diff --git a/server/src/ledgrab/api/routes/setup.py b/server/src/ledgrab/api/routes/setup.py
index 0e3a947..2cd6983 100644
--- a/server/src/ledgrab/api/routes/setup.py
+++ b/server/src/ledgrab/api/routes/setup.py
@@ -36,11 +36,13 @@ from ledgrab.api.dependencies import (
     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
@@ -116,11 +118,16 @@ def _rollback(
     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,
@@ -129,6 +136,18 @@ def _rollback(
         "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)
@@ -164,6 +183,7 @@ async def scaffold_setup(
     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.
 
@@ -192,6 +212,7 @@ async def scaffold_setup(
         picture_source_store=picture_source_store,
         css_store=css_store,
         output_target_store=output_target_store,
+        manager=manager,
     )
 
     try:
@@ -269,6 +290,16 @@ async def scaffold_setup(
         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
diff --git a/server/src/ledgrab/api/schemas/calibration.py b/server/src/ledgrab/api/schemas/calibration.py
index c13561c..ece3713 100644
--- a/server/src/ledgrab/api/schemas/calibration.py
+++ b/server/src/ledgrab/api/schemas/calibration.py
@@ -1,6 +1,6 @@
 """Pydantic schemas for the calibration session and solver API."""
 
-from typing import List, Literal
+from typing import Annotated, List, Literal
 
 from pydantic import BaseModel, Field, model_validator
 
@@ -65,14 +65,15 @@ class CalibrationSolveRequest(BaseModel):
     layout: Literal["clockwise", "counterclockwise"] = Field(
         description="Winding direction of the strip"
     )
-    corner_indices: List[int] = Field(
+    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])."
+            "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,
diff --git a/server/src/ledgrab/static/js/features/setup-wizard.ts b/server/src/ledgrab/static/js/features/setup-wizard.ts
index bb002e2..94a8338 100644
--- a/server/src/ledgrab/static/js/features/setup-wizard.ts
+++ b/server/src/ledgrab/static/js/features/setup-wizard.ts
@@ -27,6 +27,7 @@ import { t } from '../core/i18n.ts';
 import { showToast } from '../core/ui.ts';
 import { Modal } from '../core/modal.ts';
 import { mountAutoCalibration, unmountAutoCalibration } from './auto-calibration.ts';
+import { suppressGettingStartedTour } from './tutorials.ts';
 import {
     ICON_MONITOR, ICON_SPARKLES, ICON_DEVICE, ICON_OK, ICON_CHECK, ICON_ROCKET_ICON,
     ICON_CALIBRATION, ICON_START, ICON_SEARCH, ICON_PLUS,
@@ -76,8 +77,6 @@ interface WizardState {
 let _state: WizardState | null = null;
 let _modal: SetupWizardModal | null = null;
 
-const ONBOARDED_KEY = 'tour_completed';   // mirror tutorials.ts TOUR_KEY
-
 const STEPS: WizardStep[] = ['welcome', 'device', 'display', 'scaffold', 'calibrate', 'start', 'done'];
 
 // ── Modal class ────────────────────────────────────────────────────────────────
@@ -167,7 +166,7 @@ async function _markOnboarded(): Promise<void> {
     try {
         await apiPut('/preferences/onboarding', { onboarded: true });
         // Suppress tooltip tour too — wizard owns the first-run experience
-        localStorage.setItem(ONBOARDED_KEY, '1');
+        suppressGettingStartedTour();
     } catch {
         // Non-fatal: UI already moved on
     }
@@ -770,11 +769,9 @@ function _buildDoneStep(state: WizardState): string {
     </div>`;
 }
 
-function _attachStepListeners(step: WizardStep): void {
-    if (step === 'device') {
-        const form = document.getElementById('wizard-manual-form');
-        if (form) form.addEventListener('submit', (e) => void wizardAddManualDevice(e));
-    }
+function _attachStepListeners(_step: WizardStep): void {
+    // The manual device form uses onsubmit="wizardAddManualDevice(event)" inline —
+    // no duplicate addEventListener needed here.
 }
 
 // ─────────────────────────────────────────────────────────────────────────────
diff --git a/server/tests/api/routes/test_setup_routes.py b/server/tests/api/routes/test_setup_routes.py
index f5a51a8..31ab504 100644
--- a/server/tests/api/routes/test_setup_routes.py
+++ b/server/tests/api/routes/test_setup_routes.py
@@ -21,7 +21,7 @@ from __future__ import annotations
 import pytest
 from fastapi import FastAPI
 from fastapi.testclient import TestClient
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 
 # ---------------------------------------------------------------------------
@@ -89,6 +89,14 @@ def event_log():
     return log
 
 
+@pytest.fixture
+def mock_manager():
+    """A MagicMock ProcessorManager that silently accepts all calls."""
+    mgr = MagicMock()
+    mgr.remove_target = MagicMock()
+    return mgr
+
+
 @pytest.fixture
 def setup_client(
     tmp_db,
@@ -98,6 +106,7 @@ def setup_client(
     css_store,
     output_target_store,
     event_log,
+    mock_manager,
 ):
     from ledgrab.api.routes.setup import router
     from ledgrab.api.auth import verify_api_key
@@ -111,6 +120,7 @@ def setup_client(
     app.dependency_overrides[deps.get_picture_source_store] = lambda: picture_source_store
     app.dependency_overrides[deps.get_color_strip_store] = lambda: css_store
     app.dependency_overrides[deps.get_output_target_store] = lambda: output_target_store
+    app.dependency_overrides[deps.get_processor_manager] = lambda: mock_manager
 
     # Capture entity events
     def _fire(entity_type, action, entity_id):
@@ -512,3 +522,79 @@ class TestScaffoldCalibrationIntegration:
         assert updated.calibration.leds_top == 15
         assert updated.calibration.leds_right == 9
         assert updated.calibration.layout == "clockwise"
+
+
+# ---------------------------------------------------------------------------
+# Regression: scaffold registers the output target with ProcessorManager
+# ---------------------------------------------------------------------------
+
+
+class TestScaffoldRegistersWithManager:
+    def test_scaffold_calls_add_target_on_manager(
+        self,
+        setup_client,
+        sample_device,
+        mock_manager,
+    ):
+        """Blocker regression: scaffold must register the created output target
+        with the ProcessorManager so that a subsequent start call can find it.
+
+        WledOutputTarget.register_with_manager calls manager.add_target(...)
+        — we assert that method was invoked with the scaffolded target id.
+        """
+        resp = _scaffold(setup_client, device_id=sample_device.id)
+        assert resp.status_code == 201, resp.text
+        target_id = resp.json()["output_target_id"]
+
+        # register_with_manager → manager.add_target(target_id=..., ...)
+        mock_manager.add_target.assert_called_once()
+        call_kwargs = mock_manager.add_target.call_args
+        assert call_kwargs.kwargs.get("target_id") == target_id, (
+            f"manager.add_target was not called with target_id={target_id!r}; "
+            f"actual call: {call_kwargs}"
+        )
+
+    def test_scaffold_rollback_calls_remove_target_on_manager(
+        self,
+        setup_client,
+        sample_device,
+        output_target_store,
+        mock_manager,
+    ):
+        """Rollback after a post-target-creation failure must call manager.remove_target
+        so no half-registered target lingers in the ProcessorManager.
+
+        We inject a failure by making register_with_manager raise RuntimeError
+        (bypassing the ValueError-only guard), which puts the outer except branch
+        in play.  The target IS already in created_ids at that point, so rollback
+        must call manager.remove_target(target_id).
+        """
+        created_target_ids: list[str] = []
+
+        original_create = output_target_store.create_wled_target
+
+        def _spy_create(*args, **kwargs):
+            target = original_create(*args, **kwargs)
+            created_target_ids.append(target.id)
+            return target
+
+        output_target_store.create_wled_target = _spy_create
+        try:
+            # Patch register_with_manager on WledOutputTarget to raise RuntimeError —
+            # RuntimeError bypasses the ValueError guard and triggers the outer except,
+            # so rollback fires with the target already in created_ids.
+            with patch(
+                "ledgrab.storage.wled_output_target.WledOutputTarget.register_with_manager",
+                side_effect=RuntimeError("Injected registration failure for rollback test"),
+            ):
+                resp = setup_client.post(
+                    "/api/v1/setup/scaffold",
+                    json={"device_id": sample_device.id, "display_index": 0},
+                )
+            assert resp.status_code == 500, resp.text
+
+            assert len(created_target_ids) == 1, "spy did not record a created target"
+            target_id = created_target_ids[0]
+            mock_manager.remove_target.assert_called_with(target_id)
+        finally:
+            output_target_store.create_wled_target = original_create
diff --git a/server/tests/api/routes/test_setup_routes_adversarial.py b/server/tests/api/routes/test_setup_routes_adversarial.py
new file mode 100644
index 0000000..0ce95f1
--- /dev/null
+++ b/server/tests/api/routes/test_setup_routes_adversarial.py
@@ -0,0 +1,506 @@
+"""Adversarial tests for the setup scaffold and onboarding preference endpoints.
+
+Phase 2 acceptance criteria (NOT what the code happens to do):
+  - Rollback when the FINAL step (output target create) fails leaves ZERO orphans
+    AND emits ZERO "created" events.
+  - Reused capture template is NOT deleted on rollback.
+  - display_index > 63 → 422.
+  - Missing device_id → 422 (Pydantic validation before handler runs).
+  - Unknown device_id → 404.
+  - PUT onboarding false clears completed_at to null.
+  - Corrupt stored onboarding value falls back to default (onboarded=false).
+
+These fill the gaps in the existing 22 happy-path tests.
+"""
+
+from __future__ import annotations
+
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+from unittest.mock import MagicMock, patch
+
+
+# ---------------------------------------------------------------------------
+# Shared fixtures (mirrors test_setup_routes.py exactly)
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def tmp_db(tmp_path):
+    from ledgrab.storage.database import Database
+
+    db = Database(tmp_path / "adv_setup.db")
+    yield db
+    db.close()
+
+
+@pytest.fixture
+def device_store(tmp_db):
+    from ledgrab.storage import DeviceStore
+
+    return DeviceStore(tmp_db)
+
+
+@pytest.fixture
+def template_store(tmp_db):
+    from ledgrab.storage.template_store import TemplateStore
+
+    return TemplateStore(tmp_db)
+
+
+@pytest.fixture
+def picture_source_store(tmp_db):
+    from ledgrab.storage.picture_source_store import PictureSourceStore
+
+    return PictureSourceStore(tmp_db)
+
+
+@pytest.fixture
+def css_store(tmp_db):
+    from ledgrab.storage.color_strip_store import ColorStripStore
+
+    return ColorStripStore(tmp_db)
+
+
+@pytest.fixture
+def output_target_store(tmp_db):
+    from ledgrab.storage.output_target_store import OutputTargetStore
+
+    return OutputTargetStore(tmp_db)
+
+
+@pytest.fixture
+def sample_device(device_store):
+    return device_store.create_device(
+        name="Test LED Strip",
+        url="http://192.168.1.10",
+        led_count=60,
+    )
+
+
+@pytest.fixture
+def event_log():
+    log = []
+    return log
+
+
+@pytest.fixture
+def mock_manager():
+    """A MagicMock ProcessorManager that silently accepts all calls."""
+    mgr = MagicMock()
+    mgr.remove_target = MagicMock()
+    return mgr
+
+
+@pytest.fixture
+def setup_client(
+    tmp_db,
+    device_store,
+    template_store,
+    picture_source_store,
+    css_store,
+    output_target_store,
+    event_log,
+    mock_manager,
+):
+    from ledgrab.api.routes.setup import router
+    from ledgrab.api.auth import verify_api_key
+    from ledgrab.api import dependencies as deps
+
+    app = FastAPI()
+    app.include_router(router)
+    app.dependency_overrides[verify_api_key] = lambda: "test"
+    app.dependency_overrides[deps.get_device_store] = lambda: device_store
+    app.dependency_overrides[deps.get_template_store] = lambda: template_store
+    app.dependency_overrides[deps.get_picture_source_store] = lambda: picture_source_store
+    app.dependency_overrides[deps.get_color_strip_store] = lambda: css_store
+    app.dependency_overrides[deps.get_output_target_store] = lambda: output_target_store
+    app.dependency_overrides[deps.get_processor_manager] = lambda: mock_manager
+
+    def _fire(entity_type, action, entity_id):
+        event_log.append((entity_type, action, entity_id))
+
+    with patch("ledgrab.api.routes.setup.fire_entity_event", side_effect=_fire):
+        yield TestClient(app, raise_server_exceptions=False)
+
+
+def _scaffold(client, **overrides):
+    body = {"display_index": 0, **overrides}
+    return client.post("/api/v1/setup/scaffold", json=body)
+
+
+# ---------------------------------------------------------------------------
+# Rollback when the FINAL step (output target) fails
+# Criteria: "zero orphans AND zero 'created' events"
+# ---------------------------------------------------------------------------
+
+
+class TestFinalStepRollback:
+    def test_final_step_failure_leaves_zero_orphans(
+        self,
+        setup_client,
+        sample_device,
+        picture_source_store,
+        css_store,
+        output_target_store,
+    ):
+        """output_target_store.create_wled_target failing leaves NO orphaned entities."""
+        original_create = output_target_store.create_wled_target
+
+        def _fail(*args, **kwargs):
+            raise ValueError("Injected final step failure")
+
+        output_target_store.create_wled_target = _fail
+        try:
+            resp = _scaffold(setup_client, device_id=sample_device.id)
+            assert resp.status_code in (
+                400,
+                500,
+            ), f"Expected 4xx/5xx on final-step failure, got {resp.status_code}: {resp.text}"
+
+            # Zero picture sources remaining
+            remaining_ps = picture_source_store.get_all_streams()
+            assert len(remaining_ps) == 0, f"Picture sources not rolled back: {remaining_ps}"
+
+            # Zero color-strip sources remaining
+            remaining_css = css_store.get_all_sources()
+            assert len(remaining_css) == 0, f"Color-strip sources not rolled back: {remaining_css}"
+
+            # Zero output targets (trivially true since creation was never reached,
+            # but included for completeness)
+            remaining_ot = output_target_store.get_all_targets()
+            assert len(remaining_ot) == 0, f"Output targets not rolled back: {remaining_ot}"
+        finally:
+            output_target_store.create_wled_target = original_create
+
+    def test_final_step_failure_emits_zero_created_events(
+        self,
+        setup_client,
+        sample_device,
+        output_target_store,
+        event_log,
+    ):
+        """No 'created' events must be emitted when the final step fails (deferred-event contract)."""
+        original_create = output_target_store.create_wled_target
+
+        def _fail(*args, **kwargs):
+            raise ValueError("Final step injected failure")
+
+        output_target_store.create_wled_target = _fail
+        try:
+            resp = _scaffold(setup_client, device_id=sample_device.id)
+            assert resp.status_code in (400, 500)
+
+            created_events = [(et, act) for et, act, _ in event_log if act == "created"]
+            assert (
+                created_events == []
+            ), f"'created' events leaked on final-step rollback: {created_events}"
+        finally:
+            output_target_store.create_wled_target = original_create
+
+    def test_final_step_failure_reused_template_survives(
+        self,
+        setup_client,
+        sample_device,
+        template_store,
+        output_target_store,
+    ):
+        """A reused capture template must NOT be deleted when the final step fails."""
+        templates_before = {t.id for t in template_store.get_all_templates()}
+        original_create = output_target_store.create_wled_target
+
+        def _fail(*args, **kwargs):
+            raise ValueError("Final step injected failure")
+
+        output_target_store.create_wled_target = _fail
+        try:
+            _scaffold(setup_client, device_id=sample_device.id)
+        finally:
+            output_target_store.create_wled_target = original_create
+
+        templates_after = {t.id for t in template_store.get_all_templates()}
+        assert templates_before == templates_after, (
+            f"Template set changed after rollback: "
+            f"before={templates_before} after={templates_after}"
+        )
+
+    def test_final_step_failure_device_not_deleted(
+        self,
+        setup_client,
+        sample_device,
+        device_store,
+        output_target_store,
+    ):
+        """The pre-existing device must never be touched by rollback of any step."""
+        original_create = output_target_store.create_wled_target
+
+        def _fail(*args, **kwargs):
+            raise ValueError("Final step injected failure")
+
+        output_target_store.create_wled_target = _fail
+        try:
+            _scaffold(setup_client, device_id=sample_device.id)
+        finally:
+            output_target_store.create_wled_target = original_create
+
+        device = device_store.get(sample_device.id)
+        assert device is not None, "Pre-existing device was deleted during rollback"
+
+
+# ---------------------------------------------------------------------------
+# Validation: display_index bounds
+# Criteria: display_index > 63 → 422; display_index 63 → 201; negative → 422
+# ---------------------------------------------------------------------------
+
+
+class TestDisplayIndexBounds:
+    def test_display_index_64_returns_422(self, setup_client, sample_device):
+        """display_index=64 (one above max) must be rejected with 422."""
+        resp = _scaffold(setup_client, device_id=sample_device.id, display_index=64)
+        assert (
+            resp.status_code == 422
+        ), f"Expected 422 for display_index=64, got {resp.status_code}: {resp.text}"
+
+    def test_display_index_63_returns_201(self, setup_client, sample_device):
+        """display_index=63 is the maximum valid value — must be accepted."""
+        resp = _scaffold(setup_client, device_id=sample_device.id, display_index=63)
+        assert (
+            resp.status_code == 201
+        ), f"Expected 201 for display_index=63, got {resp.status_code}: {resp.text}"
+
+    def test_display_index_0_returns_201(self, setup_client, sample_device):
+        """display_index=0 is the minimum valid value."""
+        resp = _scaffold(setup_client, device_id=sample_device.id, display_index=0)
+        assert (
+            resp.status_code == 201
+        ), f"Expected 201 for display_index=0, got {resp.status_code}: {resp.text}"
+
+    def test_display_index_negative_1_returns_422(self, setup_client, sample_device):
+        """display_index=-1 must be rejected with 422."""
+        resp = _scaffold(setup_client, device_id=sample_device.id, display_index=-1)
+        assert (
+            resp.status_code == 422
+        ), f"Expected 422 for display_index=-1, got {resp.status_code}: {resp.text}"
+
+    def test_display_index_very_large_returns_422(self, setup_client, sample_device):
+        """display_index=10000 must be rejected with 422."""
+        resp = _scaffold(setup_client, device_id=sample_device.id, display_index=10000)
+        assert (
+            resp.status_code == 422
+        ), f"Expected 422 for display_index=10000, got {resp.status_code}: {resp.text}"
+
+
+# ---------------------------------------------------------------------------
+# Validation: missing / unknown device_id
+# ---------------------------------------------------------------------------
+
+
+class TestDeviceValidation:
+    def test_missing_device_id_returns_422(self, setup_client):
+        """Omitting device_id entirely must yield 422 (Pydantic required field)."""
+        resp = setup_client.post("/api/v1/setup/scaffold", json={"display_index": 0})
+        assert (
+            resp.status_code == 422
+        ), f"Expected 422 for missing device_id, got {resp.status_code}: {resp.text}"
+
+    def test_empty_string_device_id_handled(self, setup_client):
+        """device_id='' — should yield 404 (empty string not in device store) or 422."""
+        resp = _scaffold(setup_client, device_id="")
+        assert resp.status_code in (
+            404,
+            422,
+        ), f"Expected 404 or 422 for empty device_id, got {resp.status_code}: {resp.text}"
+
+    def test_unknown_device_id_returns_404(self, setup_client):
+        """device_id that does not exist in the store must yield 404."""
+        resp = _scaffold(setup_client, device_id="device_definitely_does_not_exist")
+        assert (
+            resp.status_code == 404
+        ), f"Expected 404 for unknown device_id, got {resp.status_code}: {resp.text}"
+
+    def test_none_device_id_returns_422(self, setup_client):
+        """device_id=null must yield 422 (not None-convertible to str)."""
+        resp = setup_client.post(
+            "/api/v1/setup/scaffold", json={"device_id": None, "display_index": 0}
+        )
+        assert (
+            resp.status_code == 422
+        ), f"Expected 422 for null device_id, got {resp.status_code}: {resp.text}"
+
+
+# ---------------------------------------------------------------------------
+# Rollback idempotency: calling scaffold twice with the final step failing
+# must still leave exactly zero orphans (no accumulation across calls).
+# ---------------------------------------------------------------------------
+
+
+class TestRollbackIdempotency:
+    def test_two_failed_scaffolds_leave_zero_orphans(
+        self,
+        setup_client,
+        sample_device,
+        picture_source_store,
+        css_store,
+        output_target_store,
+    ):
+        """Two sequential scaffold failures must not accumulate orphans."""
+        original_create = output_target_store.create_wled_target
+
+        def _fail(*args, **kwargs):
+            raise ValueError("Always fails")
+
+        output_target_store.create_wled_target = _fail
+        try:
+            _scaffold(setup_client, device_id=sample_device.id)
+            _scaffold(setup_client, device_id=sample_device.id)
+        finally:
+            output_target_store.create_wled_target = original_create
+
+        assert (
+            len(picture_source_store.get_all_streams()) == 0
+        ), "Picture sources accumulated across two failed scaffolds"
+        assert (
+            len(css_store.get_all_sources()) == 0
+        ), "Color-strip sources accumulated across two failed scaffolds"
+
+
+# ---------------------------------------------------------------------------
+# Onboarding: adversarial cases
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture
+def pref_client(tmp_db):
+    from ledgrab.api.routes.preferences import router
+    from ledgrab.api.auth import verify_api_key
+    from ledgrab.api import dependencies as deps
+
+    app = FastAPI()
+    app.include_router(router)
+    app.dependency_overrides[verify_api_key] = lambda: "test"
+    app.dependency_overrides[deps.get_database] = lambda: tmp_db
+
+    return TestClient(app)
+
+
+class TestOnboardingAdversarial:
+    def test_put_false_clears_completed_at(self, pref_client):
+        """PUT onboarded=false must clear completed_at to null, per criteria."""
+        # First mark as onboarded
+        r1 = pref_client.put("/api/v1/preferences/onboarding", json={"onboarded": True})
+        assert r1.status_code == 200
+        assert r1.json()["completed_at"] is not None
+
+        # Now set to false — completed_at must be cleared
+        r2 = pref_client.put("/api/v1/preferences/onboarding", json={"onboarded": False})
+        assert r2.status_code == 200
+        data = r2.json()
+        assert data["onboarded"] is False
+        assert (
+            data["completed_at"] is None
+        ), f"completed_at should be null after PUT onboarded=false, got {data['completed_at']!r}"
+
+    def test_put_false_then_get_returns_null_completed_at(self, pref_client):
+        """After PUT false, GET must also return null completed_at (persisted correctly)."""
+        pref_client.put("/api/v1/preferences/onboarding", json={"onboarded": True})
+        pref_client.put("/api/v1/preferences/onboarding", json={"onboarded": False})
+        resp = pref_client.get("/api/v1/preferences/onboarding")
+        assert resp.status_code == 200
+        assert (
+            resp.json()["completed_at"] is None
+        ), "GET after PUT false should return null completed_at"
+
+    def test_corrupt_stored_value_falls_back_to_default(self, tmp_db, pref_client):
+        """If the stored onboarding value is corrupt, GET must fall back to default.
+
+        Criteria: "corrupt stored value falls back to default".
+        We inject garbage into the db directly, then hit GET.
+        """
+        # Inject a value that is syntactically valid JSON (dict) but fails
+        # Pydantic validation because the types are wrong.
+        tmp_db.set_setting("onboarded", {"onboarded": "not_a_bool", "completed_at": 12345})
+
+        resp = pref_client.get("/api/v1/preferences/onboarding")
+        assert (
+            resp.status_code == 200
+        ), f"Expected 200 for corrupt onboarding value, got {resp.status_code}"
+        data = resp.json()
+        # Must fall back to default
+        assert (
+            data["onboarded"] is False
+        ), f"Expected onboarded=false as default after corrupt value, got {data['onboarded']!r}"
+        assert (
+            data["completed_at"] is None
+        ), f"Expected completed_at=null as default, got {data['completed_at']!r}"
+
+    def test_corrupt_stored_value_as_wrong_type_falls_back(self, tmp_db, pref_client):
+        """Stored value is a string (not a dict) — must fall back to default."""
+        tmp_db.set_setting("onboarded", "this_is_not_valid")
+
+        resp = pref_client.get("/api/v1/preferences/onboarding")
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["onboarded"] is False
+        assert data["completed_at"] is None
+
+    def test_corrupt_stored_null_falls_back_to_default(self, tmp_db, pref_client):
+        """Stored value is null/None — must return default (not crash)."""
+        tmp_db.set_setting("onboarded", None)
+
+        resp = pref_client.get("/api/v1/preferences/onboarding")
+        assert resp.status_code == 200
+        assert resp.json()["onboarded"] is False
+
+    def test_put_true_without_completed_at_stamps_timestamp(self, pref_client):
+        """PUT onboarded=true without completed_at must auto-stamp a timestamp."""
+        resp = pref_client.put("/api/v1/preferences/onboarding", json={"onboarded": True})
+        assert resp.status_code == 200
+        data = resp.json()
+        assert (
+            data["completed_at"] is not None
+        ), "Server must auto-stamp completed_at when onboarded=true is sent without a timestamp"
+        # Should be a non-empty ISO timestamp string
+        assert len(data["completed_at"]) > 10
+
+    def test_put_true_with_completed_at_preserves_it(self, pref_client):
+        """PUT onboarded=true with explicit completed_at must preserve that value."""
+        ts = "2025-03-15T10:00:00+00:00"
+        resp = pref_client.put(
+            "/api/v1/preferences/onboarding",
+            json={"onboarded": True, "completed_at": ts},
+        )
+        assert resp.status_code == 200
+        assert resp.json()["completed_at"] == ts
+
+    def test_put_false_with_completed_at_clears_it(self, pref_client):
+        """PUT onboarded=false even with a completed_at payload must clear it.
+
+        The criteria say: 'Setting onboarded=false clears completed_at to null.'
+        """
+        ts = "2025-01-01T00:00:00+00:00"
+        resp = pref_client.put(
+            "/api/v1/preferences/onboarding",
+            json={"onboarded": False, "completed_at": ts},
+        )
+        assert resp.status_code == 200
+        data = resp.json()
+        assert data["onboarded"] is False
+        assert (
+            data["completed_at"] is None
+        ), f"completed_at should be cleared to null when onboarded=false, got {data['completed_at']!r}"
+
+    def test_multiple_true_puts_only_stamp_once(self, pref_client):
+        """Two successive PUT true calls — second call must preserve the original timestamp."""
+        r1 = pref_client.put("/api/v1/preferences/onboarding", json={"onboarded": True})
+        ts1 = r1.json()["completed_at"]
+        assert ts1 is not None
+
+        # Explicitly provide the original timestamp in second call
+        r2 = pref_client.put(
+            "/api/v1/preferences/onboarding",
+            json={"onboarded": True, "completed_at": ts1},
+        )
+        assert (
+            r2.json()["completed_at"] == ts1
+        ), "Explicit completed_at should be preserved on second PUT"
diff --git a/server/tests/core/test_calibration_session_adversarial.py b/server/tests/core/test_calibration_session_adversarial.py
new file mode 100644
index 0000000..c0635cc
--- /dev/null
+++ b/server/tests/core/test_calibration_session_adversarial.py
@@ -0,0 +1,535 @@
+"""Adversarial / concurrency tests for CalibrationSession.
+
+Phase 1 acceptance criteria tested here (NOT what the code happens to do):
+  - Interleaved start/start (same device, then different device) must never
+    leave the old device without restore.
+  - Interleaved start/stop racing the idle watchdog must not leave the device
+    dark or stuck.
+  - Idle-timeout teardown restores the prior target.
+  - position() with index out of range → ValueError.
+  - stop() when idle is a safe no-op (does not call start_processing or crash).
+  - CalibrationSession lock must prevent double-teardown.
+
+All tests use a fake ProcessorManager matching the shape used in
+test_calibration_routes.py.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+import pytest_asyncio
+
+from ledgrab.core.capture.calibration_session import (
+    CalibrationSession,
+)
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+def _make_manager(device_id: str = "dev1", led_count: int = 100) -> MagicMock:
+    """Build a minimal fake ProcessorManager."""
+    mgr = MagicMock()
+    ds = MagicMock()
+    ds.led_count = led_count
+    mgr._devices = {device_id: ds}
+    mgr.get_processing_target_for_device = MagicMock(return_value=None)
+    mgr.stop_processing = AsyncMock()
+    mgr.start_processing = AsyncMock()
+    mgr.send_clear_pixels = AsyncMock()
+    mgr.set_calibration_pixel = AsyncMock()
+    return mgr
+
+
+@pytest_asyncio.fixture(autouse=True)
+async def fresh_session():
+    """Yield a brand-new CalibrationSession for each test (not the singleton)."""
+    session = CalibrationSession()
+    yield session
+    # Cleanup: cancel any lingering watchdog
+    if session._timeout_task and not session._timeout_task.done():
+        session._timeout_task.cancel()
+        try:
+            await session._timeout_task
+        except (asyncio.CancelledError, Exception):
+            pass
+
+
+# ---------------------------------------------------------------------------
+# stop() when idle is a safe no-op
+# ---------------------------------------------------------------------------
+
+
+class TestStopWhenIdle:
+    @pytest.mark.asyncio
+    async def test_stop_idle_does_not_call_start_processing(self, fresh_session):
+        """Calling stop() when no session is active must not call start_processing."""
+        mgr = _make_manager()
+        # Do NOT start a session — just stop immediately
+        await fresh_session.stop()
+        mgr.start_processing.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_stop_idle_returns_inactive_state(self, fresh_session):
+        """stop() on an idle session returns state with active=False."""
+        state = fresh_session.get_state()
+        assert state["active"] is False
+        await fresh_session.stop()  # no-op
+        assert fresh_session.is_active is False
+
+    @pytest.mark.asyncio
+    async def test_cancel_idle_safe(self, fresh_session):
+        """cancel() on idle session is also a safe no-op."""
+        await fresh_session.cancel()
+        assert fresh_session.is_active is False
+
+    @pytest.mark.asyncio
+    async def test_double_stop_is_idempotent(self, fresh_session):
+        """Calling stop() twice on an active session must not double-call start_processing."""
+        mgr = _make_manager()
+        mgr.get_processing_target_for_device = MagicMock(return_value="tgt_restore")
+        await fresh_session.start("dev1", mgr)
+        assert fresh_session.is_active is True
+
+        await fresh_session.stop()
+        assert fresh_session.is_active is False
+        # Restore called exactly once
+        mgr.start_processing.assert_awaited_once_with("tgt_restore")
+
+        # Second stop must be a no-op
+        await fresh_session.stop()
+        # start_processing should still be called exactly once (not twice)
+        assert mgr.start_processing.await_count == 1
+
+
+# ---------------------------------------------------------------------------
+# position() out of range
+# ---------------------------------------------------------------------------
+
+
+class TestPositionOutOfRange:
+    @pytest.mark.asyncio
+    async def test_position_equal_to_led_count_raises(self, fresh_session):
+        """index == led_count must raise ValueError (0-based, so out of range)."""
+        mgr = _make_manager(led_count=100)
+        await fresh_session.start("dev1", mgr)
+        with pytest.raises(ValueError, match="out of range"):
+            await fresh_session.position(100)
+
+    @pytest.mark.asyncio
+    async def test_position_above_led_count_raises(self, fresh_session):
+        """index > led_count raises ValueError."""
+        mgr = _make_manager(led_count=50)
+        await fresh_session.start("dev1", mgr)
+        with pytest.raises(ValueError, match="out of range"):
+            await fresh_session.position(999)
+
+    @pytest.mark.asyncio
+    async def test_position_negative_raises(self, fresh_session):
+        """Negative index raises ValueError."""
+        mgr = _make_manager(led_count=100)
+        await fresh_session.start("dev1", mgr)
+        with pytest.raises(ValueError):
+            await fresh_session.position(-1)
+
+    @pytest.mark.asyncio
+    async def test_position_at_led_count_minus_1_is_valid(self, fresh_session):
+        """Last valid index (led_count - 1) must succeed."""
+        mgr = _make_manager(led_count=10)
+        await fresh_session.start("dev1", mgr)
+        await fresh_session.position(9)  # must not raise
+        mgr.set_calibration_pixel.assert_awaited()
+
+    @pytest.mark.asyncio
+    async def test_position_without_active_session_raises(self, fresh_session):
+        """position() with no active session must raise RuntimeError."""
+        with pytest.raises(RuntimeError, match="No active calibration session"):
+            await fresh_session.position(5)
+
+
+# ---------------------------------------------------------------------------
+# Interleaved start/start — old device must be restored
+# ---------------------------------------------------------------------------
+
+
+class TestInterleavedStartStart:
+    @pytest.mark.asyncio
+    async def test_start_on_same_device_restores_prior_target(self, fresh_session):
+        """Starting a second session on the same device auto-stops the first.
+        The first device's prior target must be restored before the second session begins.
+        """
+        mgr = _make_manager(led_count=60)
+        mgr.get_processing_target_for_device = MagicMock(return_value="tgt_original")
+
+        # Start first session
+        await fresh_session.start("dev1", mgr)
+        assert fresh_session.is_active is True
+        assert fresh_session._prior_target_id == "tgt_original"
+
+        # Now start again on the same device
+        # The second start should stop the first (restoring tgt_original),
+        # then re-query the current running target (which will be tgt_original again
+        # since start_processing will have been called).
+        # For isolation: change what get_processing_target_for_device returns after
+        # the first stop so the second session records a fresh prior.
+        call_count = {"n": 0}
+
+        def _get_target(device_id):
+            call_count["n"] += 1
+            if call_count["n"] == 1:
+                return "tgt_original"
+            return None  # After first stop, no target running
+
+        mgr.get_processing_target_for_device = MagicMock(side_effect=_get_target)
+
+        # First session with the original target
+        fresh_session2 = CalibrationSession()
+        await fresh_session2.start("dev1", mgr)
+        assert fresh_session2.is_active is True
+
+        # Start a NEW session on the same device — must auto-stop fresh_session2
+        fresh_session3 = CalibrationSession()
+        # Inject fresh_session2's internal state into fresh_session3 to simulate
+        # the singleton pattern: replace session3's state to reflect session2 active
+        # (this mirrors the module-level singleton where only one CalibrationSession exists)
+        fresh_session3._active = fresh_session2._active
+        fresh_session3._device_id = fresh_session2._device_id
+        fresh_session3._led_count = fresh_session2._led_count
+        fresh_session3._prior_target_id = fresh_session2._prior_target_id
+        fresh_session3._last_activity = fresh_session2._last_activity
+        fresh_session3._manager = fresh_session2._manager
+        fresh_session3._timeout_task = fresh_session2._timeout_task
+        fresh_session2._timeout_task = None  # prevent double-cancel
+
+        await fresh_session3.start("dev1", mgr)
+
+        # The start must have called stop on the previous session → restore was called
+        # (mgr.start_processing was called at least once to restore the prior target)
+        assert mgr.start_processing.await_count >= 1
+
+        # Cleanup
+        await fresh_session3.stop()
+        if fresh_session2._timeout_task and not fresh_session2._timeout_task.done():
+            fresh_session2._timeout_task.cancel()
+
+    @pytest.mark.asyncio
+    async def test_new_session_on_different_device_clears_old_device(self, fresh_session):
+        """Starting a new session on a different device must clear the first device.
+
+        The first session must be stopped (its prior target restored or cleared)
+        before the second session on the new device becomes active.
+        """
+        mgr = MagicMock()
+
+        ds1 = MagicMock()
+        ds1.led_count = 30
+        ds2 = MagicMock()
+        ds2.led_count = 60
+
+        mgr._devices = {"dev1": ds1, "dev2": ds2}
+        mgr.get_processing_target_for_device = MagicMock(return_value=None)
+        mgr.stop_processing = AsyncMock()
+        mgr.start_processing = AsyncMock()
+        mgr.send_clear_pixels = AsyncMock()
+        mgr.set_calibration_pixel = AsyncMock()
+
+        # Start first session on dev1
+        await fresh_session.start("dev1", mgr)
+        assert fresh_session._device_id == "dev1"
+        assert fresh_session.is_active is True
+
+        # Now start second session on dev2 — must auto-stop dev1 first
+        await fresh_session.start("dev2", mgr)
+
+        # After the second start, session must be on dev2
+        assert fresh_session._device_id == "dev2"
+        assert fresh_session.is_active is True
+
+        # send_clear_pixels was called for dev1 (stop) AND for dev2 (start)
+        clear_calls = [call[0][0] for call in mgr.send_clear_pixels.call_args_list]
+        assert (
+            "dev1" in clear_calls
+        ), f"dev1 was never cleared during session switch; clear calls: {clear_calls}"
+        assert (
+            "dev2" in clear_calls
+        ), f"dev2 was never cleared at session start; clear calls: {clear_calls}"
+
+        # Cleanup
+        await fresh_session.stop()
+
+
+# ---------------------------------------------------------------------------
+# Idle-timeout teardown restores prior target
+# ---------------------------------------------------------------------------
+
+
+class TestIdleTimeoutRestoresPriorTarget:
+    @pytest.mark.asyncio
+    async def test_idle_timeout_calls_start_processing(self, fresh_session):
+        """When the session times out, start_processing must be called to restore the target.
+
+        We patch IDLE_TIMEOUT_SECONDS to a tiny value so the test doesn't actually
+        wait 60 seconds.
+        """
+        mgr = _make_manager(led_count=40)
+        mgr.get_processing_target_for_device = MagicMock(return_value="tgt_to_restore")
+
+        # Patch the idle timeout to 0.05 seconds
+        with patch(
+            "ledgrab.core.capture.calibration_session.IDLE_TIMEOUT_SECONDS",
+            0.05,
+        ):
+            # Also patch the watchdog sleep to something tiny
+            async def _fast_watchdog():
+                """A watchdog that checks every 0.02 seconds instead of 5."""
+                try:
+                    while True:
+                        await asyncio.sleep(0.02)
+                        if not fresh_session._active or fresh_session._last_activity is None:
+                            break
+                        from datetime import datetime, timezone
+
+                        elapsed = (
+                            datetime.now(timezone.utc) - fresh_session._last_activity
+                        ).total_seconds()
+                        if elapsed >= 0.05:
+                            async with fresh_session._lock:
+                                await fresh_session._teardown_locked(cancelled=False)
+                            break
+                except asyncio.CancelledError:
+                    pass
+
+            fresh_session._idle_watchdog = _fast_watchdog  # type: ignore[method-assign]
+
+            await fresh_session.start("dev1", mgr)
+            assert fresh_session.is_active is True
+
+            # Wait long enough for the watchdog to fire
+            await asyncio.sleep(0.25)
+
+            # Session should have been auto-stopped
+            assert (
+                fresh_session.is_active is False
+            ), "Session should have been auto-stopped by idle timeout"
+
+            # Prior target must have been restored
+            mgr.start_processing.assert_awaited_once_with("tgt_to_restore")
+
+    @pytest.mark.asyncio
+    async def test_idle_timeout_clears_device_to_black(self, fresh_session):
+        """Idle timeout must send all-black before (or after) restoring target."""
+        mgr = _make_manager(led_count=40)
+
+        async def _fast_watchdog():
+            try:
+                while True:
+                    await asyncio.sleep(0.02)
+                    if not fresh_session._active or fresh_session._last_activity is None:
+                        break
+                    from datetime import datetime, timezone
+
+                    elapsed = (
+                        datetime.now(timezone.utc) - fresh_session._last_activity
+                    ).total_seconds()
+                    if elapsed >= 0.05:
+                        async with fresh_session._lock:
+                            await fresh_session._teardown_locked(cancelled=False)
+                        break
+            except asyncio.CancelledError:
+                pass
+
+        fresh_session._idle_watchdog = _fast_watchdog  # type: ignore[method-assign]
+
+        await fresh_session.start("dev1", mgr)
+        initial_clear_count = mgr.send_clear_pixels.await_count  # from start()
+
+        await asyncio.sleep(0.25)
+
+        # send_clear_pixels must have been called at least once more during teardown
+        assert (
+            mgr.send_clear_pixels.await_count > initial_clear_count
+        ), "Device was not cleared to black during idle-timeout teardown"
+
+
+# ---------------------------------------------------------------------------
+# Concurrent start/start using asyncio.gather (true concurrency within the loop)
+# ---------------------------------------------------------------------------
+
+
+class TestConcurrentStartCalls:
+    @pytest.mark.asyncio
+    async def test_concurrent_starts_dont_corrupt_state(self, fresh_session):
+        """Two concurrent start() calls must leave the session in a consistent state.
+
+        Only one session should be active after both complete; the final device
+        must match one of the two requested devices.
+        """
+        mgr = MagicMock()
+        ds1 = MagicMock()
+        ds1.led_count = 50
+        ds2 = MagicMock()
+        ds2.led_count = 80
+        mgr._devices = {"devA": ds1, "devB": ds2}
+        mgr.get_processing_target_for_device = MagicMock(return_value=None)
+        mgr.stop_processing = AsyncMock()
+        mgr.start_processing = AsyncMock()
+        mgr.send_clear_pixels = AsyncMock()
+        mgr.set_calibration_pixel = AsyncMock()
+
+        # Fire both concurrently — the lock must ensure exactly one wins
+        results = await asyncio.gather(
+            fresh_session.start("devA", mgr),
+            fresh_session.start("devB", mgr),
+            return_exceptions=True,
+        )
+
+        # Neither call should raise (both complete without exception)
+        for r in results:
+            if isinstance(r, BaseException):
+                pytest.fail(f"Concurrent start raised unexpectedly: {r!r}")
+
+        # Exactly one session active with a valid device
+        assert fresh_session.is_active is True
+        assert fresh_session._device_id in ("devA", "devB")
+
+        # Cleanup
+        await fresh_session.stop()
+
+
+# ---------------------------------------------------------------------------
+# stop() while watchdog is still pending (concurrent stop/watchdog)
+# ---------------------------------------------------------------------------
+
+
+class TestStopVsWatchdogRace:
+    @pytest.mark.asyncio
+    async def test_explicit_stop_wins_over_watchdog(self, fresh_session):
+        """Explicit stop() must cleanly terminate before the watchdog fires.
+
+        After explicit stop, start_processing must be called exactly once
+        even if the watchdog later tries to tear down.
+        """
+        mgr = _make_manager(led_count=100)
+        mgr.get_processing_target_for_device = MagicMock(return_value="tgt_explicit")
+
+        await fresh_session.start("dev1", mgr)
+        assert fresh_session.is_active is True
+
+        # Explicitly stop — the watchdog task should be cancelled
+        await fresh_session.stop()
+
+        assert fresh_session.is_active is False
+        # start_processing called exactly once for the prior target
+        mgr.start_processing.assert_awaited_once_with("tgt_explicit")
+
+        # Give the event loop a moment to confirm the watchdog didn't double-fire
+        await asyncio.sleep(0.05)
+        # Still exactly once
+        assert (
+            mgr.start_processing.await_count == 1
+        ), "start_processing was called more than once — watchdog fired after explicit stop"
+
+
+# ---------------------------------------------------------------------------
+# start() with unknown device_id
+# ---------------------------------------------------------------------------
+
+
+class TestStartUnknownDevice:
+    @pytest.mark.asyncio
+    async def test_unknown_device_raises_valueerror(self, fresh_session):
+        """start() with a device_id not in manager._devices must raise ValueError."""
+        mgr = _make_manager(device_id="known_device")
+        with pytest.raises(ValueError, match="not found"):
+            await fresh_session.start("does_not_exist", mgr)
+
+    @pytest.mark.asyncio
+    async def test_unknown_device_leaves_session_inactive(self, fresh_session):
+        """After a failed start() the session must remain inactive."""
+        mgr = _make_manager(device_id="known_device")
+        try:
+            await fresh_session.start("no_such_device", mgr)
+        except ValueError:
+            pass
+        assert fresh_session.is_active is False
+
+    @pytest.mark.asyncio
+    async def test_failed_start_does_not_corrupt_existing_session(self, fresh_session):
+        """A failed start() attempt must not corrupt an already-active session."""
+        mgr = MagicMock()
+        ds = MagicMock()
+        ds.led_count = 100
+        mgr._devices = {"dev1": ds}
+        mgr.get_processing_target_for_device = MagicMock(return_value="prior_tgt")
+        mgr.stop_processing = AsyncMock()
+        mgr.start_processing = AsyncMock()
+        mgr.send_clear_pixels = AsyncMock()
+        mgr.set_calibration_pixel = AsyncMock()
+
+        # Start a valid session
+        await fresh_session.start("dev1", mgr)
+        assert fresh_session.is_active is True
+
+        # Now try to start on an unknown device — should fail
+        try:
+            await fresh_session.start("unknown_device", mgr)
+        except ValueError:
+            pass
+        except Exception:
+            pass  # Other errors are also acceptable
+
+        # The original session must still be active (or was cleanly stopped and
+        # replaced); either way the device must not be stuck in a broken state.
+        # The critical invariant: if active, device_id is valid.
+        if fresh_session.is_active:
+            assert fresh_session._device_id in mgr._devices
+
+        await fresh_session.stop()
+
+
+# ---------------------------------------------------------------------------
+# State snapshot integrity
+# ---------------------------------------------------------------------------
+
+
+class TestStateSnapshot:
+    @pytest.mark.asyncio
+    async def test_get_state_before_start(self, fresh_session):
+        state = fresh_session.get_state()
+        assert state["active"] is False
+        assert state["device_id"] is None
+        assert state["led_count"] == 0
+        assert state["prior_target_id"] is None
+        assert state["last_activity"] is None
+
+    @pytest.mark.asyncio
+    async def test_get_state_after_start(self, fresh_session):
+        mgr = _make_manager(led_count=60)
+        mgr.get_processing_target_for_device = MagicMock(return_value="saved_tgt")
+        await fresh_session.start("dev1", mgr)
+
+        state = fresh_session.get_state()
+        assert state["active"] is True
+        assert state["device_id"] == "dev1"
+        assert state["led_count"] == 60
+        assert state["prior_target_id"] == "saved_tgt"
+        assert state["last_activity"] is not None
+
+        await fresh_session.stop()
+
+    @pytest.mark.asyncio
+    async def test_get_state_after_stop(self, fresh_session):
+        mgr = _make_manager()
+        await fresh_session.start("dev1", mgr)
+        await fresh_session.stop()
+
+        state = fresh_session.get_state()
+        assert state["active"] is False
+        assert state["device_id"] is None
+        assert state["led_count"] == 0
+        assert state["prior_target_id"] is None
diff --git a/server/tests/core/test_calibration_solver_adversarial.py b/server/tests/core/test_calibration_solver_adversarial.py
new file mode 100644
index 0000000..c1d291a
--- /dev/null
+++ b/server/tests/core/test_calibration_solver_adversarial.py
@@ -0,0 +1,562 @@
+"""Adversarial / edge-case tests for solve_calibration() and build_segments().
+
+Criteria derived from Phase 1 acceptance criteria (NOT from what the code does):
+  - solve_calibration returns correct per-edge counts; result round-trips through
+    build_segments() and totals are preserved.
+  - An edge with 0 LEDs (two corners tapped adjacent) is valid.
+  - Wrap-around edges work correctly.
+  - Invalid inputs raise cleanly (ValueError).
+
+New adversarial cases not covered by the existing 19 happy-path tests:
+  - All four corner_indices equal (degenerate: every edge = 0) — the code must
+    either handle it gracefully (total=0) OR raise with a clear message.
+    Per criteria the total must be preserved (0 == 0) and round-trip must not crash.
+  - Descending / out-of-order corner_indices where wrap-around should apply.
+  - An edge that spans the whole strip (one edge wraps the full led_count minus
+    the contribution of the three zero-LED edges).
+  - led_count just above minimum (led_count=1) with a single LED claimed by
+    one edge.
+  - offset >= led_count: the solver stores offset verbatim; PixelMapper normalises
+    it via % total_leds.  The build_segments() round-trip must not crash.
+  - Corner indices modulo led_count are used, so indices >= led_count should be
+    accepted and reduced.
+"""
+
+import pytest
+
+from ledgrab.core.capture.calibration import (
+    EDGE_ORDER,
+    CalibrationConfig,
+    solve_calibration,
+)
+
+
+# ---------------------------------------------------------------------------
+# Helper (same as in test_calibration_solver.py — duplicated intentionally so
+# this adversarial file can run in isolation)
+# ---------------------------------------------------------------------------
+
+
+def _roundtrip(cfg: CalibrationConfig) -> None:
+    """Assert build_segments() doesn't crash and totals match."""
+    segs = cfg.build_segments()
+    total_from_segs = sum(s.led_count for s in segs)
+    expected = cfg.leds_top + cfg.leds_right + cfg.leds_bottom + cfg.leds_left
+    assert (
+        total_from_segs == expected
+    ), f"Segment total {total_from_segs} != field total {expected} for cfg={cfg!r}"
+
+
+# ---------------------------------------------------------------------------
+# Degenerate: all four corner indices are equal
+# When all four corners are tapped at the same index every consecutive pair
+# has start_idx == end_idx, so every edge gets 0 LEDs.  Total = 0 is
+# mathematically consistent with the input; the function should either
+# return a config with 0-LED edges OR raise a clear ValueError.
+# It must NOT silently return wrong counts and must NOT crash unexpectedly.
+# ---------------------------------------------------------------------------
+
+
+class TestAllFourCornersEqual:
+    """All four corner_indices equal → every edge = 0 LEDs or clean ValueError."""
+
+    @pytest.mark.parametrize(
+        "start_position,layout",
+        [(sp, lay) for sp, lay in EDGE_ORDER.keys()],
+    )
+    def test_all_equal_returns_zero_total_or_raises(self, start_position, layout):
+        """solve_calibration([k,k,k,k]) must not produce non-zero counts."""
+        led_count = 100
+        try:
+            cfg = solve_calibration(
+                led_count=led_count,
+                start_position=start_position,
+                layout=layout,
+                corner_indices=[0, 0, 0, 0],
+            )
+            # If it doesn't raise: total must be 0 (consistent with input)
+            total = cfg.leds_top + cfg.leds_right + cfg.leds_bottom + cfg.leds_left
+            assert (
+                total == 0
+            ), f"{start_position}/{layout}: all-equal corners produced total={total}; expected 0"
+            # build_segments must not crash on a 0-total config
+            segs = cfg.build_segments()
+            assert segs == [], f"Expected no segments for 0-LED config, got {segs!r}"
+        except ValueError:
+            # Raising is also acceptable — as long as it's a ValueError, not a
+            # crash/assertion/other exception.
+            pass
+
+    def test_all_equal_roundtrip_does_not_crash(self):
+        """Even if all edges are 0, build_segments() must not raise."""
+        try:
+            cfg = solve_calibration(
+                led_count=50,
+                start_position="top_left",
+                layout="clockwise",
+                corner_indices=[25, 25, 25, 25],
+            )
+            _roundtrip(cfg)  # must not crash
+        except ValueError:
+            pass  # acceptable
+
+
+# ---------------------------------------------------------------------------
+# Descending / out-of-order corner indices
+# Out-of-order but non-wrapping: e.g. [70, 50, 30, 10] for clockwise bottom_left
+# The solver uses consecutive-pair differences with wrap logic.
+# Each pair (70→50, 50→30, 30→10, 10→70) has end < start, so ALL four edges
+# get wrap-around counts.  Total must still equal led_count.
+# ---------------------------------------------------------------------------
+
+
+class TestDescendingCornerIndices:
+    """All four corners in descending order — every pair wraps."""
+
+    def test_descending_total_equals_led_count(self):
+        """Descending indices [75, 50, 25, 0] — total must == led_count."""
+        # bottom_left/clockwise: edge_order=[left,top,right,bottom]
+        # left: 75→50 = 25, top: 50→25 = 25, right: 25→0 = 25, bottom: 0→75 = 75
+        # Wait — end > start for bottom: 0→75 means 75.  But 0 < 75 so count=75-0=75.
+        # Recalculate: for bottom: start_idx=0, end_idx=75 → end>start → count=75
+        # total = 25+25+25+75 = 150 ≠ 100  → That's wrong input.
+        # Use [80, 60, 40, 20] for 100 LEDs:
+        #   left: 80→60: end<start → wrap: (100-80)+60=80  WRONG too.
+        # Actually for a valid descending case that totals 100:
+        # Monotone descending with wrap on last pair only:
+        # [75, 50, 25, 0] for bottom_left/clockwise:
+        #   left: start=75,end=50: 50<75 → wrap: (100-75)+50=75  NO
+        # All pairs descend when [75,50,25,0]:
+        #   0→75 (last pair, bottom): 75>0 → count=75-0=75
+        #   75→50 (left): 50<75 → wrap: (100-75)+50=75
+        #   Total would be 75+25+25+75=200 ≠ 100
+        # That shows descending indices don't sum to led_count in general.
+        # The critical invariant is: sum of per-edge counts == led_count when
+        # the corner indices span a single full traversal.
+        # To get descending [80,60,40,20] → sum via wrap logic:
+        #   left: 80→60: 60<80 → (100-80)+60=80
+        #   top: 60→40: 40<60 → (100-60)+40=80
+        #   right: 40→20: 20<40 → (100-40)+20=80
+        #   bottom: 20→80: 80>20 → 80-20=60
+        # total = 80+80+80+60 = 300  — still not 100.
+        #
+        # The INVARIANT of solve_calibration is: if corner_indices form a valid
+        # partition of the strip (i.e. each consecutive pair covers a segment
+        # that together span exactly led_count LEDs), the total == led_count.
+        # Descending indices that form valid partitions do sum to led_count.
+        # Example: if we want all edges wrapped, use [99, 74, 49, 24]:
+        #   left: 99→74: 74<99 → (100-99)+74=75
+        #   top: 74→49: 49<74 → (100-74)+49=75
+        #   right: 49→24: 24<49 → (100-49)+24=75
+        #   bottom: 24→99: 99>24 → 99-24=75
+        #   total = 75+75+75+75=300 ≠ 100
+        #
+        # Conclusion: descending indices don't need to sum to led_count — only
+        # a proper strip traversal (covering each LED exactly once) does.
+        # The adversarial test here is: the function must NOT crash, must NOT
+        # produce negative counts, and must round-trip cleanly.
+        cfg = solve_calibration(
+            led_count=100,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[80, 60, 40, 20],
+        )
+        counts = [cfg.leds_top, cfg.leds_right, cfg.leds_bottom, cfg.leds_left]
+        assert all(c >= 0 for c in counts), f"Negative edge counts: {counts}"
+        _roundtrip(cfg)
+
+    def test_all_four_wrap_around_non_negative(self):
+        """Every consecutive pair wraps (descending) — all counts must be >= 0."""
+        # [90, 70, 50, 30] for led_count=100, top_right/clockwise
+        cfg = solve_calibration(
+            led_count=100,
+            start_position="top_right",
+            layout="clockwise",
+            corner_indices=[90, 70, 50, 30],
+        )
+        counts = [cfg.leds_top, cfg.leds_right, cfg.leds_bottom, cfg.leds_left]
+        assert all(c >= 0 for c in counts), f"Negative edge counts: {counts}"
+        _roundtrip(cfg)
+
+
+# ---------------------------------------------------------------------------
+# Edge that wraps the whole strip
+# When only one edge is used (the other 3 are 0 LEDs), that edge should span
+# led_count LEDs.  Corner indices: [0, 0, 0, 0] gives all-zero (tested above).
+# A single-edge case: [0, 100, 100, 100] would give left=100, top=0, right=0,
+# bottom=0 for bottom_left/clockwise.  But 100 % 100 = 0 so end_idx=0,
+# start_idx=0 → count=0 for left too.  Use index 100 directly (>led_count).
+# Alternatively: for led_count=100, corners [0, 0, 0, 0] with all-zero is
+# already tested.  The wrap-all case uses non-modulo indices.
+#
+# The real case: one valid single-edge scenario:
+#   bottom_left/clockwise, 100 LEDs, EDGE_ORDER = [left,top,right,bottom]
+#   corners [0, 100, 100, 100]:  left: 0→100%100=0 → 0-LED because end==start.
+#   No, there's no way to make one edge claim all 100 LEDs with index-based input
+#   other than having it wrap around.  Test: [50, 50, 50, 50] (all equal):
+#   already tested above.
+# The closest adversarial case: one edge with count == led_count via wrap-around.
+# For bottom_left/clockwise, [50, 50, 50, 50] makes top=0,right=0,bottom=0,left=0.
+# To make ONE edge == 100: we need start_idx=50, end_idx=50 for three edges
+# and start=50, end=50 wraps to 0 for that edge... that's the all-zeros case.
+# The only way one edge gets ALL leds is:
+#   e.g. left: corners[0]=50, corners[1]=50 → 50==50 → count=0 (NOT 100).
+#   There's a subtle algorithmic distinction: adjacent indices being equal → 0 LED
+#   vs wrap-around: if the algorithm used (start+count)%N as end, then one-edge
+#   spanning the whole strip would require start=end via wrap, giving count=N.
+#   But the current algorithm: end==start → count=0.  This is the CORRECT behavior
+#   per the Phase 1 spec ("Adjacent taps on the same index → 0-LED edge").
+# ---------------------------------------------------------------------------
+
+
+class TestWholestripSingleEdge:
+    """Verify that a wrap-around edge spanning led_count-3 LEDs (max when 3 others are 1 each) works."""
+
+    def test_one_large_edge_with_minimal_others(self):
+        """One edge has led_count-3 LEDs, the other 3 each have 1 LED.
+
+        bottom_left/clockwise, led_count=100:
+        EDGE_ORDER = [left, top, right, bottom]
+        corners: [0, 97, 98, 99]
+          left: 0→97 = 97
+          top:  97→98 = 1
+          right: 98→99 = 1
+          bottom: 99→0 = 1 (wrap: (100-99)+0 = 1)
+        total = 97+1+1+1 = 100 ✓
+        """
+        cfg = solve_calibration(
+            led_count=100,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[0, 97, 98, 99],
+        )
+        assert cfg.leds_left == 97
+        assert cfg.leds_top == 1
+        assert cfg.leds_right == 1
+        assert cfg.leds_bottom == 1
+        assert (cfg.leds_top + cfg.leds_right + cfg.leds_bottom + cfg.leds_left) == 100
+        _roundtrip(cfg)
+
+    def test_last_edge_wraps_nearly_all_leds(self):
+        """The last edge (bottom in bottom_left/clockwise) wraps from 3 to 0.
+
+        corners: [0, 1, 2, 3]
+          left: 0→1 = 1
+          top:  1→2 = 1
+          right: 2→3 = 1
+          bottom: 3→0 = (100-3)+0 = 97  (wrap-around)
+        total = 100 ✓
+        """
+        cfg = solve_calibration(
+            led_count=100,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[0, 1, 2, 3],
+        )
+        assert cfg.leds_left == 1
+        assert cfg.leds_top == 1
+        assert cfg.leds_right == 1
+        assert cfg.leds_bottom == 97
+        assert (cfg.leds_top + cfg.leds_right + cfg.leds_bottom + cfg.leds_left) == 100
+        _roundtrip(cfg)
+
+
+# ---------------------------------------------------------------------------
+# led_count just above minimum
+# led_count=1: only 1 LED; corner_indices can only meaningfully be [0,0,0,0]
+# which gives all zeros.  That's the all-equal case above.
+# led_count=5 (just above conceptual minimum) with a valid single-wrap partition.
+# ---------------------------------------------------------------------------
+
+
+class TestMinimalLedCount:
+    """Edge cases around very small led_count values."""
+
+    def test_led_count_1_all_equal_indices(self):
+        """led_count=1: the only valid index is 0; all-equal → all-zero edges."""
+        try:
+            cfg = solve_calibration(
+                led_count=1,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[0, 0, 0, 0],
+            )
+            total = cfg.leds_top + cfg.leds_right + cfg.leds_bottom + cfg.leds_left
+            assert total == 0, f"Expected total=0, got {total}"
+            _roundtrip(cfg)
+        except ValueError:
+            pass  # also acceptable
+
+    def test_led_count_4_minimal_partition(self):
+        """led_count=4, each edge gets 1 LED — minimum non-trivial partition."""
+        # bottom_left/clockwise EDGE_ORDER: [left, top, right, bottom]
+        # corners: [0, 1, 2, 3]
+        #   left: 0→1=1, top: 1→2=1, right: 2→3=1, bottom: 3→0=(4-3)+0=1
+        cfg = solve_calibration(
+            led_count=4,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[0, 1, 2, 3],
+        )
+        assert cfg.leds_left == 1
+        assert cfg.leds_top == 1
+        assert cfg.leds_right == 1
+        assert cfg.leds_bottom == 1
+        _roundtrip(cfg)
+
+    def test_led_count_0_raises_valueerror(self):
+        """led_count=0 is explicitly invalid per the docstring."""
+        with pytest.raises(ValueError, match="led_count"):
+            solve_calibration(
+                led_count=0,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[0, 0, 0, 0],
+            )
+
+    def test_led_count_negative_raises_valueerror(self):
+        """led_count=-5 is invalid."""
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=-5,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[0, 0, 0, 0],
+            )
+
+    def test_led_count_5_just_above_minimum(self):
+        """led_count=5 with a valid 2/1/1/1 partition."""
+        # bottom_left/clockwise: [left,top,right,bottom]
+        # corners: [0, 2, 3, 4]
+        #   left: 0→2=2, top: 2→3=1, right: 3→4=1, bottom: 4→0=(5-4)+0=1
+        cfg = solve_calibration(
+            led_count=5,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[0, 2, 3, 4],
+        )
+        assert cfg.leds_left == 2
+        assert cfg.leds_top == 1
+        assert cfg.leds_right == 1
+        assert cfg.leds_bottom == 1
+        assert sum([cfg.leds_top, cfg.leds_right, cfg.leds_bottom, cfg.leds_left]) == 5
+        _roundtrip(cfg)
+
+
+# ---------------------------------------------------------------------------
+# Offset interactions
+# The offset is stored verbatim; PixelMapper normalises it via % total_leds.
+# solve_calibration must pass it through without modification.
+# Also test: offset == 0 (explicit), offset == led_count (should store as-is),
+# and offset >> led_count.
+# ---------------------------------------------------------------------------
+
+
+class TestOffsetInteractions:
+    """Offset is stored verbatim and must not affect per-edge counts."""
+
+    def test_offset_zero_explicit(self):
+        cfg = solve_calibration(
+            led_count=100,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[0, 25, 50, 75],
+            offset=0,
+        )
+        assert cfg.offset == 0
+        _roundtrip(cfg)
+
+    def test_offset_equals_led_count_stored_verbatim(self):
+        """offset=led_count should be stored as-is (not reduced)."""
+        cfg = solve_calibration(
+            led_count=100,
+            start_position="top_left",
+            layout="clockwise",
+            corner_indices=[0, 25, 50, 75],
+            offset=100,
+        )
+        assert cfg.offset == 100
+        _roundtrip(cfg)
+
+    def test_large_offset_stored_verbatim(self):
+        """offset >> led_count — stored verbatim, build_segments must not crash."""
+        cfg = solve_calibration(
+            led_count=60,
+            start_position="top_right",
+            layout="counterclockwise",
+            corner_indices=[0, 15, 30, 45],
+            offset=9999,
+        )
+        assert cfg.offset == 9999
+        _roundtrip(cfg)
+
+    def test_offset_does_not_change_edge_counts(self):
+        """Two calls with different offsets must produce identical edge counts."""
+        kwargs = dict(
+            led_count=100,
+            start_position="bottom_left",
+            layout="counterclockwise",
+            corner_indices=[0, 25, 50, 75],
+        )
+        cfg_no_offset = solve_calibration(**kwargs, offset=0)
+        cfg_offset_13 = solve_calibration(**kwargs, offset=13)
+        assert cfg_no_offset.leds_top == cfg_offset_13.leds_top
+        assert cfg_no_offset.leds_right == cfg_offset_13.leds_right
+        assert cfg_no_offset.leds_bottom == cfg_offset_13.leds_bottom
+        assert cfg_no_offset.leds_left == cfg_offset_13.leds_left
+
+
+# ---------------------------------------------------------------------------
+# corner_indices >= led_count (modulo reduction)
+# The solver applies % led_count to each index before computing counts.
+# Index 150 for led_count=100 should behave identically to index 50.
+# ---------------------------------------------------------------------------
+
+
+class TestCornerIndicesModuloReduction:
+    """Indices >= led_count are reduced modulo led_count before use."""
+
+    def test_indices_above_led_count_reduced(self):
+        """corner_indices [0, 25, 50, 75] and [100, 125, 150, 175] must give same result."""
+        led_count = 100
+        cfg_base = solve_calibration(
+            led_count=led_count,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[0, 25, 50, 75],
+        )
+        cfg_shifted = solve_calibration(
+            led_count=led_count,
+            start_position="bottom_left",
+            layout="clockwise",
+            corner_indices=[100, 125, 150, 175],  # each + 100 (≡ same mod 100)
+        )
+        assert cfg_base.leds_left == cfg_shifted.leds_left
+        assert cfg_base.leds_top == cfg_shifted.leds_top
+        assert cfg_base.leds_right == cfg_shifted.leds_right
+        assert cfg_base.leds_bottom == cfg_shifted.leds_bottom
+        _roundtrip(cfg_shifted)
+
+    def test_index_exactly_led_count_is_zero(self):
+        """Index = led_count reduces to 0, same as explicit 0."""
+        cfg_zero = solve_calibration(
+            led_count=100,
+            start_position="top_left",
+            layout="clockwise",
+            corner_indices=[0, 25, 50, 75],
+        )
+        cfg_hundred = solve_calibration(
+            led_count=100,
+            start_position="top_left",
+            layout="clockwise",
+            corner_indices=[100, 25, 50, 75],  # first index = led_count → reduces to 0
+        )
+        assert cfg_zero.leds_top == cfg_hundred.leds_top
+        assert cfg_zero.leds_right == cfg_hundred.leds_right
+        assert cfg_zero.leds_bottom == cfg_hundred.leds_bottom
+        assert cfg_zero.leds_left == cfg_hundred.leds_left
+
+
+# ---------------------------------------------------------------------------
+# Invalid input: wrong number of corner_indices
+# ---------------------------------------------------------------------------
+
+
+class TestInvalidCornerIndicesLength:
+    """Wrong number of corner indices must raise ValueError."""
+
+    def test_three_corners_raises(self):
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=100,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[0, 25, 50],
+            )
+
+    def test_five_corners_raises(self):
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=100,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[0, 20, 40, 60, 80],
+            )
+
+    def test_empty_corners_raises(self):
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=100,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[],
+            )
+
+    def test_one_corner_raises(self):
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=100,
+                start_position="bottom_left",
+                layout="clockwise",
+                corner_indices=[50],
+            )
+
+
+# ---------------------------------------------------------------------------
+# Invalid start_position / layout
+# ---------------------------------------------------------------------------
+
+
+class TestInvalidEnumInputs:
+    def test_invalid_start_position_raises(self):
+        with pytest.raises(ValueError, match="start_position"):
+            solve_calibration(
+                led_count=100,
+                start_position="center",
+                layout="clockwise",
+                corner_indices=[0, 25, 50, 75],
+            )
+
+    def test_invalid_layout_raises(self):
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=100,
+                start_position="bottom_left",
+                layout="diagonal",
+                corner_indices=[0, 25, 50, 75],
+            )
+
+    def test_both_invalid_raises(self):
+        with pytest.raises(ValueError):
+            solve_calibration(
+                led_count=100,
+                start_position="wrong",
+                layout="wrong",
+                corner_indices=[0, 25, 50, 75],
+            )
+
+
+# ---------------------------------------------------------------------------
+# Totals preservation invariant
+# For any valid input, sum(edge_counts) == sum via build_segments()
+# Test across all 8 combinations with a wrap-around partition.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.parametrize("start_position,layout", list(EDGE_ORDER.keys()))
+def test_total_preservation_with_wraparound(start_position, layout):
+    """For all 8 combinations with a wrap-around partition, total preserved."""
+    # The wrap-around partition: first 3 edges get 20 LEDs each, last edge
+    # gets the remaining 40 via wrap.
+    # EDGE_ORDER tells us walk order; corners: [0, 20, 40, 60] — last edge wraps to 40.
+    # bottom: 60→0 = (100-60)+0 = 40.
+    cfg = solve_calibration(
+        led_count=100,
+        start_position=start_position,
+        layout=layout,
+        corner_indices=[0, 20, 40, 60],
+    )
+    # Each of first 3 edges = 20, last = 40
+    total = cfg.leds_top + cfg.leds_right + cfg.leds_bottom + cfg.leds_left
+    assert total == 100, f"{start_position}/{layout}: expected total=100, got {total}"
+    _roundtrip(cfg)