router = APIRouter()


@@ -117,6 +135,11 @@ async def update_shutdown_action(
    """Set what happens to LED targets when the server shuts down."""
    db.set_setting("shutdown_action", {"action": body.action})
    logger.info("Shutdown action updated: %s", body.action)
+    _record_setting(
+        "settings.changed",
+        "shutdown_action",
+        f"Shutdown action set to '{body.action}'",
+    )
    return ShutdownActionResponse(action=body.action)


@@ -246,6 +269,17 @@ async def adb_connect(_: AuthRequired, request: AdbConnectRequest):
        stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=10)
        output = (stdout.decode() + stderr.decode()).strip()
        if "connected" in output.lower():
+            from ledgrab.core.activity_log.recorder import get_module_recorder
+
+            rec = get_module_recorder()
+            if rec is not None:
+                rec.record(
+                    category=ActivityCategory.DEVICE,
+                    action="device.adb_connected",
+                    severity=ActivitySeverity.INFO,
+                    message=f"ADB device connected: {sanitize_display(address)}",
+                    metadata={"address": address},
+                )
            return {"status": "connected", "address": address, "message": output}
        raise HTTPException(status_code=400, detail=output or "Connection failed")
    except FileNotFoundError:
@@ -276,6 +310,17 @@ async def adb_disconnect(_: AuthRequired, request: AdbConnectRequest):
            stderr=asyncio.subprocess.PIPE,
        )
        stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=10)
+        from ledgrab.core.activity_log.recorder import get_module_recorder
+
+        rec = get_module_recorder()
+        if rec is not None:
+            rec.record(
+                category=ActivityCategory.DEVICE,
+                action="device.adb_disconnected",
+                severity=ActivitySeverity.INFO,
+                message=f"ADB device disconnected: {sanitize_display(address)}",
+                metadata={"address": address},
+            )
        return {"status": "disconnected", "message": stdout.decode().strip()}
    except FileNotFoundError:
        raise HTTPException(status_code=500, detail="adb not found on PATH")
@@ -183,6 +183,12 @@ async def delete_template(

    Validates that no streams are currently using this template before deletion.
    """
+    _entity_name: str | None = None
+    try:
+        _entity_name = template_store.get_template(template_id).name
+    except Exception:
+        pass
+
    try:
        # Check if any streams are using this template
        streams_using_template = []
@@ -203,7 +209,7 @@ async def delete_template(

        # Proceed with deletion
        template_store.delete_template(template_id)
-        fire_entity_event("capture_template", "deleted", template_id)
+        fire_entity_event("capture_template", "deleted", template_id, entity_name=_entity_name)

    except HTTPException:
        raise  # Re-raise HTTP exceptions as-is
@@ -12,6 +12,7 @@ from ledgrab.api.schemas.update import (
    UpdateStatusResponse,
 )
 from ledgrab.core.update.update_service import UpdateService
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)
@@ -42,6 +43,17 @@ async def dismiss_update(
    service: UpdateService = Depends(get_update_service),
 ):
    service.dismiss(body.version)
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action="update.dismissed",
+            severity=ActivitySeverity.INFO,
+            message=f"Update dismissed: {body.version}",
+            metadata={"version": body.version},
+        )
    return {"ok": True}


@@ -63,6 +75,18 @@ async def apply_update(
        )
    try:
        await service.apply_update()
+        from ledgrab.core.activity_log.recorder import get_module_recorder
+
+        rec = get_module_recorder()
+        if rec is not None:
+            version = status.get("available_version", "unknown")
+            rec.record(
+                category=ActivityCategory.SYSTEM,
+                action="update.applied",
+                severity=ActivitySeverity.INFO,
+                message=f"Update applied: {version}",
+                metadata={"version": version},
+            )
        return {"ok": True, "message": "Update applied, server shutting down"}
    except Exception as exc:
        logger.error("Failed to apply update: %s", exc, exc_info=True)
@@ -83,8 +107,20 @@ async def update_update_settings(
    body: UpdateSettingsRequest,
    service: UpdateService = Depends(get_update_service),
 ):
-    return await service.update_settings(
+    result = await service.update_settings(
        enabled=body.enabled,
        check_interval_hours=body.check_interval_hours,
        include_prerelease=body.include_prerelease,
    )
+    from ledgrab.core.activity_log.recorder import get_module_recorder
+
+    rec = get_module_recorder()
+    if rec is not None:
+        rec.record(
+            category=ActivityCategory.SYSTEM,
+            action="settings.changed",
+            severity=ActivitySeverity.INFO,
+            message=f"Update settings changed (enabled={body.enabled})",
+            metadata={"setting_key": "update", "enabled": body.enabled},
+        )
+    return result
@@ -0,0 +1,93 @@
+"""Pydantic schemas for the activity-log API (Phase 4)."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+
+# ---------------------------------------------------------------------------
+# Entry + page response
+# ---------------------------------------------------------------------------
+
+
+class ActivityLogEntryResponse(BaseModel):
+    """Single audit-log entry.
+
+    Shape matches ``entry_to_dict()`` from
+    ``ledgrab.core.activity_log.recorder`` exactly — that function is the
+    single source of truth for serialisation; this schema documents the wire
+    format.
+    """
+
+    id: str = Field(description="Entry id — 'al_<8-hex>'")
+    ts: str = Field(description="ISO-8601 UTC timestamp")
+    category: str = Field(description="Broad bucket (auth, device, entity, capture, system)")
+    action: str = Field(description="Verb-object label, e.g. 'entity.created'")
+    severity: str = Field(description="info | warning | error")
+    actor: str = Field(description="API-key label or 'system' / 'anonymous'")
+    entity_type: str | None = Field(default=None, description="Affected entity type, if applicable")
+    entity_id: str | None = Field(default=None, description="Affected entity id, if applicable")
+    entity_name: str | None = Field(
+        default=None, description="Entity name at time of event, if applicable"
+    )
+    message: str = Field(description="Human-readable description")
+    metadata: dict[str, Any] = Field(default_factory=dict, description="Extra structured context")
+
+
+class ActivityLogPageResponse(BaseModel):
+    """Paginated list of audit-log entries (keyset cursor)."""
+
+    entries: list[ActivityLogEntryResponse] = Field(description="Entries on this page")
+    next_before_seq: int | None = Field(
+        default=None,
+        description=(
+            "Pass as 'before_seq' in the next request to get the following page. "
+            "None when this is the last page."
+        ),
+    )
+    has_more: bool = Field(
+        description="True when there are more entries before the first entry on this page"
+    )
+    total: int = Field(description="Total entries matching the current filters (all pages)")
+
+
+# ---------------------------------------------------------------------------
+# Settings
+# ---------------------------------------------------------------------------
+
+_MAX_DAYS_CAP = 3650  # 10 years — sanity upper bound
+_MAX_ENTRIES_CAP = 10_000_000  # 10 M rows — sanity upper bound
+
+
+class ActivityLogSettingsResponse(BaseModel):
+    """Current activity-log retention settings."""
+
+    enabled: bool = Field(description="Whether the activity log is recording")
+    max_days: int = Field(
+        ge=0,
+        le=_MAX_DAYS_CAP,
+        description="Retain entries for at most this many days (0 = no age-based pruning)",
+    )
+    max_entries: int = Field(
+        ge=0,
+        le=_MAX_ENTRIES_CAP,
+        description="Keep at most this many entries (0 = no count-based pruning)",
+    )
+
+
+class UpdateActivityLogSettingsRequest(BaseModel):
+    """Request body for PUT /settings."""
+
+    enabled: bool = Field(description="Enable or disable activity-log recording")
+    max_days: int = Field(
+        ge=0,
+        le=_MAX_DAYS_CAP,
+        description="Retain entries for at most this many days (0 = no age-based pruning)",
+    )
+    max_entries: int = Field(
+        ge=0,
+        le=_MAX_ENTRIES_CAP,
+        description="Keep at most this many entries (0 = no count-based pruning)",
+    )
@@ -36,7 +36,29 @@ class RuleSchema(BaseModel):
    )
    timezone: str | None = Field(
        None,
-        description="IANA timezone for time_of_day rule (e.g. 'Europe/Berlin'). Empty = server local.",
+        description=(
+            "IANA timezone for time_of_day / solar rules (e.g. 'Europe/Berlin'). "
+            "Empty = server local."
+        ),
+    )
+    # Solar rule fields (days_of_week / timezone above are shared with time_of_day)
+    start_event: str | None = Field(
+        None, description="'sunrise' or 'sunset' — window start anchor (for solar rule)"
+    )
+    start_offset_minutes: int | None = Field(
+        None, description="Minutes added to the start event, ±1439 (for solar rule)"
+    )
+    end_event: str | None = Field(
+        None, description="'sunrise' or 'sunset' — window end anchor (for solar rule)"
+    )
+    end_offset_minutes: int | None = Field(
+        None, description="Minutes added to the end event, ±1439 (for solar rule)"
+    )
+    latitude: float | None = Field(
+        None, description="Latitude for solar timing, -90..90 (for solar rule)"
+    )
+    longitude: float | None = Field(
+        None, description="Longitude for solar timing, -180..180 (for solar rule)"
    )
    # System idle rule fields
    idle_minutes: int | None = Field(
@@ -86,6 +108,37 @@ class RuleSchema(BaseModel):
 ConditionSchema = RuleSchema


+class ActionSchema(BaseModel):
+    """A single outbound action fired alongside scene activation/deactivation."""
+
+    action_type: str = Field(
+        max_length=32, description="Action type discriminator (e.g. 'webhook')"
+    )
+    # Webhook action fields
+    webhook_url: str | None = Field(
+        None, max_length=2048, description="Target URL for the webhook action"
+    )
+    method: str | None = Field(
+        None, max_length=8, description="'POST', 'PUT', or 'GET' (for webhook action)"
+    )
+    body_template: str | None = Field(
+        None,
+        max_length=8192,
+        description=(
+            "Request body template (for webhook action). Tokens: {{automation_name}}, "
+            "{{automation_id}}, {{event}}, {{timestamp}}."
+        ),
+    )
+    content_type: str | None = Field(
+        None,
+        max_length=128,
+        description="Content-Type header for the webhook body (default application/json)",
+    )
+    fire_on: str | None = Field(
+        None, max_length=16, description="'activate', 'deactivate', or 'both' (for webhook action)"
+    )
+
+
 class AutomationCreate(BaseModel):
    """Request to create an automation."""

@@ -101,6 +154,9 @@ class AutomationCreate(BaseModel):
        None, description="Scene preset for fallback deactivation"
    )
    tags: List[str] = Field(default_factory=list, description="User-defined tags")
+    actions: List[ActionSchema] = Field(
+        default_factory=list, description="Outbound actions (e.g. webhooks)"
+    )
    icon: str | None = Field(
        None,
        max_length=64,
@@ -126,6 +182,7 @@ class AutomationUpdate(BaseModel):
        None, description="Scene preset for fallback deactivation"
    )
    tags: List[str] | None = None
+    actions: List[ActionSchema] | None = Field(None, description="Outbound actions (e.g. webhooks)")
    icon: str | None = Field(
        None,
        max_length=64,
@@ -150,6 +207,9 @@ class AutomationResponse(BaseModel):
    deactivation_mode: str = Field(default="none", description="Deactivation behavior")
    deactivation_scene_preset_id: str | None = Field(None, description="Fallback scene preset")
    tags: List[str] = Field(default_factory=list, description="User-defined tags")
+    actions: List[ActionSchema] = Field(
+        default_factory=list, description="Outbound actions (e.g. webhooks)"
+    )
    webhook_url: str | None = Field(
        None, description="Webhook URL for the first webhook rule (if any)"
    )
@@ -179,3 +239,15 @@ class AutomationListResponse(BaseModel):

    automations: List[AutomationResponse] = Field(description="List of automations")
    count: int = Field(description="Number of automations")
+
+
+class AutomationTriggerResponse(BaseModel):
+    """Result of manually triggering an automation."""
+
+    status: str = Field(
+        description="'triggered' (scene applied / nothing to apply), 'partial' "
+        "(applied with errors), 'skipped' (rules not satisfied), or 'error'."
+    )
+    errors: List[str] = Field(
+        default_factory=list, description="Per-target error messages, if any."
+    )
@@ -0,0 +1,104 @@
+"""Pydantic schemas for the calibration session and solver API."""
+
+from typing import Annotated, List, Literal
+
+from pydantic import BaseModel, Field, model_validator
+
+
+# ── Session lifecycle ─────────────────────────────────────────────────────────
+
+
+class CalibrationSessionStartRequest(BaseModel):
+    """Request to start a calibration session on a device."""
+
+    device_id: str = Field(description="ID of the device to drive during calibration")
+
+
+class CalibrationSessionPositionRequest(BaseModel):
+    """Request to advance the chase pixel to a specific LED index."""
+
+    index: int = Field(ge=0, description="LED index to illuminate (0-based)")
+    window: int = Field(
+        default=1,
+        ge=0,
+        le=10,
+        description="Number of dim neighbour LEDs to show on each side (0 = centre only)",
+    )
+
+
+class CalibrationSessionStateResponse(BaseModel):
+    """Current calibration session state."""
+
+    active: bool = Field(description="Whether a session is currently active")
+    device_id: str | None = Field(None, description="Device being driven (null if inactive)")
+    led_count: int = Field(0, description="LED count of the active device")
+    prior_target_id: str | None = Field(
+        None, description="Target that was running before the session (will be restored on stop)"
+    )
+    last_activity: str | None = Field(
+        None, description="ISO timestamp of the last position call (null if inactive)"
+    )
+
+
+# ── Solver ────────────────────────────────────────────────────────────────────
+
+
+class CalibrationSolveRequest(BaseModel):
+    """Request to solve a CalibrationConfig from 4 corner tap indices.
+
+    Provide either *device_id* (the server derives led_count from the device)
+    or *led_count* directly.  *device_id* takes precedence.
+    """
+
+    device_id: str | None = Field(
+        None,
+        description=("Device ID to derive led_count from (preferred over led_count field)"),
+    )
+    led_count: int | None = Field(
+        None,
+        ge=1,
+        description="Total LED count (used when device_id is not provided)",
+    )
+    start_position: Literal["top_left", "top_right", "bottom_left", "bottom_right"] = Field(
+        description="Starting corner of the strip"
+    )
+    layout: Literal["clockwise", "counterclockwise"] = Field(
+        description="Winding direction of the strip"
+    )
+    corner_indices: List[Annotated[int, Field(ge=0)]] = Field(
+        description=(
+            "Four strip indices — one per screen corner — in the strip-walk order "
+            "defined by (start_position, layout).  Index 0 of the strip is the "
+            "start corner; the four tap positions are recorded in strip order "
+            "beginning from that start corner (the solver lays edges out from "
+            "led_start=0, so a non-zero physical start would require the `offset` "
+            "field rather than a shifted corner_indices[0]).  Each element must be "
+            "non-negative (ge=0); out-of-range values yield a 422."
+        ),
+        min_length=4,
+        max_length=4,
+    )
+    offset: int = Field(
+        default=0,
+        ge=0,
+        description="Physical LED offset (0 = no offset)",
+    )
+
+    @model_validator(mode="after")
+    def _require_device_or_led_count(self) -> "CalibrationSolveRequest":
+        if self.device_id is None and self.led_count is None:
+            raise ValueError("Either 'device_id' or 'led_count' must be provided")
+        return self
+
+
+class CalibrationSolvedResponse(BaseModel):
+    """Solved calibration config in simple-mode dict form."""
+
+    mode: Literal["simple"] = "simple"
+    layout: str = Field(description="Winding direction")
+    start_position: str = Field(description="Starting corner")
+    leds_top: int = Field(ge=0, description="LEDs on the top edge")
+    leds_right: int = Field(ge=0, description="LEDs on the right edge")
+    leds_bottom: int = Field(ge=0, description="LEDs on the bottom edge")
+    leds_left: int = Field(ge=0, description="LEDs on the left edge")
+    offset: int = Field(ge=0, description="Physical LED offset")
@@ -145,6 +145,10 @@ class EffectCSSResponse(_CSSResponseBase):
    scale: Any = Field(description="Spatial scale")
    mirror: bool = Field(description="Mirror/bounce mode")
    custom_palette: List[List[float]] | None = Field(None, description="Custom palette stops")
+    audio_reactive: bool = Field(False, description="Modulate output by live audio loudness")
+    reactive_audio_source_id: str = Field("", description="AudioSource id driving reactivity")
+    reactive_mode: str = Field("brightness", description="brightness | saturation | both")
+    reactive_intensity: Any = Field(None, description="Reactive modulation strength (0-1)")


 class CompositeCSSResponse(_CSSResponseBase):
@@ -332,6 +336,12 @@ class EffectCSSCreate(_CSSCreateBase):
    scale: Any = Field(default=None, description="Spatial scale (0.5-5.0)")
    mirror: bool | None = Field(None, description="Mirror/bounce mode")
    custom_palette: List[List[float]] | None = Field(None, description="Custom palette stops")
+    audio_reactive: bool | None = Field(None, description="Modulate output by live audio loudness")
+    reactive_audio_source_id: str | None = Field(None, description="AudioSource id for reactivity")
+    reactive_mode: Literal["brightness", "saturation", "both"] | None = Field(
+        None, description="brightness | saturation | both"
+    )
+    reactive_intensity: Any = Field(default=None, description="Reactive modulation strength (0-1)")


 class CompositeCSSCreate(_CSSCreateBase):
@@ -532,6 +542,12 @@ class EffectCSSUpdate(_CSSUpdateBase):
    scale: Any = Field(default=None, description="Spatial scale (0.5-5.0)")
    mirror: bool | None = Field(None, description="Mirror/bounce mode")
    custom_palette: List[List[float]] | None = Field(None, description="Custom palette stops")
+    audio_reactive: bool | None = Field(None, description="Modulate output by live audio loudness")
+    reactive_audio_source_id: str | None = Field(None, description="AudioSource id for reactivity")
+    reactive_mode: Literal["brightness", "saturation", "both"] | None = Field(
+        None, description="brightness | saturation | both"
+    )
+    reactive_intensity: Any = Field(default=None, description="Reactive modulation strength (0-1)")


 class CompositeCSSUpdate(_CSSUpdateBase):
@@ -59,6 +59,10 @@ class DeviceCreate(BaseModel):
    hue_entertainment_group_id: str | None = Field(
        None, description="Hue entertainment group/zone ID"
    )
+    hue_gradient_mode: bool | None = Field(
+        None,
+        description="Map the strip across gradient-lightstrip channels vs one record per light",
+    )
    # Yeelight fields
    yeelight_min_interval_ms: int | None = Field(
        None,
@@ -80,6 +84,10 @@ class DeviceCreate(BaseModel):
        le=10000,
        description="LIFX client-side rate limit between commands in ms (default 50)",
    )
+    lifx_per_zone: bool | None = Field(
+        None,
+        description="Stream individual zones/tiles (multizone Z/Beam, Tile/Canvas) vs single colour",
+    )
    # Govee fields
    govee_min_interval_ms: int | None = Field(
        None,
@@ -106,6 +114,9 @@ class DeviceCreate(BaseModel):
        le=10000,
        description="Nanoleaf client-side rate limit between commands in ms (default 100)",
    )
+    nanoleaf_per_panel: bool | None = Field(
+        None, description="Stream each panel individually via extControl UDP (vs single colour)"
+    )
    # SPI Direct fields
    spi_speed_hz: int | None = Field(
        None, ge=100000, le=4000000, description="SPI clock speed in Hz"
@@ -195,6 +206,10 @@ class DeviceUpdate(BaseModel):
    hue_username: str | None = Field(None, description="Hue bridge username")
    hue_client_key: str | None = Field(None, description="Hue entertainment client key")
    hue_entertainment_group_id: str | None = Field(None, description="Hue entertainment group ID")
+    hue_gradient_mode: bool | None = Field(
+        None,
+        description="Map the strip across gradient-lightstrip channels vs one record per light",
+    )
    yeelight_min_interval_ms: int | None = Field(
        None, ge=0, le=10000, description="Yeelight client-side rate limit in ms"
    )
@@ -204,6 +219,9 @@ class DeviceUpdate(BaseModel):
    lifx_min_interval_ms: int | None = Field(
        None, ge=0, le=10000, description="LIFX client-side rate limit in ms"
    )
+    lifx_per_zone: bool | None = Field(
+        None, description="Stream individual zones/tiles (multizone/matrix) vs single colour"
+    )
    govee_min_interval_ms: int | None = Field(
        None, ge=0, le=10000, description="Govee client-side rate limit in ms"
    )
@@ -212,6 +230,9 @@ class DeviceUpdate(BaseModel):
    nanoleaf_min_interval_ms: int | None = Field(
        None, ge=0, le=10000, description="Nanoleaf client-side rate limit in ms"
    )
+    nanoleaf_per_panel: bool | None = Field(
+        None, description="Stream each panel individually via extControl UDP"
+    )
    spi_speed_hz: int | None = Field(None, ge=100000, le=4000000, description="SPI clock speed")
    spi_led_type: str | None = Field(None, description="LED chipset type")
    chroma_device_type: str | None = Field(None, description="Chroma peripheral type")
@@ -344,6 +365,26 @@ class Calibration(BaseModel):
    border_width: int = Field(
        default=10, ge=1, le=100, description="Border width in pixels for edge sampling"
    )
+    roi_x: float = Field(
+        default=0.0, ge=0.0, le=1.0, description="ROI left edge as a fraction of width (0..1)"
+    )
+    roi_y: float = Field(
+        default=0.0, ge=0.0, le=1.0, description="ROI top edge as a fraction of height (0..1)"
+    )
+    roi_width: float = Field(
+        default=1.0, gt=0.0, le=1.0, description="ROI width as a fraction of width (0..1)"
+    )
+    roi_height: float = Field(
+        default=1.0, gt=0.0, le=1.0, description="ROI height as a fraction of height (0..1)"
+    )
+    linear_blend: bool = Field(
+        default=False,
+        description="Blend border pixels in linear light instead of sRGB (perceptually correct)",
+    )
+    dither: bool = Field(
+        default=False,
+        description="Spatio-temporally dither the final 8-bit output to reduce gradient banding",
+    )


 class CalibrationTestModeRequest(BaseModel):
@@ -416,11 +457,19 @@ class DeviceResponse(BaseModel):
        ),
    )
    hue_entertainment_group_id: str = Field(default="", description="Hue entertainment group ID")
+    hue_gradient_mode: bool = Field(
+        default=True,
+        description="Map the strip across gradient-lightstrip channels vs one record per light",
+    )
    yeelight_min_interval_ms: int = Field(
        default=500, description="Yeelight client-side rate limit in ms"
    )
    wiz_min_interval_ms: int = Field(default=50, description="WiZ client-side rate limit in ms")
    lifx_min_interval_ms: int = Field(default=50, description="LIFX client-side rate limit in ms")
+    lifx_per_zone: bool = Field(
+        default=False,
+        description="Stream individual zones/tiles (multizone/matrix) vs single colour",
+    )
    govee_min_interval_ms: int = Field(default=50, description="Govee client-side rate limit in ms")
    opc_channel: int = Field(default=0, description="OPC channel (0 = broadcast to all)")
    nanoleaf_paired: bool = Field(
@@ -434,6 +483,9 @@ class DeviceResponse(BaseModel):
    nanoleaf_min_interval_ms: int = Field(
        default=100, description="Nanoleaf client-side rate limit in ms"
    )
+    nanoleaf_per_panel: bool = Field(
+        default=False, description="Stream each panel individually via extControl UDP"
+    )
    spi_speed_hz: int = Field(default=800000, description="SPI clock speed in Hz")
    spi_led_type: str = Field(default="WS2812B", description="LED chipset type")
    chroma_device_type: str = Field(default="chromalink", description="Chroma peripheral type")
@@ -16,6 +16,15 @@ class MQTTSourceCreate(BaseModel):
    password: str = Field(default="", description="Broker password (optional)")
    client_id: str = Field(default="ledgrab", description="MQTT client ID")
    base_topic: str = Field(default="ledgrab", description="Base topic prefix")
+    publish_ha_discovery: bool = Field(
+        default=False, description="Publish Home Assistant MQTT auto-discovery configs"
+    )
+    discovery_prefix: str = Field(
+        default="homeassistant",
+        max_length=64,
+        pattern=r"^[A-Za-z0-9_\-/]+$",
+        description="HA MQTT discovery prefix (default 'homeassistant')",
+    )
    description: str | None = Field(None, description="Optional description", max_length=500)
    tags: List[str] = Field(default_factory=list, description="User-defined tags")
    icon: str | None = Field(
@@ -40,6 +49,15 @@ class MQTTSourceUpdate(BaseModel):
    password: str | None = Field(None, description="Broker password")
    client_id: str | None = Field(None, description="MQTT client ID")
    base_topic: str | None = Field(None, description="Base topic prefix")
+    publish_ha_discovery: bool | None = Field(
+        None, description="Publish Home Assistant MQTT auto-discovery configs"
+    )
+    discovery_prefix: str | None = Field(
+        None,
+        max_length=64,
+        pattern=r"^[A-Za-z0-9_\-/]+$",
+        description="HA MQTT discovery prefix",
+    )
    description: str | None = Field(None, description="Optional description", max_length=500)
    tags: List[str] | None = None
    icon: str | None = Field(
@@ -65,6 +83,8 @@ class MQTTSourceResponse(BaseModel):
    password_set: bool = Field(default=False, description="Whether a password is configured")
    client_id: str = Field(description="MQTT client ID")
    base_topic: str = Field(description="Base topic prefix")
+    publish_ha_discovery: bool = Field(default=False, description="HA MQTT discovery enabled")
+    discovery_prefix: str = Field(default="homeassistant", description="HA MQTT discovery prefix")
    connected: bool = Field(default=False, description="Whether the broker connection is active")
    description: str | None = Field(None, description="Description")
    tags: List[str] = Field(default_factory=list, description="User-defined tags")
@@ -0,0 +1,94 @@
+"""Scene playlist API schemas."""
+
+from datetime import datetime
+from typing import List
+
+from pydantic import BaseModel, Field
+
+from ledgrab.storage.scene_playlist import (
+    MAX_DURATION_SECONDS,
+    MIN_DURATION_SECONDS,
+)
+
+
+class PlaylistItemSchema(BaseModel):
+    scene_preset_id: str = Field(min_length=1, description="Referenced scene preset id")
+    duration_seconds: float = Field(
+        default=30.0,
+        ge=MIN_DURATION_SECONDS,
+        le=MAX_DURATION_SECONDS,
+        description="How long to hold this scene before advancing",
+    )
+
+
+class ScenePlaylistCreate(BaseModel):
+    """Create a scene playlist."""
+
+    name: str = Field(description="Playlist name", min_length=1, max_length=100)
+    description: str = Field(default="", max_length=500)
+    items: List[PlaylistItemSchema] = Field(
+        default_factory=list, description="Ordered playlist items"
+    )
+    loop: bool = Field(default=True, description="Restart from the first item after the last")
+    shuffle: bool = Field(default=False, description="Randomise item order each cycle")
+    tags: List[str] = Field(default_factory=list, description="User-defined tags")
+    icon: str | None = Field(
+        None,
+        max_length=64,
+        description="Icon id from the curated icon library. Pass empty string to clear.",
+    )
+    icon_color: str | None = Field(
+        None,
+        max_length=32,
+        description="Optional CSS color override for the icon.",
+    )
+
+
+class ScenePlaylistUpdate(BaseModel):
+    """Update scene playlist metadata, items, and playback flags."""
+
+    name: str | None = Field(None, min_length=1, max_length=100)
+    description: str | None = Field(None, max_length=500)
+    items: List[PlaylistItemSchema] | None = Field(None, description="Replace the item list")
+    loop: bool | None = None
+    shuffle: bool | None = None
+    order: int | None = None
+    tags: List[str] | None = None
+    icon: str | None = Field(None, max_length=64)
+    icon_color: str | None = Field(None, max_length=32)
+
+
+class PlaylistRuntimeStateSchema(BaseModel):
+    is_running: bool = False
+    playlist_id: str | None = None
+    playlist_name: str | None = None
+    current_index: int = 0
+    item_count: int = 0
+    current_preset_id: str | None = None
+    started_at: datetime | None = None
+    step_started_at: datetime | None = None
+    step_duration: float = 0.0
+
+
+class ScenePlaylistResponse(BaseModel):
+    """Scene playlist with items and runtime running flag."""
+
+    id: str
+    name: str
+    description: str
+    items: List[PlaylistItemSchema]
+    loop: bool
+    shuffle: bool
+    order: int
+    tags: List[str] = Field(default_factory=list)
+    icon: str | None = Field(None, max_length=64)
+    icon_color: str | None = Field(None, max_length=32)
+    is_running: bool = Field(default=False, description="True if this playlist is cycling now")
+    created_at: datetime
+    updated_at: datetime
+
+
+class ScenePlaylistListResponse(BaseModel):
+    playlists: List[ScenePlaylistResponse]
+    count: int
+    state: PlaylistRuntimeStateSchema
@@ -0,0 +1,63 @@
+"""Pydantic schemas for the setup scaffold endpoint."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from pydantic import BaseModel, Field
+
+
+class ScaffoldRequest(BaseModel):
+    """Request body for ``POST /api/v1/setup/scaffold``.
+
+    Creates a full capture-to-output chain:
+      capture template → picture source → picture color-strip source → LED output target
+
+    ``device_id`` must reference an existing, validated device (created via
+    ``POST /api/v1/devices``).  The wizard flow is: discover/create the device
+    via the canonical device endpoint first, then call scaffold with the
+    resulting ``device_id``.
+    """
+
+    # ── Existing device (required) ────────────────────────────────────────────
+    device_id: str = Field(
+        description="ID of an existing device to wire into the chain.",
+    )
+
+    # ── Capture / picture source ──────────────────────────────────────────────
+    display_index: int = Field(
+        0,
+        ge=0,
+        le=63,
+        description="Index of the monitor to capture (0 = primary; max 63).",
+    )
+
+    # ── Optional calibration override ─────────────────────────────────────────
+    calibration: dict[str, Any] | None = Field(
+        None,
+        description=(
+            "Optional CalibrationConfig dict to use for the color-strip source.  "
+            "When omitted, ``create_default_calibration(led_count)`` is used."
+        ),
+    )
+
+
+class ScaffoldResponse(BaseModel):
+    """IDs of every entity created (or reused) by the scaffold.
+
+    ``capture_template_reused`` is ``True`` when the scaffold matched an
+    existing template by engine type instead of creating a new one.
+    The device is always pre-existing (created via the canonical device endpoint
+    before calling scaffold).
+    """
+
+    device_id: str = Field(description="Device id (pre-existing).")
+    capture_template_id: str = Field(description="Capture template id.")
+    picture_source_id: str = Field(description="Raw picture source id.")
+    color_strip_source_id: str = Field(description="Picture color-strip source id.")
+    output_target_id: str = Field(description="LED output target id.")
+
+    capture_template_reused: bool = Field(
+        False,
+        description="True when an existing matching capture template was reused.",
+    )
@@ -68,6 +68,12 @@ class AuthConfig(BaseSettings):
    """Authentication configuration."""

    api_keys: dict[str, str] = {}  # label: key mapping (empty = auth disabled)
+    # When True, the OpenAPI docs routes (/docs, /redoc, /openapi.json) load
+    # WITHOUT a Bearer token from any client (loopback and LAN). This exposes
+    # the API *surface* (route paths + parameter schemas), not data — actually
+    # invoking an endpoint from Swagger still requires the token via its
+    # "Authorize" button. All other endpoints stay protected. Default off.
+    expose_docs: bool = False


 class AssetsConfig(BaseSettings):
@@ -0,0 +1,25 @@
+"""Activity / audit log core — recorder, retention engine, and actor context.
+
+Public surface
+--------------
+``ActivityRecorder``     — thread-safe facade; persists entries and fires live events.
+``ActivityLogRetentionEngine`` — background pruning engine (mirrors AutoBackupEngine).
+``current_actor``        — ``ContextVar[str]`` set by the auth layer per request.
+
+Quick start
+-----------
+Wired in ``main.py`` lifespan; injected via ``api/dependencies.py`` getters.
+Phase 3 adds the instrumentation call sites.
+"""
+
+from ledgrab.core.activity_log.context import current_actor
+from ledgrab.core.activity_log.recorder import ActivityRecorder
+from ledgrab.core.activity_log.retention import ActivityLogRetentionEngine
+from ledgrab.core.activity_log.sanitize import sanitize_display
+
+__all__ = [
+    "ActivityRecorder",
+    "ActivityLogRetentionEngine",
+    "current_actor",
+    "sanitize_display",
+]
@@ -0,0 +1,34 @@
+"""Actor context variable for the activity log.
+
+``current_actor`` is set by ``api/auth.py:verify_api_key`` so that
+``ActivityRecorder.record(...)`` can resolve the actor without requiring every
+call site to pass it explicitly.
+
+Default value is ``"system"`` — used by background engines and any code path
+that runs outside a request context (e.g. lifespan startup/shutdown, zeroconf
+discovery thread).
+
+Per-request isolation is provided by ASGI/anyio ContextVar copy semantics:
+Starlette dispatches each request in its own task whose context is a copy of
+the parent, so a ``current_actor.set(...)`` in one request is never visible to
+another request, and each request starts from the ``"system"`` default.
+
+The auth layer only *sets* (never resets) the actor: ``verify_api_key`` calls
+``current_actor.set(...)`` on the authenticated path and on the loopback-
+anonymous path. It is an ``async`` dependency on purpose — an async dependency
+runs in the same task/context as the route handler, so the ``set`` is visible
+to ``record(...)`` (a sync dependency would set it in a throwaway threadpool
+context that the handler never sees). Routes without the ``verify_api_key``
+dependency (e.g. the unauthenticated ``POST /api/v1/webhooks/{token}``) never
+set it and therefore record as ``"system"``.
+
+There is intentionally no explicit per-request reset — do not rely on one. If
+you run a recorder call in a worker thread that inherited a parent request's
+context, pass an explicit ``actor=`` to ``record(...)`` rather than trusting
+the ContextVar default.
+"""
+
+from contextvars import ContextVar
+
+#: The actor label for the current request — API-key label or ``"system"``.
+current_actor: ContextVar[str] = ContextVar("current_actor", default="system")
@@ -0,0 +1,273 @@
+"""Thread-safe ActivityRecorder facade.
+
+Responsibilities
+----------------
+1. Build an ``ActivityLogEntry`` from the caller-supplied fields.
+2. Resolve the ``actor`` from the ``current_actor`` ContextVar when not given.
+3. Persist the entry via ``ActivityLogRepository.record()`` on the event-loop
+   thread — inline if already on that thread, via
+   ``loop.call_soon_threadsafe`` if called from another thread (e.g. the
+   zeroconf discovery thread that fires ``device_discovered/lost`` events).
+4. Push a live ``activity_logged`` event via
+   ``ProcessorManager.fire_event({"type": "activity_logged", "entry": {...}})``.
+5. Never raise into the caller — audit recording is best-effort.  Failures are
+   logged at ``WARNING`` level so operators can diagnose without breaking the
+   audited action.
+
+Thread-marshal pattern mirrors ``utils/log_broadcaster.py`` (``ensure_loop`` /
+``call_soon_threadsafe``).
+
+Module accessor
+---------------
+A module-level singleton ``_recorder`` is populated by
+``set_module_recorder()`` during ``main.py`` lifespan startup and exposed via
+``get_module_recorder()``.  Background engines and other non-DI sites that need
+to call ``record()`` without FastAPI DI can use this accessor.  Phase 3
+instrumentation uses it at the ``fire_entity_event`` choke-point.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import uuid
+from datetime import datetime, timezone
+from typing import TYPE_CHECKING
+
+from ledgrab.core.activity_log.context import current_actor
+from ledgrab.storage.activity_log import ActivityLogEntry, ActivitySeverity
+from ledgrab.utils import get_logger
+
+if TYPE_CHECKING:
+    from ledgrab.core.processing.processor_manager import ProcessorManager
+    from ledgrab.storage.activity_log_repository import ActivityLogRepository
+
+logger = get_logger(__name__)
+
+
+def _new_id() -> str:
+    """Generate an activity-log entry id: ``al_<32-hex-chars>``.
+
+    Uses the full 128-bit uuid4 hex. The ``id`` column is ``UNIQUE`` and a
+    collision is silently dropped (best-effort recorder), so the entropy must
+    be high enough that a collision is astronomically unlikely even against the
+    full retention window (default 20k live rows).
+    """
+    return "al_" + uuid.uuid4().hex
+
+
+def entry_to_dict(entry: ActivityLogEntry) -> dict:
+    """Serialise an ``ActivityLogEntry`` to the canonical API/event payload dict.
+
+    Reused by Phase 4 (API response serialisation) and Phase 5 (frontend).
+    The shape is identical to the flat row codec minus the DB-only ``seq``
+    field, but with ``ts`` kept as an ISO-8601 string and ``metadata`` as a
+    real ``dict`` (not a JSON string).
+    """
+    return {
+        "id": entry.id,
+        "ts": entry.ts.isoformat(),
+        "category": entry.category,
+        "action": entry.action,
+        "severity": entry.severity,
+        "actor": entry.actor,
+        "entity_type": entry.entity_type,
+        "entity_id": entry.entity_id,
+        "entity_name": entry.entity_name,
+        "message": entry.message,
+        "metadata": entry.metadata,
+    }
+
+
+class ActivityRecorder:
+    """Thread-safe facade for persisting audit log entries.
+
+    Parameters
+    ----------
+    repo:
+        ``ActivityLogRepository`` used to persist entries.
+    processor_manager:
+        ``ProcessorManager`` whose ``fire_event`` dispatches the live
+        ``activity_logged`` event to WebSocket subscribers.
+    loop:
+        Optional: the running asyncio event loop.  If ``None``, it is
+        captured lazily on the first call that originates from an async
+        context (mirroring ``LogBroadcaster.ensure_loop``).  Pass it
+        explicitly in tests to avoid depending on a real running loop.
+    """
+
+    def __init__(
+        self,
+        repo: "ActivityLogRepository",
+        processor_manager: "ProcessorManager",
+        *,
+        loop: asyncio.AbstractEventLoop | None = None,
+    ) -> None:
+        self._repo = repo
+        self._pm = processor_manager
+        self._loop: asyncio.AbstractEventLoop | None = loop
+        self._enabled: bool = True
+
+    # ── Loop capture (mirrors LogBroadcaster.ensure_loop) ──────────────────
+
+    def ensure_loop(self) -> None:
+        """Capture the running event loop if not already stored.
+
+        Call from an async context (e.g. lifespan startup) so that
+        ``call_soon_threadsafe`` works when ``record()`` is later called from
+        non-async threads.
+        """
+        if self._loop is None:
+            try:
+                self._loop = asyncio.get_running_loop()
+            except RuntimeError:
+                pass
+
+    # ── Public API ──────────────────────────────────────────────────────────
+
+    @property
+    def enabled(self) -> bool:
+        """Whether recording is currently active."""
+        return self._enabled
+
+    @enabled.setter
+    def enabled(self, value: bool) -> None:
+        self._enabled = value
+
+    def record(
+        self,
+        category: str,
+        action: str,
+        *,
+        severity: str = ActivitySeverity.INFO,
+        actor: str | None = None,
+        entity_type: str | None = None,
+        entity_id: str | None = None,
+        entity_name: str | None = None,
+        message: str,
+        metadata: dict | None = None,
+        _bypass_enabled: bool = False,
+    ) -> None:
+        """Append one audit entry — best-effort, never raises.
+
+        Parameters
+        ----------
+        category:
+            Broad bucket — one of :class:`~ledgrab.storage.activity_log.ActivityCategory`.
+        action:
+            Verb-object label, e.g. ``"entity.created"`` or ``"server.shutting_down"``.
+        severity:
+            One of :class:`~ledgrab.storage.activity_log.ActivitySeverity`.  Defaults
+            to ``"info"``.
+        actor:
+            Who triggered the action.  When ``None`` (the common case), the
+            value is resolved from :data:`~ledgrab.core.activity_log.context.current_actor`
+            with a default of ``"system"``.
+        entity_type / entity_id / entity_name:
+            Optional entity context for entity-domain events.
+        message:
+            Human-readable description suitable for display.
+        metadata:
+            Small JSON-serialisable dict with extra context.  Defaults to ``{}``.
+        _bypass_enabled:
+            Internal flag used by the retention engine to record the
+            "audit log disabled" event even when ``enabled`` is ``False``.
+        """
+        if not self._enabled and not _bypass_enabled:
+            return
+
+        # Resolve actor from ContextVar when not explicitly supplied.
+        resolved_actor = actor if actor is not None else current_actor.get()
+
+        entry = ActivityLogEntry(
+            id=_new_id(),
+            ts=datetime.now(timezone.utc),
+            category=category,
+            action=action,
+            severity=severity,
+            actor=resolved_actor,
+            entity_type=entity_type,
+            entity_id=entity_id,
+            entity_name=entity_name,
+            message=message,
+            metadata=metadata or {},
+        )
+
+        # Determine whether we are on the event-loop thread or not.
+        loop = self._loop
+        if loop is None:
+            # Lazy capture — may fail if called before the loop is running.
+            try:
+                loop = asyncio.get_running_loop()
+                self._loop = loop
+            except RuntimeError:
+                pass
+
+        if loop is not None and loop.is_running():
+            try:
+                current = asyncio.get_event_loop()
+            except RuntimeError:
+                current = None
+
+            # If the current thread IS the event-loop thread, write inline.
+            if current is loop:
+                self._write_and_emit(entry)
+            else:
+                # Called from a non-loop thread (e.g. zeroconf discovery) —
+                # marshal onto the event-loop thread.
+                try:
+                    loop.call_soon_threadsafe(self._write_and_emit, entry)
+                except RuntimeError:
+                    # Loop has been closed (rare; happens during tests)
+                    logger.warning(
+                        "ActivityRecorder: event loop closed, dropping entry %s", entry.id
+                    )
+        else:
+            # No running loop — fall back to a direct synchronous write.
+            # This path hits in synchronous unit tests that do not start a loop.
+            self._write_and_emit(entry)
+
+    def _write_and_emit(self, entry: ActivityLogEntry) -> None:
+        """Persist *entry* and fire the live event — called on the loop thread."""
+        try:
+            self._repo.record(entry)
+        except Exception as exc:
+            logger.warning("ActivityRecorder: failed to persist entry %s: %s", entry.id, exc)
+            return  # don't emit an event for an entry that failed to persist
+
+        try:
+            self._pm.fire_event(
+                {
+                    "type": "activity_logged",
+                    "entry": entry_to_dict(entry),
+                }
+            )
+        except Exception as exc:
+            logger.warning("ActivityRecorder: failed to fire live event for %s: %s", entry.id, exc)
+
+
+# ── Module-level singleton accessor ────────────────────────────────────────
+#
+# Background engines and non-DI call sites (Phase 3's fire_entity_event hook,
+# device discovery thread) need ``record()`` without going through FastAPI DI.
+# ``set_module_recorder`` is called from ``main.py`` lifespan immediately after
+# the recorder is wired into ``init_dependencies``.
+
+_recorder: ActivityRecorder | None = None
+
+
+def set_module_recorder(recorder: ActivityRecorder) -> None:
+    """Store the application-level recorder in the module singleton.
+
+    Called once from ``main.py`` lifespan startup.
+    """
+    global _recorder
+    _recorder = recorder
+
+
+def get_module_recorder() -> ActivityRecorder | None:
+    """Return the module-level recorder, or ``None`` if not yet initialised.
+
+    Callers must guard against ``None`` — this returns ``None`` during module
+    import and early startup before ``main.py`` lifespan has run.
+    """
+    return _recorder
@@ -0,0 +1,216 @@
+"""Activity log retention engine.
+
+Mirrors ``core/backup/auto_backup.py``:
+- Settings persisted via ``db.get_setting("activity_log")`` /
+  ``db.set_setting("activity_log", {...})``.
+- ``start()`` / ``stop()`` lifecycle following the engine convention used
+  throughout the codebase.
+- Hourly background loop calling ``repo.prune(before_ts=..., max_entries=...)``.
+- ``get_settings()`` / ``async update_settings(...)`` for the Settings API
+  (Phase 4).
+
+Changing ``enabled`` to ``False`` records an ``"audit_log.disabled"`` event via
+the recorder BEFORE the flag takes effect — so the last action in the log is a
+record of the intentional disable.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from datetime import datetime, timedelta, timezone
+from typing import TYPE_CHECKING
+
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+from ledgrab.utils import get_logger
+
+if TYPE_CHECKING:
+    from ledgrab.core.activity_log.recorder import ActivityRecorder
+    from ledgrab.storage.activity_log_repository import ActivityLogRepository
+    from ledgrab.storage.database import Database
+
+logger = get_logger(__name__)
+
+DEFAULT_SETTINGS: dict = {
+    "enabled": True,
+    "max_days": 90,
+    "max_entries": 20000,
+}
+
+# Prune loop interval — run roughly once an hour.
+_PRUNE_INTERVAL_SECS = 3600
+
+
+class ActivityLogRetentionEngine:
+    """Background engine that prunes old activity log entries.
+
+    Parameters
+    ----------
+    repo:
+        The ``ActivityLogRepository`` used to prune entries.
+    db:
+        The shared ``Database`` singleton for settings persistence.
+    recorder:
+        The ``ActivityRecorder`` used to log the "audit log disabled" event
+        before disabling takes effect.
+    """
+
+    def __init__(
+        self,
+        repo: "ActivityLogRepository",
+        db: "Database",
+        recorder: "ActivityRecorder",
+    ) -> None:
+        self._repo = repo
+        self._db = db
+        self._recorder = recorder
+        self._task: asyncio.Task | None = None
+        self._settings = self._load_settings()
+        # Rehydrate the recorder's enabled flag from persisted settings so a
+        # previously-disabled log stays disabled across restarts.
+        self._recorder.enabled = self._settings["enabled"]
+
+    # ── Settings persistence ───────────────────────────────────────────────
+
+    def _load_settings(self) -> dict:
+        data = self._db.get_setting("activity_log")
+        if data:
+            return {**DEFAULT_SETTINGS, **data}
+        return dict(DEFAULT_SETTINGS)
+
+    def _save_settings(self) -> None:
+        self._db.set_setting(
+            "activity_log",
+            {
+                "enabled": self._settings["enabled"],
+                "max_days": self._settings["max_days"],
+                "max_entries": self._settings["max_entries"],
+            },
+        )
+
+    # ── Lifecycle ──────────────────────────────────────────────────────────
+
+    async def start(self) -> None:
+        """Start the retention loop if enabled."""
+        if self._settings["enabled"]:
+            self._start_loop()
+            logger.info(
+                "Activity log retention engine started " "(max_days=%d, max_entries=%d)",
+                self._settings["max_days"],
+                self._settings["max_entries"],
+            )
+        else:
+            logger.info("Activity log retention engine initialized (disabled)")
+
+    async def stop(self) -> None:
+        """Cancel the retention loop."""
+        self._cancel_loop()
+        logger.info("Activity log retention engine stopped")
+
+    def _start_loop(self) -> None:
+        self._cancel_loop()
+        self._task = asyncio.create_task(self._retention_loop())
+
+    def _cancel_loop(self) -> None:
+        if self._task is not None:
+            self._task.cancel()
+            self._task = None
+
+    # ── Prune loop ─────────────────────────────────────────────────────────
+
+    async def _retention_loop(self) -> None:
+        try:
+            while True:
+                await asyncio.sleep(_PRUNE_INTERVAL_SECS)
+                try:
+                    self._prune()
+                except Exception as exc:
+                    logger.error("Activity log retention prune failed: %s", exc, exc_info=True)
+        except asyncio.CancelledError:
+            logger.debug("Activity log retention loop cancelled")
+
+    def _prune(self) -> None:
+        """Execute one prune pass based on current settings."""
+        settings = self._settings
+        if not settings["enabled"]:
+            return
+
+        max_days: int = settings["max_days"]
+        max_entries: int = settings["max_entries"]
+
+        before_ts: datetime | None = None
+        if max_days and max_days > 0:
+            before_ts = datetime.now(timezone.utc) - timedelta(days=max_days)
+
+        max_entries_val: int | None = max_entries if max_entries and max_entries > 0 else None
+
+        deleted = self._repo.prune(before_ts=before_ts, max_entries=max_entries_val)
+        if deleted:
+            logger.info(
+                "Activity log pruned %d rows (max_days=%d, max_entries=%d)",
+                deleted,
+                max_days,
+                max_entries,
+            )
+
+    # ── Public API ─────────────────────────────────────────────────────────
+
+    def get_settings(self) -> dict:
+        """Return the current retention settings dict."""
+        return {
+            "enabled": self._settings["enabled"],
+            "max_days": self._settings["max_days"],
+            "max_entries": self._settings["max_entries"],
+        }
+
+    async def update_settings(
+        self,
+        *,
+        enabled: bool,
+        max_days: int,
+        max_entries: int,
+    ) -> dict:
+        """Persist new settings and apply them immediately.
+
+        If ``enabled`` is changing to ``False``, the disable event is recorded
+        BEFORE the flag takes effect so there is a final log entry.
+
+        Returns the new settings dict (same as ``get_settings()``).
+        """
+        was_enabled = self._settings["enabled"]
+
+        # Record the disable event before the recorder stops accepting entries.
+        if was_enabled and not enabled:
+            self._recorder.record(
+                category=ActivityCategory.SYSTEM,
+                action="audit_log.disabled",
+                severity=ActivitySeverity.WARNING,
+                actor="system",
+                message="Activity log recording disabled via settings",
+                _bypass_enabled=True,
+            )
+
+        self._settings["enabled"] = enabled
+        self._settings["max_days"] = max_days
+        self._settings["max_entries"] = max_entries
+        self._save_settings()
+
+        # Propagate enabled flag to the recorder.
+        self._recorder.enabled = enabled
+
+        if enabled:
+            self._start_loop()
+            logger.info(
+                "Activity log retention enabled (max_days=%d, max_entries=%d)",
+                max_days,
+                max_entries,
+            )
+            # Run an immediate prune pass when re-enabling.
+            try:
+                self._prune()
+            except Exception as exc:
+                logger.error("Activity log immediate prune failed: %s", exc)
+        else:
+            self._cancel_loop()
+            logger.info("Activity log retention disabled")
+
+        return self.get_settings()
@@ -0,0 +1,90 @@
+"""Log-injection sanitizer for audit-log message and display strings.
+
+Provides :func:`sanitize_display` — a dependency-free helper that strips
+characters that should not appear in a recorded ``message`` or display
+string before it is persisted to SQLite, broadcast over WebSocket, or
+exported to CSV.
+
+Design constraints
+------------------
+- **Dependency-free**: uses only the Python standard library so it can be
+  imported from any module without adding transitive weight.
+- **Conservative**: keeps printable ASCII/Unicode and normal spaces; drops
+  everything else including control chars (NUL, BEL, BS, VT, FF, ESC,
+  DEL), ANSI/CSI escape sequences (``\\x1b[...``), and carriage returns /
+  newlines / tabs which are the classic log-injection primitives.
+- **Length-capped**: truncates to *maxlen* characters and appends ``"…"``
+  so callers can rely on a bounded string without adding their own guards.
+"""
+
+from __future__ import annotations
+
+import re
+
+# Matches ANSI/VT100 escape sequences: ESC [ ... m  (CSI) and shorter forms.
+# We strip these before the printable-char filter so the bracket/letters that
+# follow the ESC don't survive stripping the ESC alone.
+_ANSI_RE = re.compile(r"\x1b(?:[@-Z\\-_]|\[[0-?]*[ -/]*[@-~])")
+
+# Characters we explicitly want to remove even if str.isprintable() would
+# let them through in some edge-case: NUL is the canonical SQL/log null-byte
+# injection; the others are kept out by the printable check but listed here
+# for documentation clarity.
+_EXPLICIT_DROP = frozenset("\x00\r\n\t")
+
+
+def sanitize_display(value: str | None, *, maxlen: int = 120) -> str:
+    """Return a sanitized, length-capped version of *value* safe for log messages.
+
+    Parameters
+    ----------
+    value:
+        The raw, potentially attacker-controlled string.  ``None`` or empty
+        returns ``""``.
+    maxlen:
+        Maximum length of the returned string (default: 120).  If the input
+        exceeds this length after sanitization, the string is truncated and
+        ``"…"`` is appended (the ellipsis counts toward *maxlen*).
+
+    Returns
+    -------
+    str
+        A string that:
+        - contains no NUL bytes (``\\x00``),
+        - contains no ANSI/CSI escape sequences,
+        - contains no carriage returns, newlines, or tab characters,
+        - contains only characters for which ``str.isprintable()`` is ``True``
+          plus the regular ASCII space (``\\x20``),
+        - is at most *maxlen* characters long.
+    """
+    if not value:
+        return ""
+
+    # 1. Strip ANSI escape sequences first so their bracket/letter tails don't
+    #    survive as stray printable characters.
+    cleaned = _ANSI_RE.sub("", value)
+
+    # 2. Drop each character that is neither printable nor a plain space.
+    #    str.isprintable() returns False for all control chars (including NUL,
+    #    BEL, BS, TAB, LF, VT, FF, CR, ESC, DEL) and True for normal letters,
+    #    digits, punctuation, and the space character.
+    cleaned = "".join(ch for ch in cleaned if ch.isprintable() or ch == " ")
+
+    # 3. Final belt-and-suspenders pass for the explicit drop set (catches NUL
+    #    that may survive if isprintable ever changes in a future Python version).
+    cleaned = "".join(ch for ch in cleaned if ch not in _EXPLICIT_DROP)
+
+    # 4. Cap length. Guard the degenerate maxlen cases: ``cleaned[: maxlen - 1]``
+    #    with maxlen <= 0 would slice from the END (keeping all-but-last char or
+    #    a negative-index tail), violating the bounded-length contract.
+    if maxlen <= 0:
+        return ""
+    if len(cleaned) > maxlen:
+        if maxlen == 1:
+            # No room for content + ellipsis; emit the ellipsis alone.
+            cleaned = "…"
+        else:
+            # Reserve one character for the ellipsis so total length == maxlen.
+            cleaned = cleaned[: maxlen - 1] + "…"
+
+    return cleaned
@@ -13,8 +13,10 @@ from ledgrab.storage.automation import (
    DisplayStateRule,
    HomeAssistantRule,
    HTTPPollRule,
+    ManualTriggerRule,
    MQTTRule,
    Rule,
+    SolarRule,
    StartupRule,
    SystemIdleRule,
    TimeOfDayRule,
@@ -23,6 +25,7 @@ from ledgrab.storage.automation import (
 from ledgrab.storage.automation_store import AutomationStore
 from ledgrab.storage.scene_preset import ScenePreset
 from ledgrab.utils import get_logger
+from ledgrab.utils.solar import compute_solar_times, utc_offset_hours_for

 logger = get_logger(__name__)

@@ -141,6 +144,11 @@ class AutomationEngine:
        self._last_deactivated: Dict[str, datetime] = {}
        # webhook_token → bool (volatile state set by webhook calls)
        self._webhook_states: Dict[str, bool] = {}
+        # True only while a single automation is being manually fired
+        # (fire_manual_trigger). The background tick never sets it, so a
+        # ManualTriggerRule reads False during normal evaluation and a
+        # manual-trigger automation never activates on its own.
+        self._manual_fire_active: bool = False
        # HA source IDs currently acquired by the engine
        self._ha_acquired: Set[str] = set()
        # MQTT source IDs currently acquired by the engine
@@ -369,6 +377,32 @@ class AutomationEngine:
            display_state,
        )

+    @staticmethod
+    def _detection_needs(rules) -> tuple[bool, bool, bool, bool, bool]:
+        """Which platform-detection probes a set of rules requires.
+
+        Returns ``(needs_running, needs_topmost, needs_fullscreen, needs_idle,
+        needs_display_state)``. Shared by the background evaluation tick and the
+        one-shot manual-trigger path so both request the same detection set.
+        """
+        match_types_used: set = set()
+        needs_idle = False
+        needs_display_state = False
+        for r in rules:
+            if isinstance(r, ApplicationRule):
+                match_types_used.add(r.match_type)
+            elif isinstance(r, SystemIdleRule):
+                needs_idle = True
+            elif isinstance(r, DisplayStateRule):
+                needs_display_state = True
+        return (
+            "running" in match_types_used,
+            bool(match_types_used & {"topmost", "topmost_fullscreen"}),
+            "fullscreen" in match_types_used,
+            needs_idle,
+            needs_display_state,
+        )
+
    async def _evaluate_all_locked(self) -> None:
        automations = self._store.get_all_automations()
        if not automations:
@@ -377,23 +411,15 @@ class AutomationEngine:
                await self._deactivate_automation(aid)
            return

-        # Determine which detection methods are actually needed
-        match_types_used: set = set()
-        needs_idle = False
-        needs_display_state = False
-        for a in automations:
-            if a.enabled:
-                for r in a.rules:
-                    if isinstance(r, ApplicationRule):
-                        match_types_used.add(r.match_type)
-                    elif isinstance(r, SystemIdleRule):
-                        needs_idle = True
-                    elif isinstance(r, DisplayStateRule):
-                        needs_display_state = True
-
-        needs_running = "running" in match_types_used
-        needs_topmost = bool(match_types_used & {"topmost", "topmost_fullscreen"})
-        needs_fullscreen = "fullscreen" in match_types_used
+        # Determine which detection methods are actually needed (across the
+        # rules of every *enabled* automation — disabled ones are skipped below).
+        (
+            needs_running,
+            needs_topmost,
+            needs_fullscreen,
+            needs_idle,
+            needs_display_state,
+        ) = self._detection_needs([r for a in automations if a.enabled for r in a.rules])

        # Single executor call for all platform detection
        (
@@ -526,6 +552,9 @@ class AutomationEngine:
    def _handle_time_of_day(self, rule: TimeOfDayRule, ctx: _RuleEvalContext) -> bool:
        return self._evaluate_time_of_day(rule)

+    def _handle_solar(self, rule: SolarRule, ctx: _RuleEvalContext) -> bool:
+        return self._evaluate_solar(rule)
+
    def _handle_system_idle(self, rule: SystemIdleRule, ctx: _RuleEvalContext) -> bool:
        return self._evaluate_idle(rule, ctx.idle_seconds)

@@ -538,12 +567,40 @@ class AutomationEngine:
    def _handle_webhook(self, rule: WebhookRule, ctx: _RuleEvalContext) -> bool:
        return self._webhook_states.get(rule.token, False)

+    def _handle_manual(self, rule: ManualTriggerRule, ctx: _RuleEvalContext) -> bool:
+        # True only while fire_manual_trigger is evaluating this one automation
+        # under the eval lock; always False during the background tick.
+        return self._manual_fire_active
+
    def _handle_home_assistant(self, rule: HomeAssistantRule, ctx: _RuleEvalContext) -> bool:
        return self._evaluate_home_assistant(rule)

    def _handle_http_poll(self, rule: HTTPPollRule, ctx: _RuleEvalContext) -> bool:
        return self._evaluate_http_poll(rule)

+    @staticmethod
+    def _weekday_window_active(
+        current: int, start: int, end: int, weekday: int, days: list
+    ) -> bool:
+        """Is ``current`` (minutes-of-day) inside the [start, end] window?
+
+        Handles the overnight wrap (start > end): the after-midnight tail
+        belongs to the window's START day, so it's matched against the
+        previous weekday. ``days`` empty = every day of the week.
+        """
+        if start <= end:
+            if not (start <= current <= end):
+                return False
+            return not days or weekday in days
+
+        # Overnight range (e.g. 22:00 → 06:00): the window belongs to its
+        # START day, so the after-midnight tail is matched against yesterday.
+        if current >= start:  # evening portion — today's window
+            return not days or weekday in days
+        if current <= end:  # early-morning portion — yesterday's window
+            return not days or ((weekday - 1) % 7) in days
+        return False
+
    @staticmethod
    def _evaluate_time_of_day(rule: TimeOfDayRule) -> bool:
        now = _now_in_tz(rule.timezone)
@@ -552,20 +609,34 @@ class AutomationEngine:
        parts_e = rule.end_time.split(":")
        start = int(parts_s[0]) * 60 + int(parts_s[1])
        end = int(parts_e[0]) * 60 + int(parts_e[1])
-        days = rule.days_of_week
+        return AutomationEngine._weekday_window_active(
+            current, start, end, now.weekday(), rule.days_of_week
+        )

-        if start <= end:
-            if not (start <= current <= end):
-                return False
-            return not days or now.weekday() in days
+    @staticmethod
+    def _evaluate_solar(rule: SolarRule) -> bool:
+        # One ``now`` drives every read: day-of-year, the UTC offset for the
+        # solar math, the current-minute compare, and the weekday.
+        now = _now_in_tz(rule.timezone)
+        day_of_year = now.timetuple().tm_yday
+        utc_offset = utc_offset_hours_for(rule.timezone, now)
+        sunrise_h, sunset_h = compute_solar_times(
+            rule.latitude, rule.longitude, day_of_year, utc_offset
+        )

-        # Overnight range (e.g. 22:00 → 06:00): the window belongs to its
-        # START day, so the after-midnight tail is matched against yesterday.
-        if current >= start:  # evening portion — today's window
-            return not days or now.weekday() in days
-        if current <= end:  # early-morning portion — yesterday's window
-            return not days or ((now.weekday() - 1) % 7) in days
-        return False
+        def _event_minutes(event: str) -> int:
+            hour = sunset_h if event == "sunset" else sunrise_h
+            return int(round(hour * 60))
+
+        # compute_solar_times clamps sunrise < sunset, so the only way to wrap
+        # past midnight is via the offsets — which ``_weekday_window_active``
+        # handles the same way it does an overnight time-of-day window.
+        start = (_event_minutes(rule.start_event) + rule.start_offset_minutes) % 1440
+        end = (_event_minutes(rule.end_event) + rule.end_offset_minutes) % 1440
+        current = now.hour * 60 + now.minute
+        return AutomationEngine._weekday_window_active(
+            current, start, end, now.weekday(), rule.days_of_week
+        )

    @staticmethod
    def _evaluate_idle(rule: SystemIdleRule, idle_seconds: float | None) -> bool:
@@ -675,6 +746,62 @@ class AutomationEngine:
        # Default: "running"
        return any(app in running_procs for app in apps_lower)

+    def _audit_activation(self, automation: Automation) -> None:
+        """Best-effort audit record for any successful automation activation.
+
+        Shared by both the normal scene path and the no-scene branch so an
+        activation is recorded uniformly regardless of whether a scene was
+        applied (mirrors the uniform recording on the deactivation side).
+        """
+        try:
+            from ledgrab.core.activity_log.recorder import get_module_recorder
+            from ledgrab.core.activity_log.sanitize import sanitize_display
+            from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+            rec = get_module_recorder()
+            if rec is not None:
+                _safe_name = sanitize_display(automation.name) if automation.name else None
+                rec.record(
+                    category=ActivityCategory.CAPTURE,
+                    action="automation.activated",
+                    severity=ActivitySeverity.INFO,
+                    actor="system",
+                    entity_type="automation",
+                    entity_id=automation.id,
+                    entity_name=_safe_name,
+                    message=f"Automation '{_safe_name or automation.id}' activated",
+                )
+        except Exception:
+            pass
+
+    def _audit_manual_trigger(self, automation: Automation) -> None:
+        """Best-effort audit record for a manual trigger.
+
+        Unlike :meth:`_audit_activation` this does NOT force ``actor='system'``
+        — the recorder resolves ``actor`` from the ``current_actor`` ContextVar
+        (set in ``verify_api_key``), so the run is attributed to the user who
+        pressed the button.
+        """
+        try:
+            from ledgrab.core.activity_log.recorder import get_module_recorder
+            from ledgrab.core.activity_log.sanitize import sanitize_display
+            from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+            rec = get_module_recorder()
+            if rec is not None:
+                _safe_name = sanitize_display(automation.name) if automation.name else None
+                rec.record(
+                    category=ActivityCategory.CAPTURE,
+                    action="automation.triggered",
+                    severity=ActivitySeverity.INFO,
+                    entity_type="automation",
+                    entity_id=automation.id,
+                    entity_name=_safe_name,
+                    message=f"Automation '{_safe_name or automation.id}' manually triggered",
+                )
+        except Exception:
+            pass
+
    async def _activate_automation(self, automation: Automation) -> None:
        if not automation.scene_preset_id:
            # No scene configured — just mark active (rules matched but nothing to do)
@@ -682,6 +809,11 @@ class AutomationEngine:
            self._last_activated[automation.id] = datetime.now(timezone.utc)
            self._fire_event(automation.id, "activated")
            logger.info(f"Automation '{automation.name}' activated (no scene configured)")
+            # Record the activation too — a no-scene activation is still a
+            # successful activation and must appear in the audit log.
+            self._audit_activation(automation)
+            await self._fire_actions(automation, "activate")
+            await self._publish_mqtt_state(automation.id, True)
            return

        if not self._scene_preset_store or not self._target_store or not self._device_store:
@@ -726,6 +858,141 @@ class AutomationEngine:
        else:
            logger.info(f"Automation '{automation.name}' activated (scene '{preset.name}' applied)")

+        # Audit record — best-effort (shared helper, also used by no-scene path).
+        self._audit_activation(automation)
+        await self._fire_actions(automation, "activate")
+        await self._publish_mqtt_state(automation.id, True)
+
+    async def _fire_actions(self, automation: Automation, event: str) -> None:
+        """Fire any outbound actions (e.g. webhooks) for this transition.
+
+        Best-effort and never raises into the activation path: a hung or
+        failing endpoint is logged/audited but must not stall the evaluation
+        loop or abort scene activation.
+        """
+        actions = getattr(automation, "actions", None)
+        if not actions:
+            return
+        from ledgrab.storage.automation import WebhookAction
+        from ledgrab.core.automations.webhook_action import fire_webhook_action, should_fire
+
+        for action in actions:
+            if not isinstance(action, WebhookAction) or not should_fire(action, event):
+                continue
+            try:
+                ok, err = await fire_webhook_action(action, automation, event)
+            except Exception as exc:  # noqa: BLE001 — defensive; fire is already best-effort
+                logger.warning(
+                    "Action fire raised for '%s': %s", automation.name, type(exc).__name__
+                )
+                ok, err = False, type(exc).__name__
+            self._audit_webhook(automation, event, ok, err)
+
+    def _audit_webhook(self, automation: Automation, event: str, ok: bool, err: str | None) -> None:
+        """Best-effort audit entry for a webhook fire (success or failure)."""
+        try:
+            from ledgrab.core.activity_log.recorder import get_module_recorder
+            from ledgrab.core.activity_log.sanitize import sanitize_display
+            from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+            rec = get_module_recorder()
+            if rec is None:
+                return
+            safe_name = sanitize_display(automation.name) if automation.name else automation.id
+            rec.record(
+                category=ActivityCategory.CAPTURE,
+                action="automation.webhook_fired",
+                severity=ActivitySeverity.INFO if ok else ActivitySeverity.WARNING,
+                actor="system",
+                entity_type="automation",
+                entity_id=automation.id,
+                entity_name=safe_name,
+                message=(
+                    f"Webhook for '{safe_name}' {'fired' if ok else 'failed'} on {event}"
+                    + ("" if ok else f" ({sanitize_display(err) if err else 'error'})")
+                ),
+            )
+        except Exception:
+            pass
+
+    async def _apply_manual_scene(self, automation: Automation) -> tuple[str, list[str]]:
+        """Apply the automation's scene once for a manual trigger.
+
+        Mirrors the scene-application core of :meth:`_activate_automation` but
+        does NOT enter the sticky ``_active_automations`` state or capture a
+        revert snapshot — a manual trigger is a one-shot apply, so the
+        background tick has nothing to reconcile away. Returns
+        ``(status, errors)`` where ``status`` is ``"triggered"`` (applied, or no
+        scene configured), ``"partial"`` (applied with errors), or ``"error"``
+        (scene stores unavailable / preset missing).
+        """
+        if not automation.scene_preset_id:
+            return ("triggered", [])
+
+        if not self._scene_preset_store or not self._target_store or not self._device_store:
+            logger.warning(
+                f"Automation '{automation.name}' triggered but scene stores not available"
+            )
+            return ("error", ["scene stores not available"])
+
+        try:
+            preset = self._scene_preset_store.get_preset(automation.scene_preset_id)
+        except ValueError:
+            logger.warning(
+                f"Automation '{automation.name}': scene preset {automation.scene_preset_id} not found"
+            )
+            return ("error", [f"scene preset {automation.scene_preset_id} not found"])
+
+        from ledgrab.core.scenes.scene_activator import apply_scene_state
+
+        status, errors = await apply_scene_state(preset, self._target_store, self._manager)
+        if errors:
+            logger.warning(
+                f"Automation '{automation.name}' manually triggered with errors: {errors}"
+            )
+        else:
+            logger.info(
+                f"Automation '{automation.name}' manually triggered (scene '{preset.name}' applied)"
+            )
+        # apply_scene_state returns "activated"/"partial"; surface "triggered"
+        # for the happy path so the API status reads naturally.
+        return ("triggered" if status == "activated" else status, errors)
+
+    async def fire_manual_trigger(self, automation: Automation) -> tuple[str, list[str]]:
+        """Manually fire an automation: evaluate its rules with the manual
+        trigger satisfied and, if it should activate, apply its scene once.
+
+        "Checks all of the rules": the automation's full rule set is evaluated
+        under its ``rule_logic`` with the ManualTriggerRule treated as True. The
+        ``enabled`` flag is intentionally ignored — it gates only the background
+        tick; a manual trigger is an explicit user action. Returns
+        ``(status, errors)``: ``"skipped"`` when the rules are not satisfied,
+        otherwise the result of :meth:`_apply_manual_scene`.
+        """
+        async with self._eval_lock:
+            detection = await asyncio.to_thread(
+                self._detect_all_sync, *self._detection_needs(automation.rules)
+            )
+            # Force the manual term True for this one evaluation, then clear it
+            # before releasing the lock so the background tick never sees it.
+            self._manual_fire_active = True
+            try:
+                should_fire = (not automation.rules) or self._evaluate_rules(automation, *detection)
+            finally:
+                self._manual_fire_active = False
+
+        if not should_fire:
+            logger.info(
+                f"Automation '{automation.name}' manual trigger skipped (rules not satisfied)"
+            )
+            return ("skipped", [])
+
+        status, errors = await self._apply_manual_scene(automation)
+        self._last_activated[automation.id] = datetime.now(timezone.utc)
+        self._fire_event(automation.id, "triggered")
+        self._audit_manual_trigger(automation)
+        return (status, errors)
+
    async def _deactivate_automation(self, automation_id: str) -> None:
        was_active = self._active_automations.pop(automation_id, False)
        if not was_active:
@@ -751,6 +1018,47 @@ class AutomationEngine:
        # Clean up any leftover snapshot
        self._pre_activation_snapshots.pop(automation_id, None)

+        # Audit record — best-effort.
+        try:
+            from ledgrab.core.activity_log.recorder import get_module_recorder
+            from ledgrab.core.activity_log.sanitize import sanitize_display
+            from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
+
+            rec = get_module_recorder()
+            if rec is not None:
+                # Reuse the automation already fetched above (no second store
+                # read); degrades to None if it was since-deleted (== None).
+                _auto_name = automation.name if automation else None
+                _safe_deact_name = sanitize_display(_auto_name) if _auto_name else None
+                rec.record(
+                    category=ActivityCategory.CAPTURE,
+                    action="automation.deactivated",
+                    severity=ActivitySeverity.INFO,
+                    actor="system",
+                    entity_type="automation",
+                    entity_id=automation_id,
+                    entity_name=_safe_deact_name,
+                    message=f"Automation '{_safe_deact_name or automation_id}' deactivated",
+                )
+        except Exception:
+            pass
+
+        # Fire any outbound deactivate actions (best-effort). Skipped when the
+        # automation was since-deleted (no actions to read).
+        if automation is not None:
+            await self._fire_actions(automation, "deactivate")
+        await self._publish_mqtt_state(automation_id, False)
+
+    async def _publish_mqtt_state(self, automation_id: str, active: bool) -> None:
+        """Best-effort publish of the automation's active state to HA discovery."""
+        mgr = self._mqtt_manager
+        if mgr is None or not hasattr(mgr, "publish_automation_state_all"):
+            return
+        try:
+            await mgr.publish_automation_state_all(automation_id, active)
+        except Exception:  # noqa: BLE001 — never raise into the engine
+            pass
+
    async def _deactivate_revert(self, automation_id: str) -> None:
        """Revert to pre-activation snapshot."""
        snapshot = self._pre_activation_snapshots.pop(automation_id, None)
@@ -855,10 +1163,12 @@ AutomationEngine._RULE_HANDLERS = {
    StartupRule: AutomationEngine._handle_startup,
    ApplicationRule: AutomationEngine._handle_application,
    TimeOfDayRule: AutomationEngine._handle_time_of_day,
+    SolarRule: AutomationEngine._handle_solar,
    SystemIdleRule: AutomationEngine._handle_system_idle,
    DisplayStateRule: AutomationEngine._handle_display_state,
    MQTTRule: AutomationEngine._handle_mqtt,
    WebhookRule: AutomationEngine._handle_webhook,
+    ManualTriggerRule: AutomationEngine._handle_manual,
    HomeAssistantRule: AutomationEngine._handle_home_assistant,
    HTTPPollRule: AutomationEngine._handle_http_poll,
 }
@@ -876,10 +1186,12 @@ def _assert_rule_handler_coverage() -> None:
        StartupRule,
        ApplicationRule,
        TimeOfDayRule,
+        SolarRule,
        SystemIdleRule,
        DisplayStateRule,
        MQTTRule,
        WebhookRule,
+        ManualTriggerRule,
        HomeAssistantRule,
        HTTPPollRule,
    }
@@ -0,0 +1,102 @@
+"""Outbound webhook action firing for the automation engine.
+
+When an automation activates or deactivates, any attached
+:class:`~ledgrab.storage.automation.WebhookAction` performs a best-effort
+outbound HTTP request (Discord / IFTTT / Zapier / Node-RED / Home Assistant
+webhooks). Firing is fire-and-forget: a hung or failing endpoint is logged and
+audited but never raises into the activation path.
+
+Security: the target URL is SSRF-gated via :func:`validate_polling_url` (LAN
+allowed so users can hit Node-RED / HA on their own network; loopback,
+link-local / cloud-metadata, multicast and reserved ranges blocked) at **both**
+save time (in the route) and fire time (here) — re-validating at fire time
+closes the DNS-rebinding window. Redirects are not followed.
+"""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+
+import httpx
+from fastapi import HTTPException
+
+from ledgrab.storage.automation import Automation, WebhookAction
+from ledgrab.utils import get_logger
+from ledgrab.utils.safe_source import validate_polling_url
+
+logger = get_logger(__name__)
+
+# A webhook must never stall the ~1 Hz evaluation loop.
+_WEBHOOK_TIMEOUT_S = 5.0
+
+
+def render_template(template: str, automation: Automation, event: str) -> str:
+    """Substitute the supported ``{{token}}`` placeholders in *template*.
+
+    Tokens: ``{{automation_name}}``, ``{{automation_id}}``, ``{{event}}``
+    (``activate``/``deactivate``), ``{{timestamp}}`` (ISO-8601 UTC). Unknown
+    tokens are left untouched.
+    """
+    replacements = {
+        "{{automation_name}}": automation.name,
+        "{{automation_id}}": automation.id,
+        "{{event}}": event,
+        "{{timestamp}}": datetime.now(timezone.utc).isoformat(),
+    }
+    out = template
+    for token, value in replacements.items():
+        out = out.replace(token, value)
+    return out
+
+
+def should_fire(action: WebhookAction, event: str) -> bool:
+    """Whether *action* fires for this transition (``activate``/``deactivate``)."""
+    return action.fire_on == event or action.fire_on == "both"
+
+
+async def fire_webhook_action(
+    action: WebhookAction,
+    automation: Automation,
+    event: str,
+) -> tuple[bool, str | None]:
+    """Fire a single webhook action. Best-effort: never raises.
+
+    Returns ``(ok, error)`` where ``ok`` is True on a 2xx response and
+    ``error`` is a short, secret-free reason on failure.
+    """
+    url = action.webhook_url.strip()
+    if not url:
+        return False, "no URL configured"
+
+    # Re-validate at fire time (DNS-rebinding window). HTTPException carries a
+    # 4xx detail; surface a short reason rather than raising into the engine.
+    try:
+        validate_polling_url(url)
+    except HTTPException as exc:
+        logger.warning("Webhook for '%s' blocked by SSRF policy: %s", automation.name, exc.detail)
+        return False, "blocked by SSRF policy"
+
+    body = render_template(action.body_template, automation, event)
+    headers = {"Content-Type": action.content_type or "application/json"}
+
+    try:
+        async with httpx.AsyncClient(timeout=_WEBHOOK_TIMEOUT_S, follow_redirects=False) as client:
+            kwargs: dict = {"headers": headers}
+            # Only attach a body for write methods with content to send.
+            if action.method in ("POST", "PUT") and body:
+                kwargs["content"] = body.encode("utf-8")
+            response = await client.request(action.method, url, **kwargs)
+    except Exception as exc:  # noqa: BLE001 — never propagate into activation
+        # Never log the rendered body or the exception repr (may carry the URL
+        # with embedded secrets) — the type name is enough to diagnose.
+        logger.warning("Webhook for '%s' failed: %s", automation.name, type(exc).__name__)
+        return False, f"request failed: {type(exc).__name__}"
+
+    ok = 200 <= response.status_code < 300
+    if not ok:
+        logger.warning("Webhook for '%s' returned HTTP %d", automation.name, response.status_code)
+        return False, f"HTTP {response.status_code}"
+    logger.info(
+        "Webhook for '%s' fired on %s (HTTP %d)", automation.name, event, response.status_code
+    )
+    return True, None
@@ -113,6 +113,23 @@ class CalibrationConfig:
    skip_leds_end: int = 0
    # Border width: how many pixels from the screen edge to sample
    border_width: int = 10
+    # Region of interest (simple mode): sample only this sub-rectangle of the
+    # frame (fractions 0..1). Defaults to the full frame. Lets a user exclude
+    # HUDs/taskbars/letterboxing from the sampled border colours.
+    roi_x: float = 0.0
+    roi_y: float = 0.0
+    roi_width: float = 1.0
+    roi_height: float = 1.0
+    # Blend border pixels in linear light (perceptually correct averaging)
+    # instead of gamma-encoded sRGB. Off by default = unchanged behaviour.
+    linear_blend: bool = False
+    # Spatio-temporal dither the final 8-bit quantization to reduce banding.
+    dither: bool = False
+
+    @property
+    def has_roi(self) -> bool:
+        """True when the ROI is narrower than the full frame."""
+        return self.roi_x > 0.0 or self.roi_y > 0.0 or self.roi_width < 1.0 or self.roi_height < 1.0

    def build_segments(self) -> List[CalibrationSegment]:
        """Derive segment list from core parameters."""
@@ -337,6 +354,8 @@ class PixelMapper:
        """
        self.calibration = calibration
        self.interpolation_mode = interpolation_mode
+        # Per-frame counter driving the temporal dither phase.
+        self._dither_frame = 0

        # Validate calibration
        self.calibration.validate()
@@ -418,7 +437,16 @@ class PixelMapper:
        Scratch buffers are cached on ``self._edge_cache`` keyed by edge name;
        the shared kernel handles all allocations on first use.
        """
-        return average_edge_to_leds(edge_pixels, edge_name, led_count, self._edge_cache, edge_name)
+        return average_edge_to_leds(
+            edge_pixels,
+            edge_name,
+            led_count,
+            self._edge_cache,
+            edge_name,
+            linear=self.calibration.linear_blend,
+            dither=self.calibration.dither,
+            frame_index=self._dither_frame,
+        )

    def map_border_to_leds(self, border_pixels: BorderPixels) -> np.ndarray:
        """Map screen border pixels to LED colors.
@@ -437,6 +465,7 @@ class PixelMapper:
        """
        led_array = self._led_buf
        led_array[:] = 0
+        self._dither_frame += 1

        # Phase 1+2: Map edges and place at offset-adjusted positions (no np.roll)
        for i, segment in enumerate(self.calibration.segments):
@@ -502,6 +531,7 @@ class AdvancedPixelMapper:
    ):
        self.calibration = calibration
        self.interpolation_mode = interpolation_mode
+        self._dither_frame = 0
        calibration.validate()

        if interpolation_mode == "average":
@@ -588,7 +618,16 @@ class AdvancedPixelMapper:
        ``cache_key`` is an integer (e.g. line index) so multiple per-line
        edges can share the same ``self._edge_cache`` dict without colliding.
        """
-        return average_edge_to_leds(edge_pixels, edge_name, led_count, self._edge_cache, cache_key)
+        return average_edge_to_leds(
+            edge_pixels,
+            edge_name,
+            led_count,
+            self._edge_cache,
+            cache_key,
+            linear=self.calibration.linear_blend,
+            dither=self.calibration.dither,
+            frame_index=self._dither_frame,
+        )

    def _map_edge_fallback(
        self,
@@ -610,6 +649,7 @@ class AdvancedPixelMapper:
        """
        led_array = self._led_buf
        led_array[:] = 0
+        self._dither_frame += 1

        for i, line in enumerate(self.calibration.lines):
            frame = frames.get(line.picture_source_id)
@@ -656,6 +696,98 @@ def create_pixel_mapper(
    return PixelMapper(calibration, interpolation_mode)


+def solve_calibration(
+    led_count: int,
+    start_position: str,
+    layout: str,
+    corner_indices: List[int],
+    offset: int = 0,
+) -> "CalibrationConfig":
+    """Derive a CalibrationConfig from 4 corner tap indices.
+
+    Given the LED-strip indices where the user tapped each physical corner of
+    the screen (in strip-walk order matching *start_position* and *layout*),
+    compute per-edge LED counts that are consistent with
+    ``EDGE_ORDER``/``EDGE_REVERSE`` and round-trip through
+    ``build_segments()``.
+
+    Args:
+        led_count: Total number of LEDs on the strip.
+        start_position: Starting corner of the strip
+            (``"top_left"``, ``"top_right"``, ``"bottom_left"``,
+            ``"bottom_right"``).
+        layout: Winding direction (``"clockwise"`` or
+            ``"counterclockwise"``).
+        corner_indices: Four strip indices, one per screen corner, in the
+            same order as the strip walk defined by ``EDGE_ORDER`` for the
+            given *(start_position, layout)* pair.  Index 0 is the start
+            corner, index 1 is the second corner reached while walking,
+            etc.  Indices may wrap around (i.e. the last segment may
+            straddle the physical end of the strip).
+        offset: Physical LED offset stored directly on the config (0 = none).
+
+    Returns:
+        ``CalibrationConfig`` in simple mode with per-edge counts filled in.
+
+    Raises:
+        ValueError: If *start_position*, *layout*, or the number of
+            corner indices is invalid.
+    """
+    key = (start_position, layout)
+    if key not in EDGE_ORDER:
+        raise ValueError(
+            f"Invalid start_position/layout combination: {start_position!r}/{layout!r}"
+        )
+    if len(corner_indices) != 4:
+        raise ValueError(f"corner_indices must have exactly 4 entries, got {len(corner_indices)}")
+    if led_count <= 0:
+        raise ValueError(f"led_count must be positive, got {led_count}")
+
+    edge_order = EDGE_ORDER[key]  # 4 edges in strip-walk order
+
+    # Compute per-edge LED counts from consecutive corner indices.
+    # The i-th edge spans from corner_indices[i] to corner_indices[(i+1) % 4],
+    # wrapping around led_count if necessary.
+    edge_counts: dict[str, int] = {}
+    for i, edge in enumerate(edge_order):
+        start_idx = corner_indices[i] % led_count
+        end_idx = corner_indices[(i + 1) % 4] % led_count
+        if end_idx > start_idx:
+            count = end_idx - start_idx
+        elif end_idx == start_idx:
+            # Adjacent taps on the same index → 0-LED edge
+            count = 0
+        else:
+            # Wrap-around: strip crosses the physical end
+            count = (led_count - start_idx) + end_idx
+        edge_counts[edge] = count
+
+    cfg = CalibrationConfig(
+        mode="simple",
+        layout=layout,
+        start_position=start_position,
+        leds_top=edge_counts.get("top", 0),
+        leds_right=edge_counts.get("right", 0),
+        leds_bottom=edge_counts.get("bottom", 0),
+        leds_left=edge_counts.get("left", 0),
+        offset=offset,
+    )
+
+    logger.info(
+        "solve_calibration: start=%s layout=%s corner_indices=%s "
+        "-> top=%d right=%d bottom=%d left=%d offset=%d",
+        start_position,
+        layout,
+        corner_indices,
+        cfg.leds_top,
+        cfg.leds_right,
+        cfg.leds_bottom,
+        cfg.leds_left,
+        offset,
+    )
+    return cfg
+
+
 def create_default_calibration(
    led_count: int,
    aspect_width: int = 16,
@@ -720,6 +852,30 @@ def create_default_calibration(
    right_count = max(1, right_count)
    left_count = max(1, left_count)

+    # The max(1, ...) floors above can push the total above led_count for
+    # small counts (e.g. led_count=5 -> top=2,right=1,bottom=2,left=1 = 6).
+    # Trim the largest edge that stays >= 1 until the total matches exactly.
+    edge_order = ["bottom", "top", "right", "left"]
+    counts = {
+        "bottom": bottom_count,
+        "top": top_count,
+        "right": right_count,
+        "left": left_count,
+    }
+    overshoot = sum(counts.values()) - led_count
+    while overshoot > 0:
+        # Pick the largest edge that can still be reduced (stays >= 1).
+        trimmable = [e for e in edge_order if counts[e] > 1]
+        if not trimmable:
+            break
+        target_edge = max(trimmable, key=lambda e: counts[e])
+        counts[target_edge] -= 1
+        overshoot -= 1
+    bottom_count = counts["bottom"]
+    top_count = counts["top"]
+    right_count = counts["right"]
+    left_count = counts["left"]
+
    config = CalibrationConfig(
        layout="clockwise",
        start_position="bottom_left",
@@ -774,6 +930,8 @@ def calibration_from_dict(data: dict) -> CalibrationConfig:
                offset=data.get("offset", 0),
                skip_leds_start=data.get("skip_leds_start", 0),
                skip_leds_end=data.get("skip_leds_end", 0),
+                linear_blend=bool(data.get("linear_blend", False)),
+                dither=bool(data.get("dither", False)),
            )
            config.validate()
            return config
@@ -799,6 +957,12 @@ def calibration_from_dict(data: dict) -> CalibrationConfig:
            skip_leds_start=data.get("skip_leds_start", 0),
            skip_leds_end=data.get("skip_leds_end", 0),
            border_width=data.get("border_width", 10),
+            roi_x=data.get("roi_x", 0.0),
+            roi_y=data.get("roi_y", 0.0),
+            roi_width=data.get("roi_width", 1.0),
+            roi_height=data.get("roi_height", 1.0),
+            linear_blend=bool(data.get("linear_blend", False)),
+            dither=bool(data.get("dither", False)),
        )

        config.validate()
@@ -843,6 +1007,10 @@ def calibration_to_dict(config: CalibrationConfig) -> dict:
            result["skip_leds_start"] = config.skip_leds_start
        if config.skip_leds_end > 0:
            result["skip_leds_end"] = config.skip_leds_end
+        if config.linear_blend:
+            result["linear_blend"] = True
+        if config.dither:
+            result["dither"] = True
        return result

    # Simple mode
@@ -870,4 +1038,14 @@ def calibration_to_dict(config: CalibrationConfig) -> dict:
        result["skip_leds_end"] = config.skip_leds_end
    if config.border_width != 10:
        result["border_width"] = config.border_width
+    # Include ROI only when it is not the full frame
+    if config.has_roi:
+        result["roi_x"] = config.roi_x
+        result["roi_y"] = config.roi_y
+        result["roi_width"] = config.roi_width
+        result["roi_height"] = config.roi_height
+    if config.linear_blend:
+        result["linear_blend"] = True
+    if config.dither:
+        result["dither"] = True
    return result
@@ -0,0 +1,422 @@
+"""Calibration session lifecycle and per-LED chase driver.
+
+Provides two things:
+1. ``set_calibration_pixel`` — direct per-index LED write for the chase
+   (added beside ``set_test_mode`` on ``ProcessorManager`` via the mixin, but
+   kept here to avoid growing device_test_mode.py further).
+2. ``CalibrationSession`` — single-active-session guard with idle timeout and
+   guaranteed stop/restore contract.
+
+Stop / restore contract (required by Phase 3 UI)
+-------------------------------------------------
+- ``start(device_id)``:
+    * If a target is currently processing on *device_id*, stop it and record
+      its ``target_id`` as ``_prior_target_id``.
+    * Send the device to black (chase start state).
+    * Record session as active with a fresh ``last_activity`` timestamp.
+    * Only one active session is allowed at a time; starting a new one on any
+      device while another is active calls ``stop()`` on the old one first.
+- ``position(index, window)``:
+    * Validates ``index < led_count``; raises ``ValueError`` on out-of-range.
+    * Sends a chase pixel (bright white centre ±window dim neighbours).
+    * Updates ``last_activity``.
+- ``stop()`` / ``cancel()``:
+    * Sends all-black to clear the device.
+    * If ``_prior_target_id`` was recorded, calls ``start_processing`` to
+      restart it.
+    * Clears the session state.
+    * NEVER leaves the device dark or stuck in chase.
+- Idle timeout (``IDLE_TIMEOUT_SECONDS``, default 60 s):
+    * A background asyncio task checks ``last_activity``; if the session has
+      been idle longer than the timeout, ``stop()`` is called automatically.
+    * The timeout task is cancelled when ``stop()`` is called explicitly.
+
+Notes
+-----
+- ``set_calibration_pixel`` reuses ``_get_idle_client`` /
+  ``_send_pixels_to_device`` from ``DeviceTestModeMixin``; no new connection
+  management is needed.
+- The session holds a reference to the ``ProcessorManager`` so it can call
+  ``stop_processing`` / ``start_processing``.
+- Thread-safety: all public methods are ``async``; the idle-timeout callback
+  schedules itself on the running event loop via ``asyncio.ensure_future``.
+"""
+
+import asyncio
+from datetime import datetime, timezone
+from typing import TYPE_CHECKING
+
+from ledgrab.utils import get_logger
+
+if TYPE_CHECKING:
+    from ledgrab.core.processing.processor_manager import ProcessorManager
+
+logger = get_logger(__name__)
+
+# ── Constants ────────────────────────────────────────────────────────────────
+
+IDLE_TIMEOUT_SECONDS: int = 60
+"""Auto-stop a calibration session after this many seconds of inactivity."""
+
+_CHASE_CENTER_COLOR: tuple[int, int, int] = (255, 255, 255)
+"""Bright white for the chase centre pixel."""
+
+_CHASE_WING_COLOR: tuple[int, int, int] = (60, 60, 60)
+"""Dim grey for ±window neighbour pixels."""
+
+
+# ── Mixin: per-index chase driver ────────────────────────────────────────────
+
+
+class CalibrationChaseMixin:
+    """Adds ``set_calibration_pixel`` to ``ProcessorManager``.
+
+    Requires the same host-class attributes as ``DeviceTestModeMixin``:
+    ``_devices``, ``_processors``, ``_idle_clients``.
+    Inherits ``_send_pixels_to_device`` and ``_get_idle_client`` from
+    ``DeviceTestModeMixin`` (both already on ``ProcessorManager``).
+    """
+
+    async def set_calibration_pixel(
+        self,
+        device_id: str,
+        index: int,
+        color: tuple[int, int, int] = _CHASE_CENTER_COLOR,
+        window: int = 1,
+    ) -> None:
+        """Light a single LED index (plus optional ±window neighbours) on a device.
+
+        Sends a full pixel array to avoid partial-frame artefacts.  The centre
+        LED is set to *color*; the ``window`` neighbours on each side are set to
+        ``_CHASE_WING_COLOR`` (dim grey) so the user can see which direction the
+        strip is wound.
+
+        Args:
+            device_id: Target device ID (must be registered).
+            index: LED index to light (0-based).  Must be < ``led_count``.
+            color: RGB tuple for the centre LED (default bright white).
+            window: Number of neighbouring LEDs to dim on each side (default 1).
+
+        Raises:
+            ValueError: If *device_id* is not registered or *index* is out of
+                range.
+        """
+        if device_id not in self._devices:
+            raise ValueError(f"Device {device_id!r} not found")
+        ds = self._devices[device_id]
+        led_count = ds.led_count
+        if led_count <= 0:
+            raise ValueError(f"Device {device_id!r} has led_count={led_count}")
+        if not (0 <= index < led_count):
+            raise ValueError(
+                f"index {index} out of range for device {device_id!r} " f"(led_count={led_count})"
+            )
+
+        pixels: list[tuple[int, int, int]] = [(0, 0, 0)] * led_count
+        pixels[index] = color
+        for offset in range(1, window + 1):
+            left = (index - offset) % led_count
+            right = (index + offset) % led_count
+            pixels[left] = _CHASE_WING_COLOR
+            pixels[right] = _CHASE_WING_COLOR
+        # Re-assign center last so on tiny strips (window >= led_count) the
+        # center LED always shows the full color rather than a wrapped wing.
+        pixels[index] = color
+
+        await self._send_pixels_to_device(device_id, pixels)
+
+        logger.debug(
+            "set_calibration_pixel: device=%s index=%d window=%d",
+            device_id,
+            index,
+            window,
+        )
+
+
+# ── Session lifecycle ─────────────────────────────────────────────────────────
+
+
+class CalibrationSession:
+    """Single-active calibration session with idle-timeout and stop/restore.
+
+    One instance is shared per application (singleton held by the API layer).
+    Only one session can be active at a time; starting a new session
+    automatically terminates the previous one.
+
+    All public methods that mutate session state acquire ``_lock`` so that
+    concurrent ``POST /session`` calls (or a ``stop`` racing with the idle
+    watchdog) cannot interleave and leave ``_prior_target_id`` stale.  The
+    watchdog calls the internal ``_teardown_locked`` helper which must only be
+    invoked when the lock is already held; if the lock is already taken the
+    watchdog simply exits, letting the holder finish teardown.
+    """
+
+    def __init__(self) -> None:
+        self._manager: "ProcessorManager | None" = None
+        self._device_id: str | None = None
+        self._led_count: int = 0
+        self._prior_target_id: str | None = None
+        self._last_activity: datetime | None = None
+        self._timeout_task: asyncio.Task | None = None
+        self._active: bool = False
+        self._lock: asyncio.Lock = asyncio.Lock()
+
+    # ── Public API ───────────────────────────────────────────────────────────
+
+    @property
+    def is_active(self) -> bool:
+        return self._active
+
+    @property
+    def device_id(self) -> str | None:
+        return self._device_id
+
+    @property
+    def led_count(self) -> int:
+        return self._led_count
+
+    @property
+    def last_activity(self) -> datetime | None:
+        return self._last_activity
+
+    async def start(self, device_id: str, manager: "ProcessorManager") -> None:
+        """Begin a calibration session on *device_id*.
+
+        If a session is already active (even on a different device), it is
+        stopped first.  If a target is currently processing on *device_id*, it
+        is stopped and remembered so it can be restored when this session ends.
+
+        Args:
+            device_id: The device to drive during calibration.
+            manager: Live ``ProcessorManager`` instance.
+
+        Raises:
+            ValueError: If *device_id* is not registered.
+        """
+        async with self._lock:
+            # Validate device before touching any state or awaiting
+            if device_id not in manager._devices:
+                raise ValueError(f"Device {device_id!r} not found")
+
+            ds = manager._devices[device_id]
+            led_count = ds.led_count
+
+            # Capture the prior running target NOW — before any await — so the
+            # value cannot be mutated by a concurrent call that sneaks in after
+            # the lock is released between awaits.
+            prior_target_id = manager.get_processing_target_for_device(device_id)
+
+            # Terminate any existing session while we still hold the lock.
+            # Call _teardown_locked directly (we already hold the lock).
+            if self._active:
+                logger.info(
+                    "CalibrationSession.start: stopping existing session on device=%s "
+                    "to start new one on device=%s",
+                    self._device_id,
+                    device_id,
+                )
+                await self._teardown_locked(cancelled=False)
+
+            # Stop any running target on this device and remember it for restore
+            if prior_target_id is not None:
+                logger.info(
+                    "CalibrationSession.start: stopping target %s on device %s for calibration",
+                    prior_target_id,
+                    device_id,
+                )
+                await manager.stop_processing(prior_target_id)
+
+            self._manager = manager
+            self._device_id = device_id
+            self._led_count = led_count
+            self._prior_target_id = prior_target_id
+            self._last_activity = datetime.now(timezone.utc)
+            self._active = True
+
+            # Clear the device to black so the chase starts from a clean state.
+            # send_clear_pixels re-raises on a double send failure; a transient
+            # failure here must NOT strand the session with _active=True and no
+            # watchdog — log and continue so the idle-timeout watchdog still gets
+            # armed (mirrors the guarded clear in _teardown_locked).
+            try:
+                await manager.send_clear_pixels(device_id)
+            except Exception as exc:
+                logger.warning(
+                    "CalibrationSession.start: failed to clear pixels on %s "
+                    "before chase (continuing): %s",
+                    device_id,
+                    exc,
+                )
+
+            # Start idle-timeout watchdog
+            self._timeout_task = asyncio.ensure_future(self._idle_watchdog())
+
+            logger.info(
+                "CalibrationSession.start: session started on device=%s led_count=%d "
+                "prior_target=%s",
+                device_id,
+                led_count,
+                prior_target_id,
+            )
+
+    async def position(self, index: int, window: int = 1) -> None:
+        """Drive the chase pixel to *index* on the active device.
+
+        Args:
+            index: LED index to illuminate (0-based, must be < led_count).
+            window: Number of dim neighbours on each side (default 1).
+
+        Raises:
+            RuntimeError: If no session is active.
+            ValueError: If *index* is out of range.
+        """
+        async with self._lock:
+            if not self._active or self._manager is None or self._device_id is None:
+                raise RuntimeError("No active calibration session")
+            if not (0 <= index < self._led_count):
+                raise ValueError(f"index {index} out of range (led_count={self._led_count})")
+
+            self._last_activity = datetime.now(timezone.utc)
+            await self._manager.set_calibration_pixel(self._device_id, index, window=window)
+
+            logger.debug(
+                "CalibrationSession.position: device=%s index=%d window=%d",
+                self._device_id,
+                index,
+                window,
+            )
+
+    async def stop(self) -> None:
+        """End the session: clear the device and restore the prior target.
+
+        Safe to call even if no session is active (no-op).
+        """
+        async with self._lock:
+            await self._teardown_locked(cancelled=False)
+
+    async def cancel(self) -> None:
+        """Alias for ``stop()`` — ends the session without applying calibration."""
+        async with self._lock:
+            await self._teardown_locked(cancelled=True)
+
+    def get_state(self) -> dict:
+        """Return a snapshot of the current session state for API responses."""
+        return {
+            "active": self._active,
+            "device_id": self._device_id,
+            "led_count": self._led_count,
+            "prior_target_id": self._prior_target_id,
+            "last_activity": (self._last_activity.isoformat() if self._last_activity else None),
+        }
+
+    # ── Internal ─────────────────────────────────────────────────────────────
+
+    async def _teardown_locked(self, cancelled: bool) -> None:
+        """Clear the device, restore the prior target, and reset state.
+
+        MUST be called with ``self._lock`` already held by the caller.
+        Safe to call when already inactive (no-op).
+        """
+        if not self._active:
+            return
+
+        device_id = self._device_id
+        manager = self._manager
+        prior_target_id = self._prior_target_id
+
+        # Cancel the idle watchdog — but only if we are NOT running inside it.
+        # Awaiting the current task would deadlock.
+        if (
+            self._timeout_task is not None
+            and self._timeout_task is not asyncio.current_task()
+            and not self._timeout_task.done()
+        ):
+            self._timeout_task.cancel()
+            try:
+                await self._timeout_task
+            except asyncio.CancelledError:
+                pass
+        self._timeout_task = None
+
+        # Reset state before side-effects so re-entrant calls are no-ops
+        self._active = False
+        self._device_id = None
+        self._led_count = 0
+        self._prior_target_id = None
+        self._last_activity = None
+        self._manager = None
+
+        if manager is None or device_id is None:
+            return
+
+        # 1. Clear the device to black
+        try:
+            await manager.send_clear_pixels(device_id)
+        except Exception as exc:
+            logger.warning(
+                "CalibrationSession._teardown: failed to clear pixels on %s: %s",
+                device_id,
+                exc,
+            )
+
+        # 2. Restore the prior target (if any)
+        if prior_target_id is not None:
+            try:
+                await manager.start_processing(prior_target_id)
+                logger.info(
+                    "CalibrationSession._teardown: restored target %s on device %s",
+                    prior_target_id,
+                    device_id,
+                )
+            except Exception as exc:
+                logger.error(
+                    "CalibrationSession._teardown: failed to restore target %s on " "device %s: %s",
+                    prior_target_id,
+                    device_id,
+                    exc,
+                )
+
+        action = "cancel" if cancelled else "stop"
+        logger.info(
+            "CalibrationSession.%s: session ended on device=%s prior_target=%s",
+            action,
+            device_id,
+            prior_target_id,
+        )
+
+    async def _idle_watchdog(self) -> None:
+        """Background task: auto-stop the session after IDLE_TIMEOUT_SECONDS.
+
+        Tries to acquire ``_lock`` when the timeout fires.  If the lock is
+        already held (e.g. a concurrent ``stop()`` is in progress) the
+        ``acquire`` will wait; once it gets the lock, ``_teardown_locked``
+        is a no-op if the session was already ended by the other caller.
+        """
+        try:
+            while True:
+                await asyncio.sleep(5)
+                if not self._active or self._last_activity is None:
+                    break
+                elapsed = (datetime.now(timezone.utc) - self._last_activity).total_seconds()
+                if elapsed >= IDLE_TIMEOUT_SECONDS:
+                    logger.warning(
+                        "CalibrationSession._idle_watchdog: session on device=%s "
+                        "idle for %.0fs — auto-stopping",
+                        self._device_id,
+                        elapsed,
+                    )
+                    async with self._lock:
+                        await self._teardown_locked(cancelled=False)
+                    break
+        except asyncio.CancelledError:
+            pass
+
+
+# ── Module-level singleton ────────────────────────────────────────────────────
+
+_session: CalibrationSession = CalibrationSession()
+
+
+def get_calibration_session() -> CalibrationSession:
+    """Return the module-level singleton ``CalibrationSession``."""
+    return _session
@@ -23,6 +23,9 @@ from typing import Any, Callable, Dict, Hashable, Tuple

 import numpy as np

+from ledgrab.utils.dither import ordered_dither_quantize
+from ledgrab.utils.linear_light import linear_to_srgb_float, linear_to_srgb_uint8, srgb_to_linear
+
 # Cache value layout — kept as a tuple for the small per-frame cost of
 # tuple unpacking vs the readability of a dataclass. The first two entries
 # are the (edge_len, led_count) signature used to detect a re-build.
@@ -75,6 +78,9 @@ def average_edge_to_leds(
    led_count: int,
    cache: Dict[Hashable, _CacheEntry],
    cache_key: Hashable,
+    linear: bool = False,
+    dither: bool = False,
+    frame_index: int = 0,
 ) -> np.ndarray:
    """Vectorised average colour per LED segment.

@@ -82,6 +88,14 @@ def average_edge_to_leds(
    over axis=0 (collapsing rows), then segment along the width; for
    left/right edges we average over axis=1 then segment along the height.

+    When ``linear`` is True the pixels are decoded to linear light before
+    averaging and re-encoded to sRGB at the end — perceptually correct
+    blending at a small extra cost (a LUT decode of the input + an analytic
+    encode of the per-LED result).
+
+    When ``dither`` is True the final 8-bit quantization is spatio-temporally
+    dithered (using ``frame_index``) to suppress gradient banding.
+
    Returns a view into the caller-owned cache's ``out_uint8`` buffer —
    do NOT retain the result across calls without copying.
    """
@@ -110,8 +124,13 @@ def average_edge_to_leds(
        out_uint8,
    ) = entry

+    # Decode to linear light first so both the row/column collapse and the
+    # per-segment mean happen in physically-linear space. ``src`` is float32
+    # in [0, 1] (linear) or the raw uint8 sRGB pixels otherwise.
+    src = srgb_to_linear(edge_pixels) if linear else edge_pixels
+
    # Mean into pre-allocated buffer (no intermediate float64 array)
-    np.mean(edge_pixels, axis=axis, out=edge_1d_buf)
+    np.mean(src, axis=axis, out=edge_1d_buf)

    # Cumulative sum so each LED segment's sum is two array lookups apart.
    cumsum_buf[0] = 0
@@ -122,8 +141,16 @@ def average_edge_to_leds(
    np.take(cumsum_buf, starts, axis=0, out=starts_buf)
    np.subtract(sums_buf, starts_buf, out=sums_buf)
    np.divide(sums_buf, lengths, out=sums_buf)
-    np.clip(sums_buf, 0, 255, out=sums_buf)
-    np.copyto(out_uint8, sums_buf, casting="unsafe")
+    if dither:
+        # sums_buf is linear [0,1] or sRGB [0,255]; quantize with dithering.
+        srgb_f = linear_to_srgb_float(sums_buf) if linear else sums_buf
+        np.copyto(out_uint8, ordered_dither_quantize(srgb_f, frame_index))
+    elif linear:
+        # sums_buf holds linear [0, 1] averages — re-encode to sRGB uint8.
+        np.copyto(out_uint8, linear_to_srgb_uint8(sums_buf))
+    else:
+        np.clip(sums_buf, 0, 255, out=sums_buf)
+        np.copyto(out_uint8, sums_buf, casting="unsafe")
    return out_uint8


@@ -159,6 +159,37 @@ def capture_display(display_index: int = 0) -> ScreenCapture:
        raise RuntimeError(f"Screen capture failed: {e}")


+def crop_screen_capture(
+    sc: ScreenCapture,
+    roi_x: float,
+    roi_y: float,
+    roi_width: float,
+    roi_height: float,
+) -> ScreenCapture:
+    """Crop a capture to a relative region-of-interest rectangle (fractions 0..1).
+
+    Sampling only a sub-rectangle of the frame lets a user exclude HUDs, task
+    bars, or letterboxing so they don't pollute the border colours. Returns the
+    original capture unchanged for a full-frame ROI (fast path). The cropped
+    image is a numpy view (no copy); out-of-range/degenerate ROIs are clamped so
+    at least a 1x1 region remains.
+    """
+    if roi_x <= 0.0 and roi_y <= 0.0 and roi_width >= 1.0 and roi_height >= 1.0:
+        return sc
+    h, w = sc.image.shape[:2]
+    x0 = max(0, min(w - 1, int(round(roi_x * w))))
+    y0 = max(0, min(h - 1, int(round(roi_y * h))))
+    x1 = max(x0 + 1, min(w, int(round((roi_x + roi_width) * w))))
+    y1 = max(y0 + 1, min(h, int(round((roi_y + roi_height) * h))))
+    cropped = sc.image[y0:y1, x0:x1]
+    return ScreenCapture(
+        image=cropped,
+        width=x1 - x0,
+        height=y1 - y0,
+        display_index=sc.display_index,
+    )
+
+
 def extract_border_pixels(screen_capture: ScreenCapture, border_width: int = 10) -> BorderPixels:
    """Extract border pixels from screen capture.

@@ -19,6 +19,17 @@ logger = get_logger(__name__)

 ARDUINO_RESET_DELAY = 2.0  # seconds to wait after opening serial for Arduino bootloader

+# Settle time between sending the final black frame and closing the port.
+# Closing the serial port deasserts DTR, which triggers the Arduino
+# auto-reset on most Adalight boards. ``flush()`` only guarantees the bytes
+# left the host UART — NOT that the board received the frame, parsed it, and
+# ran the LED ``show()`` before the reset wipes its RAM. Without this pause
+# the reset intermittently wins the race and the strip latches its last lit
+# frame instead of going dark (observed as "the strip sometimes stays on
+# after the target/automation stops"). 150 ms is far more than the few ms a
+# board needs to paint one frame, and it's a one-shot cost on teardown.
+BLACK_FRAME_SETTLE_DELAY = 0.15
+

 def parse_adalight_url(url: str) -> Tuple[str, int]:
    """Backwards-compatible alias for :func:`parse_serial_url`."""
@@ -126,6 +137,10 @@ class AdalightClient(LEDClient):
                )
                await loop.run_in_executor(executor, self._serial.write, frame)
                await loop.run_in_executor(executor, self._serial.flush)
+                # Let the board parse the frame and run show() before close()
+                # toggles DTR and resets it — otherwise the strip can latch its
+                # last lit frame instead of going dark. See BLACK_FRAME_SETTLE_DELAY.
+                await asyncio.sleep(BLACK_FRAME_SETTLE_DELAY)
                logger.info(f"Adalight black frame sent and flushed: {self._port}")
            except Exception as e:
                logger.warning(f"Failed to send black frame on close: {e}")
@@ -154,6 +154,15 @@ class DDPClient:
        all buses (observed in multi-bus setups). We reorder pixel channels
        here so the hardware receives the correct byte order directly.

+        TODO(ddp-multibus): currently UNUSED — ``send_pixels_numpy`` (the hot
+        send path) does NOT call this, so the per-bus color-order config
+        captured by ``set_buses`` is never applied to outgoing pixels. This
+        is intentionally left in place (not deleted) because it encodes real
+        multi-bus handling: if a multi-bus WLED setup needs per-bus byte
+        reordering, wire this into ``send_pixels_numpy`` before the payload
+        view is built (note it allocates a copy, so only call it when
+        ``self._buses`` actually requires reordering).
+
        Args:
            pixel_array: (N, 3) uint8 numpy array in RGB order

@@ -79,6 +79,9 @@ class HueConfig(BaseDeviceConfig):
    hue_username: str = ""
    hue_client_key: str = ""
    hue_entertainment_group_id: str = ""
+    # Map the strip across the entertainment configuration's channels
+    # (gradient-lightstrip segments) instead of one record per light.
+    hue_gradient_mode: bool = True


@dataclass(frozen=True)
@@ -115,6 +118,8 @@ class LIFXConfig(BaseDeviceConfig):

    device_type: Literal["lifx"] = "lifx"
    lifx_min_interval_ms: int = 50
+    # Per-zone/tile streaming (Z/Beam multizone, Tile/Canvas matrix) vs single colour.
+    lifx_per_zone: bool = False


@dataclass(frozen=True)
@@ -153,6 +158,8 @@ class NanoleafConfig(BaseDeviceConfig):
    device_type: Literal["nanoleaf"] = "nanoleaf"
    nanoleaf_token: str = ""
    nanoleaf_min_interval_ms: int = 100
+    # Per-panel extControl UDP streaming (addresses each panel) vs single colour.
+    nanoleaf_per_panel: bool = False


@dataclass(frozen=True)
@@ -36,6 +36,7 @@ from zeroconf.asyncio import AsyncServiceBrowser, AsyncServiceInfo, AsyncZerocon

 from ledgrab.core.devices.serial_transport import list_serial_ports
 from ledgrab.core.devices.wled_provider import WLED_MDNS_TYPE
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.utils import get_logger
 from ledgrab.utils.platform import is_android

@@ -286,3 +287,34 @@ class DiscoveryWatcher:
            )
        except Exception as e:
            logger.debug("Discovery watcher: fire_event failed: %s", e)
+
+        # Audit record — best-effort, thread-safe (recorder marshals via
+        # call_soon_threadsafe when called from the zeroconf thread).
+        try:
+            from ledgrab.core.activity_log.recorder import get_module_recorder
+            from ledgrab.core.activity_log.sanitize import sanitize_display
+
+            rec = get_module_recorder()
+            if rec is not None:
+                is_discovered = event_type == "device_discovered"
+                action = "device.discovered" if is_discovered else "device.lost"
+                severity = ActivitySeverity.INFO if is_discovered else ActivitySeverity.WARNING
+                verb = "discovered" if is_discovered else "lost"
+                # Sanitize mDNS-advertised strings before they enter the log.
+                # entry.name and entry.url are unauthenticated, attacker-controlled
+                # values; strip control chars, ANSI escapes, and NUL before use.
+                safe_name = sanitize_display(entry.name)
+                safe_url = sanitize_display(entry.url)
+                rec.record(
+                    category=ActivityCategory.DEVICE,
+                    action=action,
+                    severity=severity,
+                    actor="system",
+                    entity_type="device",
+                    entity_id=entry.url,
+                    entity_name=safe_name,
+                    message=f"Device '{safe_name}' {verb} at {safe_url}",
+                    metadata={"url": safe_url, "device_type": entry.device_type},
+                )
+        except Exception as e:
+            logger.debug("Discovery watcher: audit record failed: %s", e)
@@ -9,6 +9,7 @@ from typing import List, Tuple
 import numpy as np

 from ledgrab.core.devices.led_client import DeviceHealth, LEDClient
+from ledgrab.core.devices.pixel_reduce import resample_to_n
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)
@@ -24,15 +25,40 @@ COLOR_SPACE_RGB = 0x00
 HEADER_SIZE = 16


+def parse_entertainment_channels(config_json: dict) -> List[int]:
+    """Extract ordered channel ids from an entertainment_configuration GET.
+
+    Entertainment API v2 keys stream records by *channel*, not by light. A
+    plain bulb contributes one channel; a gradient lightstrip contributes up
+    to five (one per segment). We order channels left-to-right by their
+    spatial ``position`` (x then y) so a strip maps across the segments in
+    physical order. Channels without a position fall back to channel-id order.
+    """
+    data = config_json.get("data") or []
+    if not data:
+        return []
+    channels = data[0].get("channels") or []
+
+    def _key(c: dict) -> tuple:
+        pos = c.get("position") or {}
+        return (pos.get("x", 0.0), pos.get("y", 0.0), c.get("channel_id", 0))
+
+    ordered = sorted(channels, key=_key)
+    return [int(c["channel_id"]) for c in ordered if "channel_id" in c]
+
+
 def _build_entertainment_frame(
-    lights: List[Tuple[int, int, int]],
+    colors: List[Tuple[int, int, int]],
    brightness: int = 255,
    sequence: int = 0,
+    channel_ids: List[int] | None = None,
 ) -> bytes:
    """Build a Hue Entertainment API v2 UDP frame.

-    Each light gets 7 bytes: [light_id(2B)][R(2B)][G(2B)][B(2B)]
-    Colors are 16-bit (0-65535). We scale 8-bit RGB + brightness.
+    Each record is 7 bytes: [channel_id(1B)][R(2B)][G(2B)][B(2B)]. Colors are
+    16-bit (0-65535). Record ``i`` is keyed by ``channel_ids[i]`` when a
+    channel map is supplied (gradient-segment mode); otherwise it falls back
+    to the 0-based index (one light = one channel, the legacy behaviour).
    """
    # Header
    header = bytearray(HEADER_SIZE)
@@ -45,15 +71,15 @@ def _build_entertainment_frame(
    header[14] = COLOR_SPACE_RGB
    header[15] = 0x00  # reserved

-    # Light data
+    # Channel data
    # Note: brightness already applied by processor loop (_cached_brightness)
    data = bytearray()
-    for idx, (r, g, b) in enumerate(lights):
-        light_id = idx  # 0-based light index in entertainment group
+    for idx, (r, g, b) in enumerate(colors):
+        channel_id = channel_ids[idx] if channel_ids and idx < len(channel_ids) else idx
        r16 = int(r * 257)  # scale 0-255 to 0-65535
        g16 = int(g * 257)
        b16 = int(b * 257)
-        data += struct.pack(">BHHH", light_id, r16, g16, b16)
+        data += struct.pack(">BHHH", channel_id & 0xFF, r16, g16, b16)

    return bytes(header) + bytes(data)

@@ -62,7 +88,11 @@ class HueClient(LEDClient):
    """LED client for Philips Hue Entertainment API streaming.

    Uses UDP (optionally DTLS) to stream color data at ~25 fps to a Hue
-    entertainment group. Each light in the group is treated as one "LED".
+    entertainment group. In ``gradient_mode`` (the default) the client
+    discovers the entertainment configuration's *channels* on connect and
+    maps the strip across them, so a gradient lightstrip shows a gradient
+    instead of a single averaged colour. With gradient mode off (or if
+    discovery fails) it falls back to one record per light by index.
    """

    def __init__(
@@ -72,6 +102,7 @@ class HueClient(LEDClient):
        hue_username: str = "",
        hue_client_key: str = "",
        hue_entertainment_group_id: str = "",
+        gradient_mode: bool = True,
        **kwargs,
    ):
        self._bridge_ip = url.replace("hue://", "").rstrip("/")
@@ -79,15 +110,55 @@ class HueClient(LEDClient):
        self._username = hue_username
        self._client_key = hue_client_key
        self._group_id = hue_entertainment_group_id
+        self._gradient_mode = gradient_mode
+        self._channel_ids: List[int] = []
        self._sock: socket.socket | None = None
        self._connected = False
        self._sequence = 0
        self._dtls_sock = None

+    @property
+    def device_led_count(self) -> int | None:
+        # In gradient mode the discovered channel count is authoritative.
+        if self._gradient_mode and self._channel_ids:
+            return len(self._channel_ids)
+        return self._led_count or None
+
+    async def _fetch_channels(self) -> None:
+        """Best-effort discovery of the entertainment configuration channels."""
+        import httpx
+
+        url = (
+            f"https://{self._bridge_ip}/clip/v2/resource/"
+            f"entertainment_configuration/{self._group_id}"
+        )
+        headers = {"hue-application-key": self._username}
+        async with httpx.AsyncClient(verify=False, timeout=5.0) as client:
+            resp = await client.get(url, headers=headers)
+            resp.raise_for_status()
+            self._channel_ids = parse_entertainment_channels(resp.json())
+        if self._channel_ids:
+            logger.info(
+                "Hue %s: mapped strip across %d channels", self._bridge_ip, len(self._channel_ids)
+            )
+
    async def connect(self) -> bool:
        # Activate entertainment streaming via REST API
        await self._activate_streaming(True)

+        # Best-effort channel discovery for gradient-segment mapping. On any
+        # failure (old bridge, network) we degrade to legacy 1-light-1-record.
+        if self._gradient_mode:
+            try:
+                await self._fetch_channels()
+            except Exception as exc:  # noqa: BLE001 — degrade, never fail connect
+                logger.warning(
+                    "Hue %s: channel discovery failed, using per-light mapping (%s)",
+                    self._bridge_ip,
+                    exc,
+                )
+                self._channel_ids = []
+
        # Open UDP socket for entertainment streaming
        self._sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
        self._sock.setblocking(False)
@@ -179,12 +250,19 @@ class HueClient(LEDClient):
        if not self._connected:
            return

-        if isinstance(pixels, np.ndarray):
-            light_colors = [tuple(pixels[i]) for i in range(min(len(pixels), self._led_count))]
-        else:
-            light_colors = pixels[: self._led_count]
+        # Resample the strip to the number of addressable elements: the
+        # discovered channel count in gradient mode, else the configured
+        # light count. ``resample_to_n`` spreads the strip spatially and
+        # handles both the np.ndarray and list-of-tuples inputs.
+        target = len(self._channel_ids) if self._channel_ids else self._led_count
+        colors = resample_to_n(pixels, target)

-        frame = _build_entertainment_frame(light_colors, brightness, self._sequence)
+        frame = _build_entertainment_frame(
+            colors,
+            brightness,
+            self._sequence,
+            channel_ids=self._channel_ids or None,
+        )
        self._sequence = (self._sequence + 1) & 0xFF

        try:
@@ -50,6 +50,7 @@ class HueDeviceProvider(LEDDeviceProvider):
            hue_username=config.hue_username,
            hue_client_key=config.hue_client_key,
            hue_entertainment_group_id=config.hue_entertainment_group_id,
+            gradient_mode=getattr(config, "hue_gradient_mode", True),
        )

    async def check_health(self, url: str, http_client, prev_health=None) -> DeviceHealth:
@@ -29,6 +29,7 @@ import numpy as np

 from ledgrab.core.devices.led_client import DeviceHealth, LEDClient
 from ledgrab.core.devices.pixel_reduce import average_color as _average_color
+from ledgrab.core.devices.pixel_reduce import resample_to_n
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)
@@ -41,6 +42,21 @@ MSG_GET_SERVICE = 2
 MSG_STATE_SERVICE = 3
 MSG_SET_POWER = 21
 MSG_SET_COLOR = 102
+# Multizone (Z / Beam) and matrix (Tile / Canvas) per-pixel messages.
+MSG_GET_COLOR_ZONES = 502
+MSG_STATE_ZONE = 503
+MSG_STATE_MULTIZONE = 506
+MSG_SET_EXTENDED_COLOR_ZONES = 510
+MSG_GET_DEVICE_CHAIN = 701
+MSG_STATE_DEVICE_CHAIN = 702
+MSG_SET_TILE_STATE_64 = 715
+
+_EXT_ZONE_MAX = 82  # SetExtendedColorZones carries a fixed 82-slot HSBK array
+_TILE_PIXELS = 64  # SetTileState64 always carries 64 HSBK
+_TILE_STRUCT_SIZE = 55  # bytes per Tile struct in StateDeviceChain
+_TILE_CHAIN_MAX = 16  # StateDeviceChain always lists 16 tile slots
+# Apply field for zone writes: 0=NO_APPLY (buffer), 1=APPLY, 2=APPLY_ONLY
+_ZONE_APPLY = 1

 # Frame field byte 0 of the protocol header: tagged=1, addressable=1, protocol=1024
 _FRAME_TAGGED = 0x3400
@@ -142,6 +158,105 @@ def _build_set_power_payload(on: bool, duration_ms: int = 0) -> bytes:
    return struct.pack("<HI", 65535 if on else 0, duration_ms & 0xFFFFFFFF)


+def _build_get_color_zones_payload(start: int = 0, end: int = 255) -> bytes:
+    """GetColorZones (502) payload: start_index(1) | end_index(1)."""
+    return struct.pack("<BB", start & 0xFF, end & 0xFF)
+
+
+def _pack_hsbk_array(hsbk_list: List[Tuple[int, int, int, int]], slots: int) -> bytes:
+    """Pack a fixed-length HSBK array, zero-padding unused slots.
+
+    Both SetExtendedColorZones (82 slots) and SetTileState64 (64 slots)
+    carry a fixed-size color array regardless of how many are in use.
+    """
+    body = bytearray()
+    for h, s, b, k in hsbk_list[:slots]:
+        body += struct.pack("<HHHH", h & 0xFFFF, s & 0xFFFF, b & 0xFFFF, k & 0xFFFF)
+    used = min(len(hsbk_list), slots)
+    body += b"\x00" * (8 * (slots - used))
+    return bytes(body)
+
+
+def _build_set_extended_color_zones_payload(
+    hsbk_list: List[Tuple[int, int, int, int]],
+    *,
+    duration_ms: int = 0,
+    apply: int = _ZONE_APPLY,
+    zone_index: int = 0,
+) -> bytes:
+    """SetExtendedColorZones (510): duration(4)|apply(1)|zone_index(2)|count(1)|82×HSBK."""
+    count = min(len(hsbk_list), _EXT_ZONE_MAX)
+    header = struct.pack(
+        "<IBHB", duration_ms & 0xFFFFFFFF, apply & 0xFF, zone_index & 0xFFFF, count & 0xFF
+    )
+    return header + _pack_hsbk_array(hsbk_list, _EXT_ZONE_MAX)
+
+
+def _build_set_tile_state64_payload(
+    hsbk_list: List[Tuple[int, int, int, int]],
+    *,
+    tile_index: int = 0,
+    x: int = 0,
+    y: int = 0,
+    width: int = 8,
+    duration_ms: int = 0,
+) -> bytes:
+    """SetTileState64 (715): tile_index|length|reserved|x|y|width|duration(4)|64×HSBK."""
+    header = struct.pack(
+        "<BBBBBBI",
+        tile_index & 0xFF,
+        1,  # length: number of tiles this packet addresses
+        0,  # reserved
+        x & 0xFF,
+        y & 0xFF,
+        width & 0xFF,
+        duration_ms & 0xFFFFFFFF,
+    )
+    return header + _pack_hsbk_array(hsbk_list, _TILE_PIXELS)
+
+
+def _parse_multizone_reply(raw: bytes) -> dict | None:
+    """Parse StateZone (503) / StateMultiZone (506) → ``{"count", "index"}``.
+
+    ``count`` is the device's *total* zone count; ``index`` is where this
+    packet's run starts. Returns ``None`` for any other message type.
+    """
+    if len(raw) < 36 + 2:
+        return None
+    msg_type = struct.unpack_from("<H", raw, 32)[0]
+    if msg_type not in (MSG_STATE_ZONE, MSG_STATE_MULTIZONE):
+        return None
+    count, index = struct.unpack_from("<BB", raw, 36)
+    return {"count": int(count), "index": int(index)}
+
+
+def _parse_state_device_chain(raw: bytes) -> dict | None:
+    """Parse StateDeviceChain (702) → ``{"start_index", "tiles": [(w, h), ...]}``.
+
+    Payload: start_index(1) | 16 × Tile(55 bytes) | tile_devices_count(1).
+    Within each 55-byte Tile struct, ``width`` is byte 16 and ``height``
+    byte 17 (after accel int16×4 + user_x/user_y float32×2). Returns ``None``
+    for any other message type.
+    """
+    min_len = 36 + 1 + _TILE_STRUCT_SIZE * _TILE_CHAIN_MAX + 1
+    if len(raw) < min_len:
+        return None
+    msg_type = struct.unpack_from("<H", raw, 32)[0]
+    if msg_type != MSG_STATE_DEVICE_CHAIN:
+        return None
+    payload_off = 36
+    start_index = raw[payload_off]
+    count_off = payload_off + 1 + _TILE_STRUCT_SIZE * _TILE_CHAIN_MAX
+    total = min(raw[count_off], _TILE_CHAIN_MAX)
+    tiles: list[tuple[int, int]] = []
+    for i in range(total):
+        base = payload_off + 1 + i * _TILE_STRUCT_SIZE
+        width = raw[base + 16]
+        height = raw[base + 17]
+        tiles.append((int(width), int(height)))
+    return {"start_index": int(start_index), "tiles": tiles}
+
+
 def _parse_state_service_reply(raw: bytes) -> dict | None:
    """Parse a LIFX StateService (discovery) reply.

@@ -163,15 +278,25 @@ def _parse_state_service_reply(raw: bytes) -> dict | None:


 class _LIFXProtocol(asyncio.DatagramProtocol):
-    """Write-only datagram protocol. Inbound replies dropped silently."""
+    """Datagram protocol that buffers inbound replies for zone/tile discovery.
+
+    The streaming hot path never reads replies, but per-zone setup needs the
+    StateMultiZone / StateDeviceChain answers. The buffer is bounded so a
+    chatty bulb can't grow it without limit during steady-state streaming.
+    """
+
+    _MAX_BUFFER = 64
+
+    def __init__(self) -> None:
+        self.transport: asyncio.DatagramTransport | None = None
+        self.received: list[bytes] = []

    def connection_made(self, transport):
        self.transport = transport

    def datagram_received(self, data, addr):
-        # LIFX bulbs sometimes echo back state on broadcast. We don't need it
-        # for streaming ambilight — discard.
-        pass
+        if len(self.received) < self._MAX_BUFFER:
+            self.received.append(bytes(data))

    def error_received(self, exc):
        logger.debug("LIFX UDP error: %s", exc)
@@ -186,6 +311,7 @@ class LIFXClient(LEDClient):
        led_count: int = 1,
        *,
        min_interval_s: float = DEFAULT_MIN_INTERVAL_S,
+        per_zone: bool = False,
    ):
        host, port = parse_lifx_url(url)
        self._host = host
@@ -197,6 +323,12 @@ class LIFXClient(LEDClient):
        self._connected = False
        self._next_tx_at: float = 0.0
        self._sequence: int = 0
+        # Per-pixel state. ``_mode`` is "single" until connect() probes the
+        # device and finds multizone (Z/Beam) or tile (matrix) support.
+        self._per_zone = per_zone
+        self._mode = "single"  # "single" | "multizone" | "tile"
+        self._zone_count = 0
+        self._tiles: list[tuple[int, int]] = []

    @property
    def host(self) -> str:
@@ -212,6 +344,12 @@ class LIFXClient(LEDClient):

    @property
    def device_led_count(self) -> int | None:
+        # In per-zone streaming the device's addressable element count is
+        # authoritative (zones for multizone, total pixels for tiles).
+        if self._mode == "multizone" and self._zone_count:
+            return self._zone_count
+        if self._mode == "tile" and self._tiles:
+            return sum(w * h for w, h in self._tiles)
        return self._led_count or None

    async def connect(self) -> bool:
@@ -227,9 +365,56 @@ class LIFXClient(LEDClient):
        self._transport = transport
        self._protocol = protocol  # type: ignore[assignment]
        self._connected = True
-        logger.info("LIFXClient connected to %s:%d", self._host, self._port)
+        if self._per_zone:
+            # Best-effort: probe for multizone / tile support. On timeout or
+            # an old single-colour bulb we silently stay in single-colour mode.
+            try:
+                await self._setup_zones()
+            except Exception as exc:  # noqa: BLE001 — degrade, never fail connect
+                logger.warning(
+                    "LIFX %s: per-zone setup failed, using single colour (%s)", self._host, exc
+                )
+                self._mode = "single"
+        logger.info(
+            "LIFXClient connected to %s:%d (per_zone=%s, mode=%s)",
+            self._host,
+            self._port,
+            self._per_zone,
+            self._mode,
+        )
        return True

+    async def _setup_zones(self) -> None:
+        """Probe the device for multizone / tile support and pick a mode.
+
+        Sends GetColorZones + GetDeviceChain and waits briefly for whichever
+        reply the device returns (a strip answers StateMultiZone; a tile chain
+        answers StateDeviceChain). Leaves ``_mode`` at "single" if neither.
+        """
+        if self._protocol is None:
+            return
+        self._protocol.received.clear()
+        self._send(MSG_GET_DEVICE_CHAIN, b"")
+        self._send(MSG_GET_COLOR_ZONES, _build_get_color_zones_payload())
+        loop = asyncio.get_running_loop()
+        deadline = loop.time() + 0.6
+        while loop.time() < deadline:
+            await asyncio.sleep(0.05)
+            for raw in list(self._protocol.received):
+                chain = _parse_state_device_chain(raw)
+                if chain and chain["tiles"]:
+                    self._tiles = chain["tiles"]
+                    self._mode = "tile"
+                    logger.info("LIFX %s: tile chain (%d tiles)", self._host, len(self._tiles))
+                    return
+                zones = _parse_multizone_reply(raw)
+                if zones and zones["count"] > 1:
+                    self._zone_count = min(zones["count"], _EXT_ZONE_MAX)
+                    self._mode = "multizone"
+                    logger.info("LIFX %s: multizone (%d zones)", self._host, self._zone_count)
+                    return
+        self._mode = "single"
+
    async def close(self) -> None:
        if self._transport is not None:
            try:
@@ -255,25 +440,63 @@ class LIFXClient(LEDClient):
        )
        self._transport.sendto(packet)

+    def _hsbk_list(
+        self,
+        pixels: List[Tuple[int, int, int]] | np.ndarray,
+        n: int,
+        scale: float,
+    ) -> List[Tuple[int, int, int, int]]:
+        """Resample the strip to ``n`` pixels and convert each to HSBK."""
+        out: List[Tuple[int, int, int, int]] = []
+        for r, g, b in resample_to_n(pixels, n):
+            if scale != 1.0:
+                r, g, b = int(r * scale), int(g * scale), int(b * scale)
+            out.append(rgb_to_hsbk(r, g, b))
+        return out
+
+    def _emit_pixels(
+        self,
+        pixels: List[Tuple[int, int, int]] | np.ndarray,
+        brightness: int,
+    ) -> None:
+        """Build and send the packet(s) for the current mode (single/zone/tile)."""
+        scale = max(0, min(255, brightness)) / 255.0 if brightness < 255 else 1.0
+        if self._mode == "multizone" and self._zone_count > 0:
+            hsbk = self._hsbk_list(pixels, self._zone_count, scale)
+            self._send(MSG_SET_EXTENDED_COLOR_ZONES, _build_set_extended_color_zones_payload(hsbk))
+            return
+        if self._mode == "tile" and self._tiles:
+            total = sum(w * h for w, h in self._tiles)
+            full = self._hsbk_list(pixels, total, scale)
+            offset = 0
+            for ti, (w, h) in enumerate(self._tiles):
+                n = w * h
+                chunk = full[offset : offset + n]
+                offset += n
+                self._send(
+                    MSG_SET_TILE_STATE_64,
+                    _build_set_tile_state64_payload(chunk, tile_index=ti, width=w),
+                )
+            return
+        # Single-colour fallback (every non-multizone/tile bulb).
+        r, g, b = _average_color(pixels)
+        if scale != 1.0:
+            r, g, b = int(r * scale), int(g * scale), int(b * scale)
+        h, s, br, k = rgb_to_hsbk(r, g, b)
+        self._send(MSG_SET_COLOR, _build_set_color_payload(h, s, br, k, duration_ms=0))
+
    async def send_pixels(
        self,
        pixels: List[Tuple[int, int, int]] | np.ndarray,
        brightness: int = 255,
    ) -> bool:
-        """Average the strip → HSBK → SetColor."""
+        """Stream per-zone/tile when detected, else average the strip → SetColor."""
        if not self.is_connected:
            raise RuntimeError("LIFXClient not connected")
        now = time.monotonic()
        if now < self._next_tx_at:
            return True
-        r, g, b = _average_color(pixels)
-        if brightness < 255:
-            scale = max(0, min(255, brightness)) / 255.0
-            r = int(r * scale)
-            g = int(g * scale)
-            b = int(b * scale)
-        h, s, br, k = rgb_to_hsbk(r, g, b)
-        self._send(MSG_SET_COLOR, _build_set_color_payload(h, s, br, k, duration_ms=0))
+        self._emit_pixels(pixels, brightness)
        self._next_tx_at = now + self._min_interval_s
        return True

@@ -288,14 +511,7 @@ class LIFXClient(LEDClient):
        now = time.monotonic()
        if now < self._next_tx_at:
            return
-        r, g, b = _average_color(pixels)
-        if brightness < 255:
-            scale = max(0, min(255, brightness)) / 255.0
-            r = int(r * scale)
-            g = int(g * scale)
-            b = int(b * scale)
-        h, s, br, k = rgb_to_hsbk(r, g, b)
-        self._send(MSG_SET_COLOR, _build_set_color_payload(h, s, br, k, duration_ms=0))
+        self._emit_pixels(pixels, brightness)
        self._next_tx_at = now + self._min_interval_s

    @property
@@ -51,6 +51,7 @@ class LIFXDeviceProvider(LEDDeviceProvider):
            config.device_url,
            led_count=config.led_count,
            min_interval_s=max(0.0, config.lifx_min_interval_ms / 1000.0),
+            per_zone=getattr(config, "lifx_per_zone", False),
        )

    async def check_health(self, url: str, http_client, prev_health=None) -> DeviceHealth:
@@ -6,12 +6,13 @@ handshake: the user holds the controller's power button for 5 seconds to
 open a 30-second pairing window, then we POST to ``/api/v1/new`` to claim
 an auth token. The token is long-lived and gets stored on the device.

-Once paired, color control is a simple ``PUT /api/v1/{token}/state`` with
-HSBT (hue / saturation / brightness; kelvin only matters when sat=0).
-LedGrab averages the incoming strip to one HSB triple. Per-panel streaming
-mode (``extControl`` UDP, ~60 Hz, addresses each panel individually) is
-documented but not implemented here — the MVP keeps the device acting as
-a single-pixel target like Yeelight / Hue.
+Two output modes:
+  * **Single-colour** (default): average the strip to one HSB triple and
+    ``PUT /api/v1/{token}/state`` — matches every other consumer-bulb driver.
+  * **Per-panel** (``per_panel=True``): enable ``extControl`` v2 and stream a
+    UDP packet per frame to port 60222, addressing each panel individually
+    (resampled from the strip). Enabled when the device config opts in; falls
+    back to single-colour if the controller/firmware can't stream.

 URL scheme: ``nanoleaf://<host>``. Port is fixed at 16021 on the protocol
 side. The auth token is stored separately on the device config, not in
@@ -23,6 +24,8 @@ Reference: https://forum.nanoleaf.me/docs/openapi
 from __future__ import annotations

 import asyncio
+import socket
+import struct
 from datetime import datetime, timezone
 from typing import List, Tuple
 from urllib.parse import urlparse
@@ -37,9 +40,62 @@ from ledgrab.utils import get_logger
 logger = get_logger(__name__)

 NANOLEAF_PORT = 16021
+# extControl v2 UDP streaming target port (fixed by the protocol).
+NANOLEAF_STREAM_PORT = 60222
 DEFAULT_MIN_INTERVAL_S = 0.1  # 10 Hz; HTTP per frame, plenty for averaged ambilight


+def order_panels(position_data: List[dict]) -> List[int]:
+    """Order panel IDs left-to-right, top-to-bottom from ``panelLayout`` data.
+
+    Each ``positionData`` entry has ``panelId``/``x``/``y``. We sort by (x, y)
+    so an ambient strip maps across the panels spatially. Entries without an
+    integer ``panelId`` (or the controller/rhythm panel, panelId 0) are dropped.
+    """
+    panels = [
+        p for p in position_data if isinstance(p.get("panelId"), int) and p.get("panelId", 0) != 0
+    ]
+    panels.sort(key=lambda p: (p.get("x", 0), p.get("y", 0)))
+    return [int(p["panelId"]) for p in panels]
+
+
+def map_pixels_to_panels(
+    pixels: List[Tuple[int, int, int]] | np.ndarray,
+    panel_ids: List[int],
+) -> List[Tuple[int, int, int, int]]:
+    """Resample an N-pixel strip to one ``(panel_id, r, g, b)`` per panel.
+
+    Nearest-neighbour resample: panel ``i`` of ``M`` samples strip pixel
+    ``floor(i * N / M)``. Returns black for an empty strip.
+    """
+    arr = np.asarray(pixels, dtype=np.uint8).reshape(-1, 3)
+    n_pix = len(arr)
+    n_panels = max(1, len(panel_ids))
+    out: List[Tuple[int, int, int, int]] = []
+    for i, pid in enumerate(panel_ids):
+        if n_pix == 0:
+            out.append((pid, 0, 0, 0))
+            continue
+        idx = min(n_pix - 1, (i * n_pix) // n_panels)
+        px = arr[idx]
+        out.append((pid, int(px[0]), int(px[1]), int(px[2])))
+    return out
+
+
+def build_extcontrol_v2_packet(panels: List[Tuple[int, int, int, int]]) -> bytes:
+    """Build a Nanoleaf extControl **v2** UDP streaming packet.
+
+    Wire format (all multi-byte fields big-endian):
+      ``uint16 nPanels`` then per panel
+      ``uint16 panelId, uint8 R, uint8 G, uint8 B, uint8 W, uint16 transitionTime``
+    ``transitionTime`` is in 100 ms units; 1 = a 100 ms ease for smooth motion.
+    """
+    parts = [struct.pack(">H", len(panels))]
+    for pid, r, g, b in panels:
+        parts.append(struct.pack(">HBBBBH", pid & 0xFFFF, r & 0xFF, g & 0xFF, b & 0xFF, 0, 1))
+    return b"".join(parts)
+
+
 def parse_nanoleaf_url(url: str) -> str:
    """Pull the host out of ``nanoleaf://host`` or accept a bare host.

@@ -131,6 +187,7 @@ class NanoleafClient(LEDClient):
        auth_token: str = "",
        min_interval_s: float = DEFAULT_MIN_INTERVAL_S,
        request_timeout_s: float = 3.0,
+        per_panel: bool = False,
    ):
        self._host = parse_nanoleaf_url(url)
        self._token = auth_token
@@ -140,6 +197,11 @@ class NanoleafClient(LEDClient):
        self._http: httpx.AsyncClient | None = None
        self._connected = False
        self._next_tx_at: float = 0.0
+        # Per-panel extControl streaming state.
+        self._per_panel = per_panel
+        self._streaming = False
+        self._panel_ids: List[int] = []
+        self._udp: socket.socket | None = None

    @property
    def host(self) -> str:
@@ -157,6 +219,9 @@ class NanoleafClient(LEDClient):

    @property
    def device_led_count(self) -> int | None:
+        # In per-panel streaming mode the panel count is authoritative.
+        if self._streaming and self._panel_ids:
+            return len(self._panel_ids)
        return self._led_count or None

    def _state_url(self) -> str:
@@ -169,9 +234,63 @@ class NanoleafClient(LEDClient):
            raise RuntimeError("NanoleafClient requires an auth_token; pair the device first")
        self._http = httpx.AsyncClient(timeout=self._request_timeout_s)
        self._connected = True
-        logger.info("NanoleafClient connected to %s:%d", self._host, NANOLEAF_PORT)
+        if self._per_panel:
+            # Best-effort: if streaming can't be enabled (old firmware, network)
+            # we silently fall back to the single-colour HTTP path.
+            try:
+                await self._setup_streaming()
+            except Exception as exc:  # noqa: BLE001 — degrade, never fail connect
+                logger.warning(
+                    "Nanoleaf %s: per-panel streaming unavailable, using single-colour (%s)",
+                    self._host,
+                    exc,
+                )
+                self._streaming = False
+        logger.info(
+            "NanoleafClient connected to %s:%d (per_panel=%s, streaming=%s)",
+            self._host,
+            NANOLEAF_PORT,
+            self._per_panel,
+            self._streaming,
+        )
        return True

+    async def _setup_streaming(self) -> None:
+        """Fetch the panel layout and enable extControl v2 UDP streaming."""
+        if self._http is None:
+            raise RuntimeError("NanoleafClient not connected")
+        base = f"http://{self._host}:{NANOLEAF_PORT}/api/v1/{self._token}"
+        # 1) Panel layout → ordered panel IDs.
+        resp = await self._http.get(f"{base}/panelLayout/layout")
+        resp.raise_for_status()
+        position_data = resp.json().get("positionData", [])
+        panel_ids = order_panels(position_data)
+        if not panel_ids:
+            raise RuntimeError("Nanoleaf reported no addressable panels")
+        # 2) Switch the controller into external (UDP) control, v2.
+        enable = await self._http.put(
+            f"{base}/effects",
+            json={
+                "write": {
+                    "command": "display",
+                    "animType": "extControl",
+                    "extControlVersion": "v2",
+                }
+            },
+        )
+        if enable.status_code not in (200, 204):
+            raise RuntimeError(
+                f"Nanoleaf refused extControl enable ({enable.status_code}): {enable.text[:200]}"
+            )
+        # 3) Fire-and-forget UDP socket for the per-frame stream.
+        self._udp = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+        self._udp.setblocking(False)
+        self._panel_ids = panel_ids
+        self._streaming = True
+        logger.info(
+            "Nanoleaf %s: extControl v2 streaming over %d panels", self._host, len(panel_ids)
+        )
+
    async def close(self) -> None:
        if self._http is not None:
            try:
@@ -179,6 +298,13 @@ class NanoleafClient(LEDClient):
            except (httpx.HTTPError, RuntimeError):
                pass
        self._http = None
+        if self._udp is not None:
+            try:
+                self._udp.close()
+            except OSError:
+                pass
+            self._udp = None
+        self._streaming = False
        self._connected = False

    async def _put_state(self, body: dict) -> None:
@@ -197,12 +323,28 @@ class NanoleafClient(LEDClient):
        pixels: List[Tuple[int, int, int]] | np.ndarray,
        brightness: int = 255,
    ) -> bool:
-        """Average the strip and PUT a single HSB state update."""
+        """Stream per-panel (extControl) when enabled, else average to one HSB state."""
        if not self.is_connected:
            raise RuntimeError("NanoleafClient not connected")
-        loop_now = asyncio.get_event_loop().time()
+        loop_now = asyncio.get_running_loop().time()
        if loop_now < self._next_tx_at:
            return True
+
+        if self._streaming and self._udp is not None and self._panel_ids:
+            scale = max(0, min(255, brightness)) / 255.0 if brightness < 255 else 1.0
+            arr = np.asarray(pixels, dtype=np.float32).reshape(-1, 3)
+            if scale != 1.0:
+                arr = arr * scale
+            scaled = np.clip(arr, 0, 255).astype(np.uint8)
+            panels = map_pixels_to_panels(scaled, self._panel_ids)
+            packet = build_extcontrol_v2_packet(panels)
+            try:
+                self._udp.sendto(packet, (self._host, NANOLEAF_STREAM_PORT))
+            except OSError as exc:
+                logger.debug("Nanoleaf %s: UDP stream send failed (%s)", self._host, exc)
+            self._next_tx_at = loop_now + self._min_interval_s
+            return True
+
        r, g, b = _average_color(pixels)
        if brightness < 255:
            scale = max(0, min(255, brightness)) / 255.0
@@ -61,6 +61,7 @@ class NanoleafDeviceProvider(LEDDeviceProvider):
            led_count=config.led_count,
            auth_token=config.nanoleaf_token,
            min_interval_s=max(0.0, config.nanoleaf_min_interval_ms / 1000.0),
+            per_panel=getattr(config, "nanoleaf_per_panel", False),
        )

    async def pair_device(self, url: str) -> dict:
@@ -40,3 +40,29 @@ def average_color(
        total_b += b
    n = len(pixels)
    return total_r // n, total_g // n, total_b // n
+
+
+def resample_to_n(
+    pixels: List[Tuple[int, int, int]] | np.ndarray,
+    n: int,
+) -> List[Tuple[int, int, int]]:
+    """Nearest-neighbour resample an N-pixel strip to exactly ``n`` pixels.
+
+    Output pixel ``i`` of ``n`` samples input pixel ``floor(i * N / n)``, so
+    the strip spreads spatially across a multi-element device (LIFX zones/
+    tiles, Nanoleaf panels, Hue segments). Black is returned for an empty
+    strip; ``n <= 0`` yields an empty list.
+    """
+    if n <= 0:
+        return []
+    arr = np.asarray(pixels, dtype=np.uint8).reshape(-1, 3)
+    n_pix = len(arr)
+    out: List[Tuple[int, int, int]] = []
+    for i in range(n):
+        if n_pix == 0:
+            out.append((0, 0, 0))
+            continue
+        idx = min(n_pix - 1, (i * n_pix) // n)
+        px = arr[idx]
+        out.append((int(px[0]), int(px[1]), int(px[2])))
+    return out
@@ -260,7 +260,12 @@ class CS2Adapter(GameAdapter):

        auth_section = payload.get("auth", {})
        actual_token = auth_section.get("token", "")
-        return bool(actual_token and actual_token == expected_token)
+        if not actual_token:
+            return False
+        # Constant-time comparison to avoid a timing oracle. Compare UTF-8 bytes
+        # so an attacker-controlled non-ASCII payload token returns False rather
+        # than raising TypeError out of secrets.compare_digest (→ 500).
+        return secrets.compare_digest(actual_token.encode("utf-8"), expected_token.encode("utf-8"))

    @classmethod
    def get_config_schema(cls) -> dict[str, Any]:
@@ -177,7 +177,12 @@ class Dota2Adapter(GameAdapter):

        auth_section = payload.get("auth", {})
        actual_token = auth_section.get("token", "")
-        return bool(actual_token and actual_token == expected_token)
+        if not actual_token:
+            return False
+        # Constant-time comparison to avoid a timing oracle. Compare UTF-8 bytes
+        # so an attacker-controlled non-ASCII payload token returns False rather
+        # than raising TypeError out of secrets.compare_digest (→ 500).
+        return secrets.compare_digest(actual_token.encode("utf-8"), expected_token.encode("utf-8"))

    @classmethod
    def get_config_schema(cls) -> dict[str, Any]:
@@ -4,6 +4,7 @@ Allows users to define custom JSON path mappings via the adapter_config
 rather than a YAML file. Delegates all parsing logic to MappingAdapter.
 """

+import secrets
 from typing import Any, ClassVar

 from ledgrab.core.game_integration.base_adapter import GameAdapter
@@ -54,11 +55,18 @@ class GenericWebhookAdapter(GameAdapter):
        payload: dict[str, Any],
        adapter_config: dict[str, Any],
    ) -> bool:
-        """Validate auth using a configurable header token."""
+        """Validate auth using a configurable header token.
+
+        Secure-by-default: this adapter is explicitly network-facing
+        (it accepts unauthenticated HTTP POSTs from anywhere on the LAN),
+        so a missing/empty ``auth_token`` REJECTS the request rather than
+        accepting it. A token must be configured for ingestion to work.
+        """
        expected_token = adapter_config.get("auth_token")
        if not expected_token:
-            # No auth configured
-            return True
+            # No token configured — reject (secure-by-default for a
+            # network-facing adapter; an open webhook is a LAN attack surface).
+            return False

        auth_header = adapter_config.get("auth_header", "Authorization")
        actual_value = headers.get(auth_header, "")
@@ -67,18 +75,32 @@ class GenericWebhookAdapter(GameAdapter):
        if actual_value.startswith("Bearer "):
            actual_value = actual_value[7:]

-        return bool(actual_value and actual_value == expected_token)
+        if not actual_value:
+            return False
+
+        # Constant-time comparison to avoid a token-length/timing oracle.
+        # Compare UTF-8 byte encodings: secrets.compare_digest raises TypeError
+        # on non-ASCII str, and the header value is attacker-controlled
+        # (Starlette latin-1-decodes header bytes to a possibly-non-ASCII str).
+        # Byte comparison is well-defined for any input and stays constant-time,
+        # so a non-ASCII token cleanly returns False instead of raising a 500.
+        return secrets.compare_digest(actual_value.encode("utf-8"), expected_token.encode("utf-8"))

    @classmethod
    def get_config_schema(cls) -> dict[str, Any]:
        """Return generic webhook config schema."""
        return {
            "type": "object",
+            "required": ["auth_token"],
            "properties": {
                "auth_token": {
                    "type": "string",
                    "title": "Auth Token",
-                    "description": "Optional token for authenticating incoming webhooks.",
+                    "description": (
+                        "Required token for authenticating incoming webhooks. "
+                        "Without it, ingestion is rejected (this adapter is "
+                        "network-facing and secure-by-default)."
+                    ),
                },
                "auth_header": {
                    "type": "string",
@@ -136,15 +158,17 @@ class GenericWebhookAdapter(GameAdapter):
            "HTTP POST requests with JSON payloads.\n\n"
            "**Steps:**\n"
            "1. Configure your event mappings above — map JSON paths to standard events\n"
-            "2. Set an auth token (optional but recommended)\n"
+            "2. Set an auth token (REQUIRED — without it incoming webhooks are rejected)\n"
            "3. Point your game/application to:\n"
            "   `POST http://<YOUR_IP>:8080/api/v1/game-integrations/<ID>/event`\n\n"
            "**Mapping example:**\n"
            "- Source path: `player.stats.health` → Event: `health` (min: 0, max: 100)\n"
            "- Source path: `events.kill_count` → Event: `kill` (trigger: on_increase)\n\n"
-            "**Auth:**\n"
+            "**Auth (required):**\n"
            "- Set `Authorization: Bearer <token>` header in your webhook sender\n"
            "- Or configure a custom auth header name in the adapter config\n"
+            "- A token is mandatory: this endpoint is reachable from the LAN, so\n"
+            "  an unconfigured token rejects all incoming events.\n"
        )


@@ -0,0 +1,90 @@
+"""Runtime manager that owns LoL Live-Client-Data polling threads.
+
+Unlike GSI adapters (CS2, Dota 2) that push events into the HTTP ingest
+endpoint, League of Legends only exposes a *local* poll API. The
+:class:`~ledgrab.core.game_integration.adapters.lol_adapter.LoLPoller` knows how
+to poll it; this manager owns the poller lifecycle, starting one daemon poller
+per enabled ``lol`` integration and reconciling on config changes.
+
+``sync()`` is the single reconciliation entry point — call it at startup and
+after any integration create/update/delete so runtime pollers always match the
+enabled ``lol`` integrations in the store.
+"""
+
+from __future__ import annotations
+
+import threading
+from typing import Any, Iterable
+
+from ledgrab.core.game_integration.adapters.lol_adapter import LoLAdapter, LoLPoller
+from ledgrab.core.game_integration.runtime_state import process_payload
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+
+
+class LoLPollManager:
+    """Owns one :class:`LoLPoller` per enabled League of Legends integration."""
+
+    def __init__(self, event_bus: Any) -> None:
+        self._event_bus = event_bus
+        self._lock = threading.Lock()
+        self._pollers: dict[str, LoLPoller] = {}
+        # Last adapter_config a poller was started with, so a config edit
+        # (e.g. poll interval) triggers a restart rather than going unnoticed.
+        self._configs: dict[str, dict[str, Any]] = {}
+
+    def sync(self, integrations: Iterable[Any]) -> None:
+        """Reconcile running pollers against the enabled ``lol`` integrations."""
+        desired = {
+            c.id: c
+            for c in integrations
+            if getattr(c, "enabled", False)
+            and getattr(c, "adapter_type", None) == LoLAdapter.ADAPTER_TYPE
+        }
+        with self._lock:
+            # Stop pollers whose integration is gone, disabled, or reconfigured.
+            for integration_id in list(self._pollers):
+                cfg = desired.get(integration_id)
+                if cfg is None or self._configs.get(integration_id) != dict(cfg.adapter_config):
+                    self._stop_locked(integration_id)
+            # Start pollers for anything enabled that isn't already running.
+            for integration_id, cfg in desired.items():
+                if integration_id not in self._pollers:
+                    self._start_locked(integration_id, cfg)
+
+    def stop_all(self) -> None:
+        """Stop every poller (shutdown hook)."""
+        with self._lock:
+            for integration_id in list(self._pollers):
+                self._stop_locked(integration_id)
+
+    @property
+    def active_count(self) -> int:
+        with self._lock:
+            return len(self._pollers)
+
+    # ── internals (call under self._lock) ──────────────────────────────────
+
+    def _start_locked(self, integration_id: str, cfg: Any) -> None:
+        adapter_config = dict(cfg.adapter_config)
+        poller = LoLPoller(adapter_config, self._make_callback(integration_id, adapter_config))
+        poller.start()
+        self._pollers[integration_id] = poller
+        self._configs[integration_id] = adapter_config
+        logger.info("Started LoL poller for integration %s", integration_id)
+
+    def _stop_locked(self, integration_id: str) -> None:
+        poller = self._pollers.pop(integration_id, None)
+        self._configs.pop(integration_id, None)
+        if poller is not None:
+            poller.stop()
+            logger.info("Stopped LoL poller for integration %s", integration_id)
+
+    def _make_callback(self, integration_id: str, adapter_config: dict[str, Any]):
+        event_bus = self._event_bus
+
+        def _on_poll(data: dict[str, Any]) -> None:
+            process_payload(integration_id, LoLAdapter, adapter_config, data, event_bus)
+
+        return _on_poll
@@ -10,6 +10,7 @@ The MappingAdapter class is a concrete GameAdapter whose behavior is
 entirely driven by the parsed YAML definition.
 """

+import secrets
 import time
 from pathlib import Path
 from typing import Any
@@ -241,7 +242,14 @@ class MappingAdapter(GameAdapter):
            expected_key = "auth_token"
            expected_value = adapter_config.get(expected_key, "")
            actual_value = headers.get(header_name, "")
-            return bool(expected_value and actual_value == expected_value)
+            if not (expected_value and actual_value):
+                return False
+            # Constant-time comparison to avoid a timing oracle. Compare UTF-8
+            # bytes so an attacker-controlled non-ASCII header value returns
+            # False rather than raising TypeError out of compare_digest (→ 500).
+            return secrets.compare_digest(
+                actual_value.encode("utf-8"), expected_value.encode("utf-8")
+            )

        logger.warning(f"Unknown auth type '{auth_type}' in mapping adapter '{self._name}'")
        return False
@@ -0,0 +1,120 @@
+"""Shared in-memory runtime state + payload processing for game integrations.
+
+Both the HTTP ingest route (``api/routes/game_integration.py``) and the
+poll-based LoL manager (``lol_poll_manager.py``) feed adapter payloads through
+the SAME per-integration prev-state / stats here, so a polled integration's
+status counters look identical to a pushed one's.
+
+Lives in ``core/`` (not the route) so the poll manager can reuse it without a
+route → core layering inversion.
+"""
+
+from __future__ import annotations
+
+import threading
+from typing import Any
+
+from ledgrab.core.game_integration.events import GameEvent
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+
+_state_lock = threading.Lock()
+
+# integration_id -> prev_state dict for diff-based trigger detection
+_prev_states: dict[str, dict[str, Any]] = {}
+
+# integration_id -> runtime stats
+_integration_stats: dict[str, dict[str, Any]] = {}
+
+
+def get_prev_state(integration_id: str) -> dict[str, Any]:
+    """Get or create the prev_state dict for an integration."""
+    with _state_lock:
+        if integration_id not in _prev_states:
+            _prev_states[integration_id] = {}
+        return _prev_states[integration_id]
+
+
+def set_prev_state(integration_id: str, state: dict[str, Any]) -> None:
+    """Update the prev_state dict for an integration."""
+    with _state_lock:
+        _prev_states[integration_id] = state
+
+
+def record_events(integration_id: str, events: list[GameEvent]) -> None:
+    """Record event stats for an integration."""
+    with _state_lock:
+        if integration_id not in _integration_stats:
+            _integration_stats[integration_id] = {
+                "event_count": 0,
+                "event_counts_by_type": {},
+                "last_event_time": None,
+            }
+        stats = _integration_stats[integration_id]
+        for event in events:
+            stats["event_count"] += 1
+            stats["event_counts_by_type"][event.event_type] = (
+                stats["event_counts_by_type"].get(event.event_type, 0) + 1
+            )
+            stats["last_event_time"] = event.timestamp
+
+
+def get_stats(integration_id: str) -> dict[str, Any]:
+    """Get a snapshot of runtime stats for an integration.
+
+    Returns a fresh copy (incl. a copied ``event_counts_by_type``) so the caller
+    can read/iterate it on the event loop while the poll thread keeps mutating
+    the live dict under the lock — otherwise a concurrent insert raises
+    ``RuntimeError: dictionary changed size during iteration``.
+    """
+    with _state_lock:
+        stats = _integration_stats.get(integration_id)
+        if stats is None:
+            return {"event_count": 0, "event_counts_by_type": {}, "last_event_time": None}
+        return {
+            "event_count": stats["event_count"],
+            "event_counts_by_type": dict(stats["event_counts_by_type"]),
+            "last_event_time": stats["last_event_time"],
+        }
+
+
+def cleanup_state(integration_id: str) -> None:
+    """Remove runtime state for a deleted integration."""
+    with _state_lock:
+        _prev_states.pop(integration_id, None)
+        _integration_stats.pop(integration_id, None)
+
+
+def process_payload(
+    integration_id: str,
+    adapter_cls: Any,
+    adapter_config: dict[str, Any],
+    data: dict[str, Any],
+    event_bus: Any,
+) -> list[GameEvent]:
+    """Parse a raw adapter payload, publish the events, and record stats.
+
+    Mirrors the body of the HTTP ingest route so a polled payload produces
+    identical events. Unlike the route (which returns HTTP 400 on a bad
+    payload), parse failures here are logged and swallowed — a poll loop must
+    keep running across a transient malformed frame. Returns the events
+    published (empty on parse failure).
+    """
+    prev_state = get_prev_state(integration_id)
+    try:
+        events, new_state = adapter_cls.parse_payload(data, adapter_config, prev_state)
+    except Exception:
+        logger.exception(
+            "Adapter %s failed to parse polled payload for %s",
+            getattr(adapter_cls, "ADAPTER_TYPE", "?"),
+            integration_id,
+        )
+        return []
+
+    set_prev_state(integration_id, new_state)
+    for event in events:
+        event_bus.publish(event)
+    if events:
+        record_events(integration_id, events)
+    return events
@@ -0,0 +1,141 @@
+"""Home Assistant MQTT auto-discovery publisher.
+
+When an :class:`~ledgrab.storage.mqtt_source.MQTTSource` has
+``publish_ha_discovery`` enabled, this publishes retained
+``<discovery_prefix>/<component>/<node_id>/<object_id>/config`` topics so an
+MQTT-only Home Assistant install gets LedGrab entities automatically — no YAML.
+
+Scope (read-only telemetry): a ``binary_sensor`` per automation (active /
+inactive) plus a connectivity ``binary_sensor`` for LedGrab itself, all tied to
+the broker's birth/will ``<base_topic>/status`` availability topic. This is
+deliberately one-way (LedGrab → HA): there is no inbound command surface, so it
+adds no new attack surface. Controllable ``light``/``switch`` entities (HA →
+LedGrab) are a documented follow-up.
+
+Cleanup: disabling discovery or deleting the source publishes an empty retained
+payload to every previously-published config topic, so HA drops the entities
+rather than leaving orphans behind forever.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import TYPE_CHECKING
+
+from ledgrab.utils import get_logger
+
+if TYPE_CHECKING:
+    from ledgrab.core.mqtt.mqtt_runtime import MQTTRuntime
+    from ledgrab.storage.automation_store import AutomationStore
+    from ledgrab.storage.mqtt_source import MQTTSource
+
+logger = get_logger(__name__)
+
+
+def _node_id(source_id: str) -> str:
+    """Namespace every entity by source so multiple brokers don't collide."""
+    return f"ledgrab_{source_id}"
+
+
+class HADiscoveryPublisher:
+    """Builds and publishes HA discovery configs for one MQTT runtime."""
+
+    def __init__(
+        self,
+        runtime: "MQTTRuntime",
+        source: "MQTTSource",
+        automation_store: "AutomationStore",
+        version: str = "",
+    ) -> None:
+        self._runtime = runtime
+        self._source = source
+        self._automation_store = automation_store
+        self._version = version
+        self._base = source.base_topic
+        self._prefix = source.discovery_prefix or "homeassistant"
+        # Config topics we have published, so remove_all() can clear exactly them.
+        self._published: set[str] = set()
+
+    # ── Config builders (pure — unit tested) ──────────────────────
+
+    def _device_block(self) -> dict:
+        return {
+            "identifiers": [_node_id(self._source.id)],
+            "name": "LedGrab",
+            "manufacturer": "LedGrab",
+            "model": "Ambient",
+            "sw_version": self._version or "unknown",
+        }
+
+    def _config_topic(self, component: str, object_id: str) -> str:
+        return f"{self._prefix}/{component}/{_node_id(self._source.id)}/{object_id}/config"
+
+    def build_connectivity_config(self) -> tuple[str, dict]:
+        """A connectivity ``binary_sensor`` reflecting LedGrab's birth/will."""
+        sid = self._source.id
+        topic = self._config_topic("binary_sensor", "connectivity")
+        payload = {
+            "unique_id": f"ledgrab_{sid}_connectivity",
+            "name": "LedGrab",
+            "device_class": "connectivity",
+            "state_topic": f"{self._base}/status",
+            "payload_on": "online",
+            "payload_off": "offline",
+            "device": self._device_block(),
+        }
+        return topic, payload
+
+    def build_automation_config(self, automation) -> tuple[str, dict]:
+        """A ``binary_sensor`` reflecting an automation's active state."""
+        sid = self._source.id
+        topic = self._config_topic("binary_sensor", f"automation_{automation.id}")
+        payload = {
+            "unique_id": f"ledgrab_{sid}_automation_{automation.id}",
+            "name": automation.name,
+            "state_topic": f"{self._base}/automation/{automation.id}/state",
+            "value_template": "{{ value_json.action }}",
+            "payload_on": "active",
+            "payload_off": "inactive",
+            "availability_topic": f"{self._base}/status",
+            "payload_available": "online",
+            "payload_not_available": "offline",
+            "device": self._device_block(),
+        }
+        return topic, payload
+
+    # ── Publish / remove ──────────────────────────────────────────
+
+    async def publish_all(self) -> None:
+        """Publish (retained) every discovery config + an initial state snapshot."""
+        topic, payload = self.build_connectivity_config()
+        await self._runtime.publish(topic, json.dumps(payload), retain=True)
+        self._published.add(topic)
+
+        count = 0
+        for automation in self._automation_store.get_all():
+            topic, payload = self.build_automation_config(automation)
+            await self._runtime.publish(topic, json.dumps(payload), retain=True)
+            self._published.add(topic)
+            # Seed an initial "inactive" state; the engine flips it live on the
+            # next activate/deactivate transition.
+            await self._runtime.publish_automation_state(automation.id, "inactive")
+            count += 1
+        logger.info(
+            "HA discovery published for source %s: %d automation sensor(s) + connectivity",
+            self._source.id,
+            count,
+        )
+
+    async def remove_all(self) -> None:
+        """Clear every previously-published config (empty retained payload)."""
+        for topic in list(self._published):
+            await self._runtime.publish(topic, "", retain=True)
+        # Also clear automations that may have been published in a prior run but
+        # since deleted — recompute the current set and clear those too.
+        topic, _ = self.build_connectivity_config()
+        await self._runtime.publish(topic, "", retain=True)
+        for automation in self._automation_store.get_all():
+            topic, _ = self.build_automation_config(automation)
+            await self._runtime.publish(topic, "", retain=True)
+        self._published.clear()
+        logger.info("HA discovery removed for source %s", self._source.id)
@@ -22,11 +22,27 @@ class MQTTManager:
    Multiple consumers share the same runtime via acquire/release.
    """

-    def __init__(self, store: MQTTSourceStore) -> None:
+    def __init__(
+        self,
+        store: MQTTSourceStore,
+        automation_store=None,
+        version: str = "",
+    ) -> None:
        self._store = store
+        # Optional deps for HA discovery publishing (injected from main.py).
+        self._automation_store = automation_store
+        self._version = version
        # source_id -> (runtime, ref_count)
        self._runtimes: Dict[str, tuple] = {}
+        # Sources for which we hold a discovery acquire() reference.
+        self._discovery_sources: set[str] = set()
        self._lock = asyncio.Lock()
+        # Serializes the discovery reconcile (ensure/disable) check-then-acquire so
+        # two concurrent sync_discovery() calls for the same source can't both see
+        # "first time" and double-acquire (ref-count leak). Separate from _lock
+        # because ensure/disable call acquire()/release() which take _lock, and
+        # asyncio.Lock is not re-entrant.
+        self._discovery_lock = asyncio.Lock()

    async def acquire(self, source_id: str) -> MQTTRuntime:
        """Get or create a runtime for the given MQTT source. Increments ref count."""
@@ -100,6 +116,88 @@ class MQTTManager:
        except Exception as e:
            logger.warning("Failed to update MQTT runtime %s: %s", source_id, e)

+    # ===== Home Assistant MQTT discovery =====
+
+    def _make_publisher(self, runtime: MQTTRuntime, source):
+        from ledgrab.core.mqtt.ha_discovery import HADiscoveryPublisher
+
+        return HADiscoveryPublisher(runtime, source, self._automation_store, version=self._version)
+
+    async def bootstrap_discovery(self) -> None:
+        """On startup, ensure a runtime + publish discovery for every enabled source."""
+        if self._automation_store is None:
+            return
+        for source in self._store.get_all():
+            if getattr(source, "publish_ha_discovery", False):
+                try:
+                    await self.ensure_discovery(source.id)
+                except Exception as exc:  # noqa: BLE001 — best-effort bootstrap
+                    logger.warning("HA discovery bootstrap failed for %s: %s", source.id, exc)
+
+    async def ensure_discovery(self, source_id: str) -> None:
+        """Hold a runtime open for *source_id* and publish its discovery configs.
+
+        Idempotent: a second call re-publishes (configs are retained) without
+        leaking a second acquire reference.
+        """
+        if self._automation_store is None:
+            return
+        try:
+            source = self._store.get(source_id)
+        except Exception:
+            return
+        if not getattr(source, "publish_ha_discovery", False):
+            return
+        async with self._discovery_lock:
+            first_time = source_id not in self._discovery_sources
+            if first_time:
+                runtime = await self.acquire(source_id)
+                self._discovery_sources.add(source_id)
+            else:
+                runtime = self.get_runtime(source_id)
+                if runtime is None:
+                    return
+        await self._make_publisher(runtime, source).publish_all()
+
+    async def disable_discovery(self, source_id: str) -> None:
+        """Clear a source's discovery configs and drop our runtime reference."""
+        async with self._discovery_lock:
+            if source_id not in self._discovery_sources:
+                return
+            self._discovery_sources.discard(source_id)
+            runtime = self.get_runtime(source_id)
+            if runtime is not None:
+                try:
+                    source = self._store.get(source_id)
+                    await self._make_publisher(runtime, source).remove_all()
+                except Exception as exc:  # noqa: BLE001 — best-effort cleanup
+                    logger.warning("HA discovery cleanup failed for %s: %s", source_id, exc)
+            await self.release(source_id)
+
+    async def sync_discovery(self, source_id: str) -> None:
+        """Reconcile discovery state after a source is created/updated."""
+        try:
+            source = self._store.get(source_id)
+        except Exception:
+            return
+        if getattr(source, "publish_ha_discovery", False):
+            await self.ensure_discovery(source_id)
+        else:
+            await self.disable_discovery(source_id)
+
+    async def publish_automation_state_all(self, automation_id: str, active: bool) -> None:
+        """Fan an automation's active state out to every discovery-enabled runtime."""
+        if not self._discovery_sources:
+            return
+        action = "active" if active else "inactive"
+        for source_id in list(self._discovery_sources):
+            runtime = self.get_runtime(source_id)
+            if runtime is not None:
+                try:
+                    await runtime.publish_automation_state(automation_id, action)
+                except Exception:  # noqa: BLE001 — never raise into the engine
+                    pass
+
    def get_connection_status(self) -> List[Dict[str, Any]]:
        """Get status of all active MQTT connections (for dashboard indicators)."""
        result = []
@@ -0,0 +1,115 @@
+"""AudioEnergyTap — a lightweight read-only tap on a shared audio capture.
+
+Lets a non-audio stream (e.g. a procedural effect) react to live loudness
+without owning audio capture. It resolves an ``audio_source_id`` to capture
+params exactly like ``AudioColorStripStream`` does, acquires the SHARED stream
+via the ``AudioCaptureManager`` (ref-counted, so it piggybacks on any audio
+visualiser already capturing the same device), and exposes a smoothed 0–1
+energy scalar.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+
+# RMS of typical program audio sits around 0.05–0.3; scale so a moderately
+# loud signal reaches ~1.0 before the per-effect intensity slider is applied.
+_RMS_GAIN = 4.0
+
+
+class AudioEnergyTap:
+    """Resolve → acquire → read smoothed loudness from a shared audio capture."""
+
+    def __init__(
+        self,
+        audio_capture_manager: Any,
+        audio_source_store: Any = None,
+        audio_template_store: Any = None,
+    ) -> None:
+        self._mgr = audio_capture_manager
+        self._audio_source_store = audio_source_store
+        self._audio_template_store = audio_template_store
+        self._stream = None
+        self._source_id = ""
+        self._device_index = -1
+        self._loopback = True
+        self._engine_type = None
+        self._engine_config = None
+        self._smoothed = 0.0
+
+    @property
+    def available(self) -> bool:
+        return self._mgr is not None
+
+    @property
+    def active(self) -> bool:
+        return self._stream is not None
+
+    def configure(self, audio_source_id: str) -> None:
+        """Resolve capture params for ``audio_source_id`` (no acquire yet)."""
+        self._source_id = audio_source_id or ""
+        self._device_index = -1
+        self._loopback = True
+        self._engine_type = None
+        self._engine_config = None
+        if not self._source_id or not self._audio_source_store:
+            return
+        try:
+            resolved = self._audio_source_store.resolve_audio_source(self._source_id)
+            self._device_index = resolved.device_index
+            self._loopback = resolved.is_loopback
+            if resolved.audio_template_id and self._audio_template_store:
+                try:
+                    tpl = self._audio_template_store.get_template(resolved.audio_template_id)
+                    self._engine_type = tpl.engine_type
+                    self._engine_config = tpl.engine_config
+                except ValueError as e:
+                    logger.warning(
+                        "AudioEnergyTap: template %s missing: %s", resolved.audio_template_id, e
+                    )
+        except ValueError as e:
+            logger.warning(
+                "AudioEnergyTap: failed to resolve audio source %s: %s", self._source_id, e
+            )
+
+    def start(self) -> None:
+        if self._mgr is None or self._stream is not None:
+            return
+        try:
+            self._stream = self._mgr.acquire(
+                self._device_index,
+                self._loopback,
+                engine_type=self._engine_type,
+                engine_config=self._engine_config,
+            )
+        except Exception as e:  # acquisition is best-effort — never break the effect
+            logger.warning("AudioEnergyTap: failed to acquire audio capture: %s", e)
+            self._stream = None
+
+    def stop(self) -> None:
+        if self._mgr is not None and self._stream is not None:
+            try:
+                self._mgr.release(self._device_index, self._loopback, engine_type=self._engine_type)
+            except Exception:  # release is best-effort
+                pass
+        self._stream = None
+        self._smoothed = 0.0
+
+    def energy(self, smoothing: float = 0.4) -> float:
+        """Return smoothed loudness in [0, 1] (0 when no capture/analysis yet).
+
+        ``smoothing`` is the inertia of the EMA: 0 = instantaneous, →1 = sluggish.
+        """
+        if self._stream is None:
+            return 0.0
+        analysis = self._stream.get_latest_analysis()
+        raw = 0.0
+        if analysis is not None:
+            raw = min(1.0, max(0.0, float(getattr(analysis, "rms", 0.0)) * _RMS_GAIN))
+        a = max(0.0, min(0.99, smoothing))
+        self._smoothed = self._smoothed * a + raw * (1.0 - a)
+        return self._smoothed
@@ -18,7 +18,7 @@ from ledgrab.core.capture.calibration import (
    CalibrationConfig,
    create_pixel_mapper,
 )
-from ledgrab.core.capture.screen_capture import extract_border_pixels
+from ledgrab.core.capture.screen_capture import crop_screen_capture, extract_border_pixels
 from ledgrab.storage.bindable import bfloat
 from ledgrab.utils import get_logger
 from ledgrab.utils.frame_limiter import FrameLimiter
@@ -296,7 +296,19 @@ class PictureColorStripStream(ColorStripStream):
                            t1 = time.perf_counter()
                            led_colors = mapper.map_lines_to_leds(frames_dict)
                        else:
-                            border_pixels = extract_border_pixels(frame, calibration.border_width)
+                            src = frame
+                            bw = calibration.border_width
+                            if calibration.has_roi:
+                                src = crop_screen_capture(
+                                    frame,
+                                    calibration.roi_x,
+                                    calibration.roi_y,
+                                    calibration.roi_width,
+                                    calibration.roi_height,
+                                )
+                                # Border width must stay within the cropped size.
+                                bw = max(1, min(bw, min(src.width, src.height) // 4))
+                            border_pixels = extract_border_pixels(src, bw)
                            t1 = time.perf_counter()
                            led_colors = mapper.map_border_to_leds(border_pixels)
                        t2 = time.perf_counter()
@@ -246,6 +246,13 @@ class ColorStripStreamManager:
            # Inject gradient store for palette resolution
            if self._gradient_store and hasattr(css_stream, "set_gradient_store"):
                css_stream.set_gradient_store(self._gradient_store)
+            # Inject audio capture deps for audio-reactive effects
+            if self._audio_capture_manager and hasattr(css_stream, "set_audio_capture"):
+                css_stream.set_audio_capture(
+                    self._audio_capture_manager,
+                    self._audio_source_store,
+                    self._audio_template_store,
+                )
            # Inject asset store for notification sound playback
            if self._asset_store and hasattr(css_stream, "set_asset_store"):
                css_stream.set_asset_store(self._asset_store)
@@ -69,6 +69,9 @@ class CompositeColorStripStream(ColorStripStream):

        # (src_len, dst_len) -> (src_x, dst_x, buffer) cache for zone resizing
        self._resize_cache: Dict[tuple, tuple] = {}
+        # (src_len, target_n) -> (src_x, dst_x) cache for full-strip resizing
+        # (output reuses the preallocated self._resize_buf from _ensure_pool)
+        self._resize_linspace_cache: Dict[tuple, tuple] = {}
        # layer_index -> (source_id, consumer_id, stream)
        self._sub_streams: Dict[int, tuple] = {}
        # layer_index -> (vs_id, value_stream)
@@ -314,8 +317,14 @@ class CompositeColorStripStream(ColorStripStream):
        n_src = len(colors)
        if n_src == target_n:
            return colors
-        src_x = np.linspace(0, 1, n_src)
-        dst_x = np.linspace(0, 1, target_n)
+        # Cache the (src_x, dst_x) linspace arrays keyed by (n_src, target_n)
+        # exactly like the zone path, so they are not reallocated every frame.
+        lkey = (n_src, target_n)
+        linspaces = self._resize_linspace_cache.get(lkey)
+        if linspaces is None:
+            linspaces = (np.linspace(0, 1, n_src), np.linspace(0, 1, target_n))
+            self._resize_linspace_cache[lkey] = linspaces
+        src_x, dst_x = linspaces
        buf = self._resize_buf
        for ch in range(3):
            np.copyto(
@@ -10,7 +10,6 @@ to the user's location and the current season.
 """

 import datetime
-import math
 import threading
 import time

@@ -84,71 +83,16 @@ _daylight_lut: np.ndarray | None = None


 # ── Solar position helpers ──────────────────────────────────────────────
-
-
-def _compute_solar_times(
-    latitude: float,
-    longitude: float,
-    day_of_year: int,
-    utc_offset_hours: float = 0.0,
-) -> tuple:
-    """Return (sunrise_hour, sunset_hour) in the user's wall-clock time.
-
-    Uses simplified NOAA solar equations:
-      - declination: decl = 23.45 * sin(2π * (284 + doy) / 365)
-      - hour angle:  cos(ha) = -tan(lat) * tan(decl)
-      - solar noon (UTC): 12 - longitude/15
-      - wall-clock sunrise/sunset: solar_noon_utc + utc_offset ∓ ha/15
-
-    Polar day and polar night are clamped to visible ranges.
-    """
-    deg2rad = math.pi / 180.0
-
-    decl_deg = 23.45 * math.sin(2.0 * math.pi * (284 + day_of_year) / 365.0)
-    decl_rad = decl_deg * deg2rad
-    lat_rad = latitude * deg2rad
-
-    cos_ha = -math.tan(lat_rad) * math.tan(decl_rad)
-    solar_noon_utc = 12.0 - longitude / 15.0
-    solar_noon_local = solar_noon_utc + utc_offset_hours
-
-    if cos_ha <= -1.0:
-        # Polar day — sun never sets; fake a long visible window
-        sunrise = solar_noon_local - 9.0
-        sunset = solar_noon_local + 9.0
-    elif cos_ha >= 1.0:
-        # Polar night — sun never rises; collapse to noon
-        sunrise = solar_noon_local
-        sunset = solar_noon_local
-    else:
-        ha_hours = math.acos(cos_ha) / (deg2rad * 15.0)
-        sunrise = solar_noon_local - ha_hours
-        sunset = solar_noon_local + ha_hours
-
-    # Clamp to a safe range the LUT builder can render. With reasonable
-    # tz/longitude pairs sunrise lands in (3..10) and sunset in (14..21);
-    # we widen the clamp so weird tz/lon combinations still produce a
-    # usable curve instead of dividing by zero.
-    sunrise = max(0.5, min(11.5, sunrise))
-    sunset = max(12.5, min(23.5, sunset))
-    return sunrise, sunset
-
-
-def _utc_offset_hours_for(tz_name: str, when: datetime.datetime | None = None) -> float:
-    """Return the UTC offset (in hours) for the given IANA timezone.
-
-    Empty/unknown tz falls back to the system local offset for ``when``.
-    """
-    when = when or datetime.datetime.now()
-    if tz_name and ZoneInfo is not None:
-        try:
-            offset = when.replace(tzinfo=None).astimezone(ZoneInfo(tz_name)).utcoffset()
-            if offset is not None:
-                return offset.total_seconds() / 3600.0
-        except ZoneInfoNotFoundError:
-            pass
-    local_offset = when.astimezone().utcoffset()
-    return local_offset.total_seconds() / 3600.0 if local_offset else 0.0
+#
+# ``compute_solar_times`` / ``utc_offset_hours_for`` moved to
+# ``ledgrab.utils.solar`` (pure math, no project imports) so the automation
+# engine can reuse them without importing this processing module. Imported
+# under their old private names here so the rest of this file (and
+# ``value_stream``, which imports them from here) is unchanged.
+from ledgrab.utils.solar import (  # noqa: E402
+    compute_solar_times as _compute_solar_times,
+    utc_offset_hours_for as _utc_offset_hours_for,
+)


 def _build_lut_for_solar_times(sunrise: float, sunset: float) -> np.ndarray:
@@ -11,6 +11,8 @@ from ledgrab.core.devices.led_client import (
    check_device_health,
    get_device_capabilities,
 )
+from ledgrab.core.activity_log.sanitize import sanitize_display
+from ledgrab.storage.activity_log import ActivityCategory, ActivitySeverity
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)
@@ -128,6 +130,41 @@ class DeviceHealthMixin:
                    "latency_ms": state.health.latency_ms,
                }
            )
+            # Audit record for device online/offline transition — best-effort.
+            # Wrapped so an instrumentation/import regression can never escape
+            # into _health_check_loop, whose top-level handler exits the loop
+            # permanently with no re-spawn (mirrors discovery_watcher._emit).
+            try:
+                from ledgrab.core.activity_log.recorder import get_module_recorder
+
+                rec = get_module_recorder()
+                if rec is not None:
+                    is_online = state.health.online
+                    # Best-effort name lookup from the device store.
+                    device_name: str | None = None
+                    try:
+                        if self._device_store is not None:
+                            device_name = self._device_store.get_device(device_id).name
+                    except Exception:
+                        pass
+                    safe_name = sanitize_display(device_name) if device_name else None
+                    display = safe_name or device_id
+                    action = "device.online" if is_online else "device.offline"
+                    severity = ActivitySeverity.INFO if is_online else ActivitySeverity.WARNING
+                    status_word = "came online" if is_online else "went offline"
+                    rec.record(
+                        category=ActivityCategory.DEVICE,
+                        action=action,
+                        severity=severity,
+                        actor="system",
+                        entity_type="device",
+                        entity_id=device_id,
+                        entity_name=safe_name,
+                        message=f"Device '{display}' {status_word}",
+                        metadata={"latency_ms": state.health.latency_ms},
+                    )
+            except Exception as e:
+                logger.debug("Device health: audit record failed for %s: %s", device_id, e)

        # Auto-sync LED count
        reported = state.health.device_led_count
@@ -137,16 +137,34 @@ class DeviceTestModeMixin:
        await self._send_pixels_to_device(device_id, pixels)

    async def _send_clear_pixels(self, device_id: str) -> None:
-        """Send all-black pixels to clear LED output."""
+        """Send all-black pixels to clear LED output.
+
+        This is the explicit teardown path — unlike the per-frame
+        ``_send_pixels_to_device`` swallow, a clear that must actually take
+        effect retries once before giving up, so a single transient send
+        error doesn't leave the device lit after a session ends.
+        """
        ds = self._devices[device_id]
        pixels = [(0, 0, 0)] * ds.led_count
-        await self._send_pixels_to_device(device_id, pixels)
+        try:
+            client = await self._get_idle_client(device_id)
+            await client.send_pixels(pixels)
+        except Exception as e:
+            logger.warning(f"Clear send to {device_id} failed, retrying once: {e}")
+            client = await self._get_idle_client(device_id)
+            await client.send_pixels(pixels)

    async def _send_pixels_to_device(self, device_id: str, pixels) -> None:
        """Send pixels to a device via cached idle client.

        Reuses a cached connection to avoid repeated serial reconnections
        (which trigger Arduino bootloader reset on Adalight devices).
+
+        Send failures are logged and swallowed (best-effort per-frame send).
+        Callers that need a *guaranteed* clear — e.g. session teardown that
+        must "never leave the device dark" — must NOT rely on this returning
+        cleanly; use ``_send_clear_pixels`` (which retries once) and treat a
+        propagated exception as a failed clear.
        """
        try:
            client = await self._get_idle_client(device_id)
@@ -304,6 +304,13 @@ class EffectColorStripStream(ColorStripStream):
        # Sparkle rain state
        self._sparkle_state: np.ndarray | None = None  # per-LED brightness 0..1
        self._gradient_store = None  # injected by stream manager
+        # Audio-reactive modulation (tap injected by stream manager via
+        # set_audio_capture; defaults keep the effect working with no audio).
+        self._audio_tap = None
+        self._audio_reactive = False
+        self._reactive_mode = "brightness"
+        self._reactive_audio_source_id = ""
+        self._reactive_intensity = 0.7
        self._update_from_source(source)

    def set_gradient_store(self, gradient_store) -> None:
@@ -344,9 +351,36 @@ class EffectColorStripStream(ColorStripStream):
        self._intensity = bfloat(getattr(source, "intensity", 1.0), 1.0)
        self._scale = bfloat(getattr(source, "scale", 1.0), 1.0)
        self._mirror = bool(getattr(source, "mirror", False))
+
+        # Audio-reactive params
+        self._audio_reactive = bool(getattr(source, "audio_reactive", False))
+        self._reactive_mode = getattr(source, "reactive_mode", "brightness") or "brightness"
+        self._reactive_intensity = bfloat(getattr(source, "reactive_intensity", 0.7), 0.7)
+        self._reactive_audio_source_id = getattr(source, "reactive_audio_source_id", "") or ""
+        if self._audio_tap is not None:
+            self._audio_tap.configure(self._reactive_audio_source_id)
+
        with self._colors_lock:
            self._colors: np.ndarray | None = None

+    def set_audio_capture(
+        self, audio_capture_manager, audio_source_store=None, audio_template_store=None
+    ) -> None:
+        """Inject audio capture deps so the effect can react to loudness.
+
+        Called by the stream manager after construction (mirrors
+        ``set_gradient_store``). Builds the tap and resolves the referenced
+        audio source; a missing manager just leaves the effect non-reactive.
+        """
+        if audio_capture_manager is None:
+            return
+        from ledgrab.core.processing.audio_energy_tap import AudioEnergyTap
+
+        self._audio_tap = AudioEnergyTap(
+            audio_capture_manager, audio_source_store, audio_template_store
+        )
+        self._audio_tap.configure(self._reactive_audio_source_id)
+
    def configure(self, device_led_count: int) -> None:
        if self._auto_size and device_led_count > 0:
            new_count = max(self._led_count, device_led_count)
@@ -369,6 +403,8 @@ class EffectColorStripStream(ColorStripStream):
    def start(self) -> None:
        if self._running:
            return
+        if self._audio_reactive and self._audio_tap is not None:
+            self._audio_tap.start()
        self._running = True
        self._thread = threading.Thread(
            target=self._animate_loop,
@@ -387,6 +423,8 @@ class EffectColorStripStream(ColorStripStream):
            if self._thread.is_alive():
                logger.warning("EffectColorStripStream animate thread did not terminate within 5s")
            self._thread = None
+        if self._audio_tap is not None:
+            self._audio_tap.stop()
        self._heat = None
        self._heat_n = 0
        logger.info("EffectColorStripStream stopped")
@@ -399,12 +437,39 @@ class EffectColorStripStream(ColorStripStream):
        from ledgrab.storage.color_strip_source import EffectColorStripSource

        if isinstance(source, EffectColorStripSource):
+            # Release the audio capture before reconfiguring so a changed
+            # audio source / loopback is re-acquired cleanly.
+            tap_was_active = self._audio_tap is not None and self._audio_tap.active
+            if tap_was_active:
+                self._audio_tap.stop()
            prev_led_count = self._led_count if self._auto_size else None
            self._update_from_source(source)
            if prev_led_count and self._auto_size:
                self._led_count = prev_led_count
+            if self._running and self._audio_reactive and self._audio_tap is not None:
+                self._audio_tap.start()
            logger.info("EffectColorStripStream params updated in-place")

+    def _apply_audio_modulation(self, buf: np.ndarray) -> None:
+        """Scale brightness and/or saturation of the rendered frame by loudness.
+
+        Quiet audio dims/desaturates toward ``1 - intensity``; loud audio drives
+        full brightness (and a saturation boost up to ``1 + intensity``).
+        """
+        k = max(0.0, min(1.0, self.resolve("reactive_intensity", self._reactive_intensity)))
+        if k <= 0.0:
+            return
+        e = self._audio_tap.energy()
+        f = buf.astype(np.float32)
+        if self._reactive_mode in ("brightness", "both"):
+            f *= (1.0 - k) + k * e
+        if self._reactive_mode in ("saturation", "both"):
+            sat = (1.0 - k) + 2.0 * k * e
+            lum = (f[:, 0] * 0.299 + f[:, 1] * 0.587 + f[:, 2] * 0.114)[:, None]
+            f = lum + (f - lum) * sat
+        np.clip(f, 0.0, 255.0, out=f)
+        buf[:] = f.astype(np.uint8)
+
    def set_clock(self, clock) -> None:
        """Set or clear the sync clock runtime. Thread-safe (read atomically by loop)."""
        self._clock = clock
@@ -476,6 +541,9 @@ class EffectColorStripStream(ColorStripStream):
                            continue
                        render_fn(self, buf, n, anim_time)

+                        if self._audio_reactive and self._audio_tap is not None:
+                            self._apply_audio_modulation(buf)
+
                        with self._colors_lock:
                            self._colors = buf
                    except Exception as e:
@@ -973,13 +1041,18 @@ class EffectColorStripStream(ColorStripStream):
        # Use noise at very low frequency for blob movement
        np.multiply(self._s_arange, scale * 0.03, out=self._s_f32_a)

-        # Two blob layers at different speeds for organic movement
+        # Two blob layers at different speeds for organic movement.
+        # fbm() returns a shared internal buffer that the next fbm() call
+        # overwrites, so each layer must be copied out — write into the
+        # preallocated scratch buffers instead of allocating per frame.
        self._s_f32_a += t * speed * 0.1
-        layer1 = self._noise.fbm(self._s_f32_a, octaves=3).copy()
+        layer1 = self._s_layer1
+        np.copyto(layer1, self._noise.fbm(self._s_f32_a, octaves=3))

        np.multiply(self._s_arange, scale * 0.05, out=self._s_f32_a)
        self._s_f32_a += t * speed * 0.07 + 100.0
-        layer2 = self._noise.fbm(self._s_f32_a, octaves=2).copy()
+        layer2 = self._s_layer2
+        np.copyto(layer2, self._noise.fbm(self._s_f32_a, octaves=2))

        # Combine: create blob-like shapes with soft edges
        combined = self._s_f32_a
@@ -44,6 +44,7 @@ from ledgrab.core.processing.sync_clock_manager import SyncClockManager
 from ledgrab.core.weather.weather_manager import WeatherManager
 from ledgrab.core.processing.device_health import DeviceHealthMixin
 from ledgrab.core.processing.device_test_mode import DeviceTestModeMixin
+from ledgrab.core.capture.calibration_session import CalibrationChaseMixin
 from ledgrab.utils import get_logger

 logger = get_logger(__name__)
@@ -106,7 +107,9 @@ class DeviceState:
    zone_mode: str = "combined"


-class ProcessorManager(AutoRestartMixin, DeviceHealthMixin, DeviceTestModeMixin):
+class ProcessorManager(
+    AutoRestartMixin, DeviceHealthMixin, DeviceTestModeMixin, CalibrationChaseMixin
+):
    """Manages devices and delegates target processing to TargetProcessor instances.

    Devices are registered for health monitoring.
@@ -739,10 +739,19 @@ class WledTargetProcessor(TargetProcessor):

        self._last_preview_data = data

-        async def _send_safe(ws):
+        # Bound each per-client send to roughly one frame interval so a slow or
+        # backpressured preview WebSocket can never throttle the device send
+        # cadence. Clients that time out or error are dropped from the set.
+        eff_fps = self._effective_fps if self._effective_fps > 0 else 30
+        send_timeout = 1.0 / eff_fps
+
+        async def _send_safe(ws) -> bool:
            try:
-                await ws.send_bytes(data)
+                await asyncio.wait_for(ws.send_bytes(data), timeout=send_timeout)
                return True
+            except asyncio.TimeoutError:
+                logger.debug("LED preview broadcast WS send timed out (slow client dropped)")
+                return False
            except Exception as e:
                logger.debug("LED preview broadcast WS send failed: %s", e)
                return False
@@ -0,0 +1,323 @@
+"""Playlist engine — background loop that auto-cycles a scene playlist.
+
+A playlist is an ordered, timed sequence of scene presets. The engine drives
+**at most one** playlist at a time: starting a new playlist transparently stops
+any currently-running one. Each cycle re-reads the playlist from the store, so
+edits (and deletion) take effect at the next cycle boundary without a restart.
+
+The actual state application reuses ``scene_activator.apply_scene_state`` — the
+same code path the scene-presets API and the automation engine use — so a
+playlist step behaves exactly like manually activating that preset.
+"""
+
+import asyncio
+import random
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import List
+
+from ledgrab.storage.scene_playlist import ScenePlaylist, clamp_duration
+from ledgrab.utils import get_logger
+
+logger = get_logger(__name__)
+
+
+@dataclass
+class PlaylistRuntimeState:
+    """Volatile runtime state of the (single) active playlist. Not persisted."""
+
+    playlist_id: str
+    playlist_name: str
+    current_index: int
+    item_count: int
+    current_preset_id: str | None
+    started_at: datetime
+    step_started_at: datetime
+    step_duration: float
+
+    def to_dict(self) -> dict:
+        return {
+            "is_running": True,
+            "playlist_id": self.playlist_id,
+            "playlist_name": self.playlist_name,
+            "current_index": self.current_index,
+            "item_count": self.item_count,
+            "current_preset_id": self.current_preset_id,
+            "started_at": self.started_at.isoformat(),
+            "step_started_at": self.step_started_at.isoformat(),
+            "step_duration": self.step_duration,
+        }
+
+
+_IDLE_STATE = {
+    "is_running": False,
+    "playlist_id": None,
+    "playlist_name": None,
+    "current_index": 0,
+    "item_count": 0,
+    "current_preset_id": None,
+    "started_at": None,
+    "step_started_at": None,
+    "step_duration": 0.0,
+}
+
+
+class PlaylistError(Exception):
+    """Raised when a playlist cannot be started (empty / not found)."""
+
+
+class PlaylistEngine:
+    """Cycles a scene playlist's presets on a timer, one playlist at a time."""
+
+    def __init__(
+        self,
+        playlist_store,
+        scene_preset_store,
+        target_store,
+        processor_manager,
+    ):
+        self._playlist_store = playlist_store
+        self._scene_preset_store = scene_preset_store
+        self._target_store = target_store
+        self._manager = processor_manager
+
+        self._task: asyncio.Task | None = None
+        self._state: PlaylistRuntimeState | None = None
+        # Serialises start/stop so overlapping API calls can't leave two
+        # cycling tasks alive at once.
+        self._lifecycle_lock = asyncio.Lock()
+
+    # ===== Public control API =====
+
+    async def start_playlist(self, playlist_id: str) -> PlaylistRuntimeState:
+        """Start cycling ``playlist_id``, stopping any current playlist first.
+
+        Raises ``PlaylistError`` if the playlist is unknown or has no items.
+        """
+        try:
+            playlist = self._playlist_store.get_playlist(playlist_id)
+        except Exception as exc:  # EntityNotFoundError / ValueError
+            raise PlaylistError(f"Playlist not found: {playlist_id}") from exc
+
+        if not playlist.items:
+            raise PlaylistError(f"Playlist '{playlist.name}' has no items")
+
+        async with self._lifecycle_lock:
+            await self._cancel_task()
+
+            now = datetime.now(timezone.utc)
+            first_item = playlist.items[0]
+            self._state = PlaylistRuntimeState(
+                playlist_id=playlist.id,
+                playlist_name=playlist.name,
+                current_index=0,
+                item_count=len(playlist.items),
+                current_preset_id=first_item.scene_preset_id,
+                started_at=now,
+                step_started_at=now,
+                step_duration=clamp_duration(first_item.duration_seconds),
+            )
+            self._task = asyncio.create_task(self._run(playlist.id))
+
+        self._fire_event("started")
+        logger.info("Playlist '%s' started (%d items)", playlist.name, len(playlist.items))
+        return self._state
+
+    async def stop(self) -> None:
+        """Stop the active playlist (if any). Leaves the last scene applied."""
+        async with self._lifecycle_lock:
+            was_running = self._task is not None
+            await self._cancel_task()
+            stopped_id = self._state.playlist_id if self._state else None
+            self._state = None
+        if was_running:
+            self._fire_event("stopped", playlist_id=stopped_id)
+            logger.info("Playlist stopped")
+
+    async def stop_if_running(self, playlist_id: str) -> None:
+        """Stop the playlist only if ``playlist_id`` is the one running.
+
+        Used when a playlist is deleted or edited so a stale snapshot can't keep
+        cycling. The read-compare and the stop happen atomically under the
+        lifecycle lock so a concurrent natural-end / start can't slip a
+        different playlist in between the check and the stop.
+        """
+        async with self._lifecycle_lock:
+            state = self._state
+            if state is None or state.playlist_id != playlist_id:
+                return
+            was_running = self._task is not None
+            await self._cancel_task()
+            stopped_id = self._state.playlist_id if self._state else playlist_id
+            self._state = None
+        if was_running:
+            self._fire_event("stopped", playlist_id=stopped_id)
+            logger.info("Playlist stopped")
+
+    # ===== Query API (used by routes) =====
+
+    def is_running(self) -> bool:
+        # Snapshot the task ref so a concurrent clear (set to None at an await
+        # boundary) can't turn the deref into an attribute error mid-read.
+        task = self._task
+        return task is not None and not task.done()
+
+    def get_running_playlist_id(self) -> str | None:
+        state = self._state
+        return state.playlist_id if state else None
+
+    def get_state(self) -> dict:
+        # Snapshot both refs once: the event loop won't preempt this sync method
+        # between the reads, but snapshotting also guards against ever returning
+        # a half-cleared state (running True while _state is already None).
+        state = self._state
+        task = self._task
+        if state is not None and task is not None and not task.done():
+            return state.to_dict()
+        return dict(_IDLE_STATE)
+
+    # ===== Internal =====
+
+    async def _cancel_task(self) -> None:
+        """Cancel and await the cycling task. Caller holds the lifecycle lock."""
+        task = self._task
+        self._task = None
+        if task is None:
+            return
+        task.cancel()
+        try:
+            await task
+        except asyncio.CancelledError:
+            pass
+        except Exception as exc:  # pragma: no cover - defensive
+            logger.error("Playlist task raised on cancel: %s", exc, exc_info=True)
+
+    async def _run(self, playlist_id: str) -> None:
+        """Cycle the playlist until cancelled, the playlist ends, or it errors."""
+        try:
+            while True:
+                # Re-read each cycle so edits/deletes apply at the boundary.
+                try:
+                    playlist = self._playlist_store.get_playlist(playlist_id)
+                except Exception:
+                    logger.info("Playlist %s removed while running; stopping", playlist_id)
+                    break
+
+                if not playlist.items:
+                    logger.info("Playlist '%s' has no items; stopping", playlist.name)
+                    break
+
+                applied_any = await self._run_cycle(playlist)
+
+                if not playlist.loop:
+                    break
+                if not applied_any:
+                    # Every item referenced a missing preset — a looping
+                    # playlist would otherwise spin with no dwell. Bail out.
+                    logger.warning(
+                        "Playlist '%s' applied no valid presets this cycle; stopping",
+                        playlist.name,
+                    )
+                    break
+
+            # Natural end (non-loop or guard). Clear state without recursing
+            # through stop() (which would try to cancel this very task). Take
+            # the lifecycle lock so the clear + 'stopped' event are atomic with
+            # respect to a concurrent start/stop/stop_if_running — otherwise the
+            # two could interleave and emit duplicate or contradictory terminal
+            # events. _run never calls _cancel_task and never otherwise holds
+            # this lock, so acquiring it here cannot deadlock; if a canceller
+            # holding the lock cancels us while we wait to acquire it, the
+            # acquire raises CancelledError and we fall through to the handler.
+            ended_id = None
+            should_fire = False
+            async with self._lifecycle_lock:
+                # Re-check under the lock: a concurrent start_playlist may have
+                # replaced us while we waited. Only clear if we're still current.
+                if self._task is asyncio.current_task():
+                    self._task = None
+                    ended_id = self._state.playlist_id if self._state else None
+                    self._state = None
+                    should_fire = True
+            if should_fire:
+                self._fire_event("stopped", playlist_id=ended_id)
+                logger.info("Playlist '%s' finished", playlist_id)
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:  # pragma: no cover - defensive
+            logger.error("Playlist run loop error: %s", exc, exc_info=True)
+
+    async def _run_cycle(self, playlist: ScenePlaylist) -> bool:
+        """Run one pass over the playlist's items. Returns True if any applied."""
+        order = self._resolve_order(playlist)
+        applied_any = False
+
+        for orig_index, item in order:
+            duration = clamp_duration(item.duration_seconds)
+
+            applied = await self._apply_item(item.scene_preset_id)
+            if applied:
+                # Only advertise runtime state for a step we actually applied
+                # and are about to dwell on — a missing/skipped preset must not
+                # briefly show up in get_state(). ``current_index`` is the
+                # ORIGINAL persisted items[] index (not the shuffled position)
+                # so clients can correlate it back to the items array.
+                if self._state is not None:
+                    self._state.current_index = orig_index
+                    self._state.current_preset_id = item.scene_preset_id
+                    self._state.step_started_at = datetime.now(timezone.utc)
+                    self._state.step_duration = duration
+
+                applied_any = True
+                self._fire_event("advanced", index=orig_index, preset_id=item.scene_preset_id)
+                # Only dwell on scenes we actually applied; skip missing ones
+                # immediately so the cycle doesn't stall on a dead reference.
+                await asyncio.sleep(duration)
+
+        return applied_any
+
+    def _resolve_order(self, playlist: ScenePlaylist) -> List:
+        """Return ``(orig_index, item)`` pairs, optionally shuffled.
+
+        The original persisted index travels with each item so ``current_index``
+        in the runtime state always maps back into ``playlist.items`` even when
+        shuffle reorders the cycle.
+        """
+        indexed = list(enumerate(playlist.items))
+        if playlist.shuffle and len(indexed) > 1:
+            random.shuffle(indexed)  # noqa: S311 - cosmetic ordering, not security
+        return indexed
+
+    async def _apply_item(self, preset_id: str) -> bool:
+        """Apply one scene preset. Returns False if it could not be applied."""
+        if not self._scene_preset_store or not self._target_store or not self._manager:
+            logger.warning("Playlist engine missing stores; cannot apply %s", preset_id)
+            return False
+        try:
+            preset = self._scene_preset_store.get_preset(preset_id)
+        except Exception:
+            logger.warning("Playlist references missing scene preset %s (skipped)", preset_id)
+            return False
+
+        from ledgrab.core.scenes.scene_activator import apply_scene_state
+
+        _status, errors = await apply_scene_state(preset, self._target_store, self._manager)
+        if errors:
+            logger.warning("Playlist step '%s' applied with errors: %s", preset.name, errors)
+        return True
+
+    def _fire_event(self, action: str, **extra) -> None:
+        if self._manager is None:
+            return
+        try:
+            self._manager.fire_event(
+                {
+                    "type": "playlist_state_changed",
+                    "action": action,
+                    "playlist_id": extra.get("playlist_id")
+                    or (self._state.playlist_id if self._state else None),
+                    **{k: v for k, v in extra.items() if k != "playlist_id"},
+                }
+            )
+        except Exception as exc:
+            logger.error("Playlist event fire failed: %s", exc, exc_info=True)
@@ -35,6 +35,7 @@ import ledgrab.core.audio  # noqa: F401 — trigger engine auto-registration
 from ledgrab.storage.value_source_store import ValueSourceStore
 from ledgrab.storage.automation_store import AutomationStore
 from ledgrab.storage.scene_preset_store import ScenePresetStore
+from ledgrab.storage.scene_playlist_store import ScenePlaylistStore
 from ledgrab.storage.sync_clock_store import SyncClockStore
 from ledgrab.storage.color_strip_processing_template_store import (
    ColorStripProcessingTemplateStore,
@@ -47,8 +48,10 @@ from ledgrab.core.weather.weather_manager import WeatherManager
 from ledgrab.storage.home_assistant_store import HomeAssistantStore
 from ledgrab.core.home_assistant.ha_manager import HomeAssistantManager
 from ledgrab.core.automations.automation_engine import AutomationEngine
+from ledgrab.core.scenes.playlist_engine import PlaylistEngine
 from ledgrab.storage.game_integration_store import GameIntegrationStore
 from ledgrab.core.game_integration.event_bus import GameEventBus
+from ledgrab.core.game_integration.lol_poll_manager import LoLPollManager
 import ledgrab.core.game_integration.adapters  # noqa: F401 — register built-in adapters
 from ledgrab.core.game_integration.community_loader import register_community_adapters
 from ledgrab.core.mqtt.mqtt_manager import MQTTManager
@@ -58,6 +61,9 @@ from ledgrab.storage.audio_processing_template_store import AudioProcessingTempl
 from ledgrab.storage.pattern_template_store import PatternTemplateStore
 import ledgrab.core.audio.filters  # noqa: F401 — trigger audio filter auto-registration
 from ledgrab.core.backup.auto_backup import AutoBackupEngine
+from ledgrab.core.activity_log.recorder import ActivityRecorder, set_module_recorder
+from ledgrab.core.activity_log.retention import ActivityLogRetentionEngine
+from ledgrab.storage.activity_log_repository import ActivityLogRepository
 from ledgrab.core.processing.os_notification_listener import OsNotificationListener
 from ledgrab.core.devices.discovery_watcher import DiscoveryWatcher
 from ledgrab.core.update.update_service import UpdateService
@@ -157,6 +163,7 @@ audio_template_store = AudioTemplateStore(db)
 value_source_store = ValueSourceStore(db)
 automation_store = AutomationStore(db)
 scene_preset_store = ScenePresetStore(db)
+scene_playlist_store = ScenePlaylistStore(db)
 sync_clock_store = SyncClockStore(db)
 cspt_store = ColorStripProcessingTemplateStore(db)
 gradient_store = GradientStore(db)
@@ -173,14 +180,22 @@ weather_manager = WeatherManager(weather_source_store)
 ha_store = HomeAssistantStore(db)
 ha_manager = HomeAssistantManager(ha_store)
 mqtt_source_store = MQTTSourceStore(db)
-mqtt_manager = MQTTManager(mqtt_source_store)
+mqtt_manager = MQTTManager(
+    mqtt_source_store, automation_store=automation_store, version=__version__
+)
 http_endpoint_store = HTTPEndpointStore(db)
 audio_processing_template_store = AudioProcessingTemplateStore(db)
 game_integration_store = GameIntegrationStore(db)
 pattern_template_store = PatternTemplateStore(db)
 game_event_bus = GameEventBus()
+# Owns LoL Live-Client-Data poll threads (League is poll-only, not GSI-push).
+lol_poll_manager = LoLPollManager(game_event_bus)
 register_community_adapters()

+# Activity log repository — constructed at module level like other stores so
+# it migrates the DB schema (``002_add_activity_log``) on import.
+activity_log_repo = ActivityLogRepository(db)
+
 processor_manager = ProcessorManager(
    ProcessorDependencies(
        picture_source_store=picture_source_store,
@@ -254,6 +269,15 @@ async def lifespan(app: FastAPI):
        client_labels = ", ".join(config.auth.api_keys.keys())
        logger.info(f"Authorized clients: {client_labels}")

+    # Warn when the OpenAPI docs surface is exposed without a token.
+    if config.auth.expose_docs:
+        logger.warning(
+            "auth.expose_docs is ON: /docs, /redoc and /openapi.json load "
+            "without an API key (loopback and LAN). The API surface (routes + "
+            "schemas) is readable by anyone who can reach the server; endpoint "
+            "calls still require a token."
+        )
+
    # One-shot migration: legacy global ``mqtt:`` config block → first MQTTSource.
    # No-op once the store has any entries.
    try:
@@ -278,6 +302,26 @@ async def lifespan(app: FastAPI):
        value_source_store=value_source_store,
    )

+    # Create playlist engine — auto-cycles scene presets, one playlist at a
+    # time. Idle (no background task) until a playlist is started via the API.
+    playlist_engine = PlaylistEngine(
+        playlist_store=scene_playlist_store,
+        scene_preset_store=scene_preset_store,
+        target_store=output_target_store,
+        processor_manager=processor_manager,
+    )
+
+    # Create activity recorder + retention engine.  The recorder needs the
+    # processor_manager to fire live events, so it is built after that is
+    # already constructed at module level.
+    activity_recorder = ActivityRecorder(activity_log_repo, processor_manager)
+    activity_recorder.ensure_loop()
+    activity_log_retention_engine = ActivityLogRetentionEngine(
+        repo=activity_log_repo,
+        db=db,
+        recorder=activity_recorder,
+    )
+
    # Create auto-backup engine — derive paths from database location so that
    # demo mode auto-backups go to data/demo/ instead of data/.
    _data_dir = Path(config.storage.database_file).parent
@@ -314,7 +358,9 @@ async def lifespan(app: FastAPI):
        value_source_store=value_source_store,
        automation_store=automation_store,
        scene_preset_store=scene_preset_store,
+        scene_playlist_store=scene_playlist_store,
        automation_engine=automation_engine,
+        playlist_engine=playlist_engine,
        auto_backup_engine=auto_backup_engine,
        sync_clock_store=sync_clock_store,
        sync_clock_manager=sync_clock_manager,
@@ -328,12 +374,19 @@ async def lifespan(app: FastAPI):
        ha_manager=ha_manager,
        game_integration_store=game_integration_store,
        game_event_bus=game_event_bus,
+        lol_poll_manager=lol_poll_manager,
        mqtt_store=mqtt_source_store,
        mqtt_manager=mqtt_manager,
        http_endpoint_store=http_endpoint_store,
        audio_processing_template_store=audio_processing_template_store,
        pattern_template_store=pattern_template_store,
+        activity_recorder=activity_recorder,
+        activity_log_repo=activity_log_repo,
+        activity_log_retention_engine=activity_log_retention_engine,
    )
+    # Expose the recorder via the module singleton so non-DI sites
+    # (fire_entity_event, device threads) can call record() without FastAPI DI.
+    set_module_recorder(activity_recorder)

    # Register devices in processor manager for health monitoring
    devices = device_store.get_all_devices()
@@ -373,12 +426,25 @@ async def lifespan(app: FastAPI):
    # Start automation engine (evaluates conditions and activates scenes)
    await automation_engine.start()

+    # Publish Home Assistant MQTT discovery for any source that opted in.
+    await mqtt_manager.bootstrap_discovery()
+
    # Start auto-backup engine (periodic configuration backups)
    await auto_backup_engine.start()

+    # Start activity log retention engine (hourly prune of old entries)
+    await activity_log_retention_engine.start()
+
    # Start update checker (periodic release polling)
    await update_service.start()

+    # Start LoL Live-Client-Data pollers for any enabled League integrations
+    # (League is poll-only; GSI games push into the ingest endpoint instead).
+    try:
+        lol_poll_manager.sync(game_integration_store.get_all_integrations())
+    except Exception as e:
+        logger.error(f"Failed to start LoL pollers: {e}")
+
    # Start OS notification listener (Windows toast → notification CSS streams)
    os_notif_listener = OsNotificationListener(
        color_strip_store=color_strip_store,
@@ -424,6 +490,19 @@ async def lifespan(app: FastAPI):
        except Exception as e:
            logger.error("Shutdown step '%s' raised: %s", label, e)

+    # Record the shutdown event FIRST — before any engine teardown — so there
+    # is always a final log entry on graceful shutdown.
+    try:
+        activity_recorder.record(
+            category="system",
+            action="server.shutting_down",
+            severity="info",
+            actor="system",
+            message="Server is shutting down",
+        )
+    except Exception as e:
+        logger.error("Failed to record shutdown event: %s", e)
+
    # Legacy hook — SQLite stores are write-through so this only logs.
    # Durability comes from PRAGMA synchronous=FULL + the explicit
    # wal_checkpoint(TRUNCATE) in Database.close() at the end of this block.
@@ -436,6 +515,22 @@ async def lifespan(app: FastAPI):
    # would talk to processors mid-shutdown.
    await _bounded("automation_engine.stop", automation_engine.stop(), timeout=1.5)

+    # Stop the playlist engine so its cycling task can't apply scenes mid-shutdown.
+    await _bounded("playlist_engine.stop", playlist_engine.stop(), timeout=1.0)
+
+    # Stop LoL poller threads so they stop publishing game events.
+    try:
+        lol_poll_manager.stop_all()
+    except Exception as e:
+        logger.error(f"Error stopping LoL pollers: {e}")
+
+    # Tear down any active calibration session BEFORE stop_all so the device
+    # isn't left stuck in the white-chase and its prior target is restored.
+    # stop() is a no-op when no session is active.
+    from ledgrab.core.capture.calibration_session import get_calibration_session
+
+    await _bounded("calibration_session.stop", get_calibration_session().stop(), timeout=1.0)
+
    # Stop discovery watcher and OS notification listener so they stop
    # firing events into a shutting-down processor manager.
    if discovery_watcher is not None:
@@ -486,6 +581,7 @@ async def lifespan(app: FastAPI):

    await _bounded("update_service.stop", update_service.stop(), timeout=0.5)
    await _bounded("auto_backup_engine.stop", auto_backup_engine.stop(), timeout=0.5)
+    await _bounded("activity_log_retention.stop", activity_log_retention_engine.stop(), timeout=0.5)

    # Close the DB last so it runs a TRUNCATE checkpoint, flushing the WAL
    # into the main file. Without this, writes can survive a graceful app
@@ -606,25 +702,27 @@ async def _access_log(request: Request, call_next):


 # ── Auth-gated OpenAPI surface ────────────────────────────────────────────
-# Re-add the docs endpoints we disabled above, now protected by the same
-# Bearer auth as the rest of the API. When auth is unconfigured, loopback
-# clients still get in anonymously (per ``verify_api_key`` policy).
+# Re-add the docs endpoints we disabled above, protected by the same Bearer
+# auth as the rest of the API. The ``DocsAccess`` dependency relaxes this to
+# anonymous access (loopback + LAN) when ``auth.expose_docs`` is set, so the
+# docs can be viewed in a browser without a token. When auth is unconfigured,
+# loopback clients still get in anonymously (per ``verify_api_key`` policy).
 from fastapi.openapi.docs import get_redoc_html, get_swagger_ui_html  # noqa: E402
-from ledgrab.api.auth import AuthRequired  # noqa: E402
+from ledgrab.api.auth import DocsAccess  # noqa: E402


@app.get("/openapi.json", include_in_schema=False)
-async def _openapi(_auth: AuthRequired):
+async def _openapi(_auth: DocsAccess):
    return JSONResponse(app.openapi())


@app.get("/docs", include_in_schema=False)
-async def _swagger_docs(_auth: AuthRequired):
+async def _swagger_docs(_auth: DocsAccess):
    return get_swagger_ui_html(openapi_url="/openapi.json", title=f"{app.title} — API docs")


@app.get("/redoc", include_in_schema=False)
-async def _redoc_docs(_auth: AuthRequired):
+async def _redoc_docs(_auth: DocsAccess):
    return get_redoc_html(openapi_url="/openapi.json", title=f"{app.title} — API docs")


@@ -0,0 +1,770 @@
+/* ─────────────────────────────────────────────────────────────────────────
+   Activity Log — audit viewer tab
+   Design language: precision-instrument / ledger. Monospaced timestamps,
+   color-coded severity rail, thin category pills. Clean "terminal" feel
+   without being cold — the primary green accent anchors the live-update dot.
+   ───────────────────────────────────────────────────────────────────────── */
+
+/* ── Panel wrapper ───────────────────────────────────────────────────────── */
+
+.al-panel {
+    display: flex;
+    flex-direction: column;
+    gap: 0;
+    max-width: 1400px;
+    margin: 0 auto;
+    padding: var(--space-lg) var(--space-lg) var(--space-xl);
+    min-height: 0;
+}
+
+/* ── Filter toolbar ──────────────────────────────────────────────────────── */
+
+.al-toolbar {
+    display: flex;
+    flex-direction: column;
+    gap: var(--space-sm);
+    padding: var(--space-md) var(--space-md);
+    /* Match the elevated card surface used by entity cards (.dashboard-target),
+       not the near-black --bg-secondary, so the panel reads as one of the app's
+       cards rather than a separate flat sheet. */
+    background: var(--lux-bg-1, var(--card-bg));
+    border: var(--lux-hairline) solid var(--lux-line, var(--border-color));
+    border-radius: var(--radius-md);
+    margin-bottom: var(--space-md);
+}
+
+.al-toolbar-row {
+    display: flex;
+    align-items: center;
+    flex-wrap: wrap;
+    gap: var(--space-sm);
+}
+
+.al-toolbar-search {
+    gap: var(--space-sm);
+}
+
+/* Search input */
+.al-search-wrap {
+    position: relative;
+    flex: 1;
+    min-width: 200px;
+    max-width: 420px;
+}
+
+.al-search-icon {
+    position: absolute;
+    left: 10px;
+    top: 50%;
+    transform: translateY(-50%);
+    pointer-events: none;
+    color: var(--text-muted);
+    display: flex;
+    align-items: center;
+}
+
+.al-search-icon .icon {
+    width: 15px;
+    height: 15px;
+}
+
+.al-search-input {
+    width: 100%;
+    padding: 6px 10px 6px 34px;
+    background: var(--card-bg);
+    border: var(--lux-hairline) solid var(--border-color);
+    border-radius: var(--radius-sm);
+    color: var(--text-color);
+    font-size: 0.875rem;
+    font-family: var(--font-body);
+    transition: border-color var(--duration-fast) var(--ease-out),
+                box-shadow var(--duration-fast) var(--ease-out);
+    outline: none;
+}
+
+.al-search-input:focus {
+    border-color: var(--primary-color);
+    box-shadow: 0 0 0 2px rgba(76, 175, 80, 0.15);
+}
+
+.al-search-input::placeholder { color: var(--text-muted); }
+
+/* Quick presets */
+.al-presets {
+    display: flex;
+    gap: 4px;
+    flex-wrap: wrap;
+}
+
+.al-preset-btn {
+    padding: 4px 10px;
+    font-size: 0.75rem;
+    font-weight: 500;
+    background: var(--card-bg);
+    border: var(--lux-hairline) solid var(--border-color);
+    border-radius: var(--radius-pill);
+    color: var(--text-secondary);
+    cursor: pointer;
+    transition: background var(--duration-fast), color var(--duration-fast), border-color var(--duration-fast);
+    white-space: nowrap;
+}
+
+.al-preset-btn:hover {
+    background: var(--primary-color);
+    border-color: var(--primary-color);
+    color: var(--primary-contrast);
+}
+
+/* Clear button */
+.al-clear-btn {
+    padding: 4px 6px;
+    margin-left: auto;
+    color: var(--text-secondary);
+}
+
+.al-clear-btn:hover { color: var(--danger-color); border-color: var(--danger-color); }
+
+/* Export button + dropdown */
+.al-export-wrap {
+    position: relative;
+    margin-left: auto;
+}
+
+.al-export-btn {
+    display: flex;
+    align-items: center;
+    gap: 5px;
+    padding: 5px 12px;
+    font-size: 0.8125rem;
+}
+
+.al-export-btn .icon { width: 14px; height: 14px; }
+
+/* Caret signals this button opens a menu (rather than firing a direct action),
+   and rotates to point up while the menu is open. */
+.al-export-caret {
+    display: inline-flex;
+    margin-left: 1px;
+    transition: transform var(--duration-fast) var(--ease-out);
+}
+.al-export-caret .icon { width: 12px; height: 12px; }
+.al-export-wrap.open .al-export-caret { transform: rotate(180deg); }
+
+.al-export-menu {
+    display: none;
+    position: absolute;
+    right: 0;
+    top: calc(100% + 4px);
+    background: var(--card-bg);
+    border: var(--lux-hairline) solid var(--border-color);
+    border-radius: var(--radius-sm);
+    box-shadow: 0 4px 12px var(--shadow-color);
+    z-index: 100;
+    min-width: 140px;
+    overflow: hidden;
+}
+
+.al-export-wrap.open .al-export-menu { display: block; }
+
+.al-export-menu button {
+    display: block;
+    width: 100%;
+    padding: 8px 14px;
+    text-align: left;
+    background: none;
+    border: none;
+    color: var(--text-color);
+    font-size: 0.8125rem;
+    cursor: pointer;
+}
+
+.al-export-menu button:hover,
+.al-export-menu button:focus-visible { background: var(--bg-secondary); outline: none; }
+
+/* Filter label */
+.al-filter-label {
+    font-size: 0.75rem;
+    font-weight: 600;
+    color: var(--text-muted);
+    text-transform: uppercase;
+    letter-spacing: 0.05em;
+    white-space: nowrap;
+    flex-shrink: 0;
+}
+
+.al-filter-label-sep { margin-left: var(--space-sm); }
+
+/* Category / severity chips */
+.al-chip-group {
+    display: flex;
+    gap: 4px;
+    flex-wrap: wrap;
+}
+
+.al-chip {
+    display: inline-flex;
+    align-items: center;
+    gap: 4px;
+    padding: 3px 9px;
+    font-size: 0.75rem;
+    font-weight: 500;
+    border: var(--lux-hairline) solid var(--border-color);
+    border-radius: var(--radius-pill);
+    background: var(--card-bg);
+    color: var(--text-secondary);
+    cursor: pointer;
+    transition: background var(--duration-fast), color var(--duration-fast), border-color var(--duration-fast);
+    white-space: nowrap;
+}
+
+.al-chip .icon { width: 12px; height: 12px; }
+
+.al-chip:hover {
+    background: var(--bg-secondary);
+    border-color: var(--text-secondary);
+    color: var(--text-color);
+}
+
+.al-chip.active {
+    background: var(--primary-color);
+    border-color: var(--primary-color);
+    color: var(--primary-contrast);
+}
+
+/* Severity chip colors when active */
+.al-sev-chip-error.active  { background: var(--danger-color);  border-color: var(--danger-color); }
+.al-sev-chip-warning.active { background: var(--warning-color); border-color: var(--warning-color); }
+.al-sev-chip-info.active   { background: var(--info-color);    border-color: var(--info-color); }
+
+/* Advanced field row */
+.al-toolbar-advanced {
+    gap: var(--space-sm);
+    padding-top: var(--space-xs);
+    border-top: var(--lux-hairline) solid var(--border-color);
+}
+
+.al-field-group {
+    display: flex;
+    flex-direction: column;
+    gap: 2px;
+    min-width: 160px;
+    flex: 1;
+}
+
+.al-field-label {
+    font-size: 0.6875rem;
+    font-weight: 600;
+    color: var(--text-muted);
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+}
+
+.al-field-input {
+    padding: 4px 8px;
+    background: var(--card-bg);
+    border: var(--lux-hairline) solid var(--border-color);
+    border-radius: var(--radius-sm);
+    color: var(--text-color);
+    font-size: 0.8125rem;
+    font-family: var(--font-body);
+    outline: none;
+    transition: border-color var(--duration-fast);
+}
+
+.al-field-input:focus {
+    border-color: var(--primary-color);
+    box-shadow: 0 0 0 2px rgba(76, 175, 80, 0.12);
+}
+
+/* ── List header (count + live dot) ─────────────────────────────────────── */
+
+.al-list-header {
+    display: flex;
+    align-items: center;
+    justify-content: space-between;
+    padding: var(--space-xs) 0;
+    margin-bottom: var(--space-xs);
+}
+
+.al-count {
+    font-size: 0.8125rem;
+    color: var(--text-muted);
+    font-variant-numeric: tabular-nums;
+}
+
+.al-live-indicator {
+    display: flex;
+    align-items: center;
+    gap: 5px;
+    font-size: 0.75rem;
+    font-weight: 500;
+    color: var(--primary-text-color);
+    letter-spacing: 0.02em;
+}
+
+.al-live-dot {
+    width: 7px;
+    height: 7px;
+    background: var(--primary-color);
+    border-radius: 50%;
+    animation: al-pulse 2s infinite;
+}
+
+@keyframes al-pulse {
+    0%, 100% { opacity: 1; transform: scale(1); }
+    50%       { opacity: 0.55; transform: scale(0.85); }
+}
+
+/* ── Entry rows ─────────────────────────────────────────────────────────── */
+
+.al-list {
+    display: flex;
+    flex-direction: column;
+    gap: 1px;
+}
+
+.al-entry {
+    /* Same elevated surface + hairline as entity cards (see .al-toolbar). */
+    background: var(--lux-bg-1, var(--card-bg));
+    border: var(--lux-hairline) solid var(--lux-line, var(--border-color));
+    border-radius: var(--radius-sm);
+    overflow: hidden;
+    transition: border-color var(--duration-fast);
+}
+
+.al-entry:hover { border-color: var(--text-muted); }
+
+/* New-entry flash — settles on the card surface (animation-fill-mode: forwards
+   holds the 100% frame, so it must match the .al-entry background exactly). */
+@keyframes al-new-flash {
+    0%   { background: color-mix(in srgb, var(--primary-color) 18%, var(--lux-bg-1, var(--card-bg))); }
+    100% { background: var(--lux-bg-1, var(--card-bg)); }
+}
+
+.al-entry-new.al-entry-appear { animation: al-new-flash 1.8s var(--ease-out) forwards; }
+
+.al-entry-row {
+    display: grid;
+    /* icon | time | badge | actor | message | entity | chevron
+       badge is fixed so all category names (AUTH…CAPTURE) occupy identical
+       width; actor is capped so long usernames don't push the message over;
+       message takes all remaining space. */
+    grid-template-columns: 24px 80px 78px minmax(0, 110px) 1fr auto 20px;
+    align-items: center;
+    gap: var(--space-sm);
+    padding: var(--space-xs) var(--space-md);
+    cursor: pointer;
+    min-height: 36px;
+    outline: none;
+}
+
+.al-entry-row:focus-visible { box-shadow: inset 0 0 0 2px var(--primary-color); }
+
+/* Severity rail icon */
+.al-sev { display: flex; align-items: center; justify-content: center; }
+.al-sev .icon { width: 14px; height: 14px; }
+.al-sev-info    .icon { color: var(--info-color); }
+.al-sev-warning .icon { color: var(--warning-color); }
+.al-sev-error   .icon { color: var(--danger-color); }
+
+/* Time */
+.al-time {
+    font-size: 0.75rem;
+    font-variant-numeric: tabular-nums;
+    font-family: var(--font-mono);
+    color: var(--text-muted);
+    white-space: nowrap;
+}
+
+/* Category badge */
+.al-cat-badge {
+    display: inline-block;
+    padding: 1px 7px;
+    font-size: 0.6875rem;
+    font-weight: 600;
+    letter-spacing: 0.03em;
+    text-transform: uppercase;
+    border-radius: var(--radius-pill);
+    white-space: nowrap;
+    border: var(--lux-hairline) solid transparent;
+}
+
+/* Per-category colors — subtle tinted backgrounds */
+.al-cat-auth    { background: rgba(33, 150, 243, 0.12); color: var(--info-color); border-color: rgba(33, 150, 243, 0.25); }
+.al-cat-device  { background: rgba(156, 39, 176, 0.10); color: #ab47bc;           border-color: rgba(156, 39, 176, 0.22); }
+.al-cat-entity  { background: rgba(76, 175, 80, 0.12);  color: var(--primary-text-color); border-color: rgba(76, 175, 80, 0.25); }
+.al-cat-capture { background: rgba(255, 152, 0, 0.12);  color: var(--warning-color);  border-color: rgba(255, 152, 0, 0.25); }
+.al-cat-system  { background: rgba(120, 120, 120, 0.12); color: var(--text-secondary); border-color: rgba(120, 120, 120, 0.25); }
+
+[data-theme="light"] .al-cat-auth    { background: rgba(33, 150, 243, 0.08);  }
+/* Darker purple text in light theme — the dark-theme #ab47bc fails AA contrast
+   on the pale tinted background at this small badge size. */
+[data-theme="light"] .al-cat-device  { background: rgba(156, 39, 176, 0.08); color: #8e24aa; }
+[data-theme="light"] .al-cat-entity  { background: rgba(76, 175, 80, 0.08);   }
+[data-theme="light"] .al-cat-capture { background: rgba(255, 152, 0, 0.08);   }
+[data-theme="light"] .al-cat-system  { background: rgba(120, 120, 120, 0.08); }
+
+/* Actor — constrained by its grid column (minmax(0, 110px)) */
+.al-actor {
+    font-size: 0.8125rem;
+    font-family: var(--font-mono);
+    color: var(--text-secondary);
+    white-space: nowrap;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    min-width: 0;
+}
+
+/* Message — min-width:0 lets the 1fr column actually truncate */
+.al-msg {
+    font-size: 0.8125rem;
+    color: var(--text-color);
+    overflow: hidden;
+    text-overflow: ellipsis;
+    white-space: nowrap;
+    min-width: 0;
+    text-align: left;
+}
+
+/* Entity crosslink */
+.al-entity { display: flex; align-items: center; }
+
+.al-entity-link {
+    font-size: 0.75rem;
+    color: var(--primary-text-color);
+    background: none;
+    border: none;
+    padding: 0;
+    cursor: pointer;
+    text-decoration: underline dotted;
+    text-underline-offset: 2px;
+    white-space: nowrap;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    max-width: 120px;
+    display: block;
+}
+
+.al-entity-link:hover { color: var(--primary-hover); text-decoration-style: solid; }
+
+.al-entity-name {
+    font-size: 0.75rem;
+    color: var(--text-muted);
+    white-space: nowrap;
+    overflow: hidden;
+    text-overflow: ellipsis;
+}
+
+/* Expand chevron */
+.al-expand-chevron {
+    font-size: 0.6875rem;
+    color: var(--text-muted);
+    justify-self: end;
+    user-select: none;
+}
+
+/* ── Entry detail drawer ─────────────────────────────────────────────────── */
+
+.al-detail {
+    padding: var(--space-sm) var(--space-md) var(--space-md);
+    border-top: var(--lux-hairline) solid var(--border-color);
+    background: var(--bg-secondary);
+    animation: al-detail-open var(--duration-fast) var(--ease-out);
+}
+
+@keyframes al-detail-open {
+    from { opacity: 0; transform: translateY(-4px); }
+    to   { opacity: 1; transform: translateY(0); }
+}
+
+.al-detail-grid {
+    display: grid;
+    grid-template-columns: max-content 1fr;
+    gap: 4px var(--space-md);
+    font-size: 0.8125rem;
+}
+
+.al-detail-grid dt {
+    color: var(--text-muted);
+    font-weight: 600;
+    font-size: 0.75rem;
+    text-transform: uppercase;
+    letter-spacing: 0.04em;
+    align-self: start;
+    padding-top: 2px;
+    white-space: nowrap;
+}
+
+.al-detail-grid dd {
+    color: var(--text-color);
+    word-break: break-all;
+}
+
+.al-detail-grid code {
+    font-family: var(--font-mono);
+    font-size: 0.8125rem;
+    background: var(--bg-color);
+    padding: 1px 5px;
+    border-radius: var(--radius-sm);
+    border: var(--lux-hairline) solid var(--border-color);
+}
+
+.al-meta-pre {
+    font-family: var(--font-mono);
+    font-size: 0.75rem;
+    background: var(--bg-color);
+    border: var(--lux-hairline) solid var(--border-color);
+    border-radius: var(--radius-sm);
+    padding: var(--space-sm);
+    overflow-x: auto;
+    white-space: pre;
+    color: var(--text-secondary);
+    max-height: 220px;
+    overflow-y: auto;
+    margin: 0;
+}
+
+/* ── Load More ───────────────────────────────────────────────────────────── */
+
+.al-load-more {
+    display: block;
+    width: 100%;
+    margin-top: var(--space-md);
+    text-align: center;
+    font-size: 0.875rem;
+}
+
+/* ── Empty / loading / error states ─────────────────────────────────────── */
+
+.al-state {
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+    justify-content: center;
+    gap: var(--space-sm);
+    padding: var(--space-xl);
+    color: var(--text-muted);
+    font-size: 0.875rem;
+    text-align: center;
+}
+
+.al-state-icon .icon {
+    width: 36px;
+    height: 36px;
+    opacity: 0.35;
+}
+
+.al-loading { flex-direction: row; padding: var(--space-lg); }
+
+.al-spinner {
+    width: 20px;
+    height: 20px;
+    border: 2px solid var(--border-color);
+    border-top-color: var(--primary-color);
+    border-radius: 50%;
+    animation: al-spin 0.6s linear infinite;
+    flex-shrink: 0;
+}
+
+@keyframes al-spin { to { transform: rotate(360deg); } }
+
+.al-error .al-state-icon .icon { color: var(--danger-color); opacity: 0.6; }
+
+/* ── List container ──────────────────────────────────────────────────────── */
+
+.al-list-container {
+    flex: 1;
+    min-height: 0;
+}
+
+/* Subtle busy state while a slow re-query is in flight: the current rows stay
+   visible (no spinner flash) but dim slightly and stop accepting clicks until
+   the fresh results swap in. Only applied after a short delay, so instant
+   filtering shows nothing. */
+.al-list-container.al-busy {
+    opacity: 0.55;
+    pointer-events: none;
+    transition: opacity var(--duration-fast) var(--ease-out);
+}
+
+/* ── Tabular-nums utility ────────────────────────────────────────────────── */
+
+.tabular-nums { font-variant-numeric: tabular-nums; }
+
+/* ── Responsive ──────────────────────────────────────────────────────────── */
+
+@media (max-width: 900px) {
+    .al-entry-row {
+        /* icon | time | badge (fixed) | message | chevron — actor+entity hidden */
+        grid-template-columns: 20px 70px 78px 1fr 18px;
+    }
+    /* Hide actor and entity link at small widths */
+    .al-actor,
+    .al-entity { display: none; }
+}
+
+@media (max-width: 600px) {
+    .al-panel { padding: var(--space-sm); }
+
+    .al-entry-row {
+        grid-template-columns: 20px 1fr auto 18px;
+        grid-template-rows: auto auto;
+        gap: 4px var(--space-xs);
+    }
+
+    /* Row 1: [sev] [message] [badge] [chevron]; Row 2: [time] under the message.
+       message stays in its own 1fr column so it never overlaps the badge. */
+    .al-time     { grid-column: 2; grid-row: 2; font-size: 0.6875rem; }
+    .al-cat-badge{ grid-column: 3; grid-row: 1; }
+    .al-msg      { grid-column: 2; grid-row: 1; }
+
+    .al-toolbar-advanced .al-field-group { min-width: 100%; }
+}
+
+/* ─────────────────────────────────────────────────────────────────────────
+   Dashboard "Recent Activity" widget  (.dal-*)
+   Compact, consistent with the precision-instrument language of the full tab.
+   Rows are tighter than the full viewer — just sev icon + relative time + msg.
+   ───────────────────────────────────────────────────────────────────────── */
+
+/* List container */
+.dal-list {
+    display: flex;
+    flex-direction: column;
+    gap: 0;
+}
+
+/* Compact entry row */
+.al-compact-row {
+    display: grid;
+    grid-template-columns: 18px 52px 1fr;
+    align-items: center;
+    gap: 0 8px;
+    padding: 5px 4px;
+    border-bottom: 1px solid var(--border-color);
+    min-height: 28px;
+    transition: background 0.12s;
+}
+
+.al-compact-row:last-child {
+    border-bottom: none;
+}
+
+.al-compact-row:hover {
+    background: var(--bg-secondary);
+}
+
+.al-compact-icon .icon {
+    width: 14px;
+    height: 14px;
+    display: block;
+    flex-shrink: 0;
+}
+
+.al-compact-time {
+    font-family: var(--font-mono, monospace);
+    font-size: 0.6875rem;
+    color: var(--text-muted);
+    white-space: nowrap;
+    font-variant-numeric: tabular-nums;
+}
+
+.al-compact-msg {
+    font-size: 0.8125rem;
+    color: var(--text-color);
+    white-space: nowrap;
+    overflow: hidden;
+    text-overflow: ellipsis;
+}
+
+/* Severity color on the row itself (row inherits .al-sev-* from renderCompactEntry) */
+.al-compact-row.al-sev-error   .al-compact-icon .icon { color: var(--danger-color); }
+.al-compact-row.al-sev-warning .al-compact-icon .icon { color: var(--warning-color); }
+.al-compact-row.al-sev-info    .al-compact-icon .icon { color: var(--info-color); }
+
+/* Empty state inside widget */
+.dal-empty {
+    padding: 16px 8px;
+}
+.dal-empty p {
+    font-size: 0.8125rem;
+    color: var(--text-muted);
+}
+
+/* Loading state placeholder */
+.dal-loading {
+    padding: 16px 8px;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+}
+
+/* Footer — "View all →" link */
+.dal-footer {
+    padding: 6px 4px 2px;
+    display: flex;
+    justify-content: flex-end;
+}
+
+.dal-view-all {
+    font-size: 0.8125rem;
+    color: var(--primary-text-color);
+    text-decoration: none;
+    border: none;
+    background: none;
+    cursor: pointer;
+    padding: 2px 4px;
+    transition: opacity 0.15s;
+}
+
+.dal-view-all:hover {
+    opacity: 0.75;
+    text-decoration: underline;
+}
+
+/* ─────────────────────────────────────────────────────────────────────────
+   Settings panel helpers  (ds-info-note, ds-inline-link)
+   These are general enough to live here but scoped tightly enough to not
+   bleed into the rest of the settings layout.
+   ───────────────────────────────────────────────────────────────────────── */
+
+.ds-info-note {
+    display: flex;
+    align-items: flex-start;
+    gap: 8px;
+    padding: 10px 14px;
+    margin-bottom: 16px;
+    background: color-mix(in srgb, var(--info-color) 8%, var(--bg-secondary));
+    border: 1px solid color-mix(in srgb, var(--info-color) 25%, var(--border-color));
+    border-radius: var(--radius-sm, 4px);
+    font-size: 0.8125rem;
+    color: var(--text-color);
+    line-height: 1.5;
+}
+
+.ds-info-note .icon {
+    width: 15px;
+    height: 15px;
+    flex-shrink: 0;
+    margin-top: 1px;
+    color: var(--info-color);
+}
+
+/* Inline text button that looks like a link (used in ds-info-note, hints) */
+.ds-inline-link {
+    background: none;
+    border: none;
+    padding: 0;
+    margin: 0;
+    cursor: pointer;
+    color: var(--primary-text-color);
+    font-size: inherit;
+    font-family: inherit;
+    text-decoration: underline;
+    text-underline-offset: 2px;
+    display: inline;
+}
+
+.ds-inline-link:hover {
+    opacity: 0.8;
+}
@@ -19,5 +19,6 @@
@import './graph-editor.css';
@import './appearance.css';
@import './game-integration.css';
+@import './activity-log.css';
@import './mobile.css';
@import './tv.css';
--- a/Show More
+++ b/Show More