perf(visualizer): cut spectrum + track-switch CPU significantly

Frontend hot path (player.js, background.js):
- visualizer rAF: drop per-frame getComputedStyle('--accent') (cached on
  applyAccentColor), build canvas LinearGradient once per accent change
  instead of 32× per frame, batch all bars into a single beginPath/fill
- FPS-gate canvas redraw via frequencyDataVersion so 60-144 Hz monitors
  stop re-rendering identical frames produced at 30 Hz on the backend
- editorial spectrum bars: replace style.height (layout) with
  transform: scaleY (compositor-only); cache bar refs, pre-compute
  per-bar gain/range, dedup writes at 1/1000 quantization
- coalesce VU needle into the visualizer rAF; cache vuNeedle ref;
  dedup angle writes at 0.1°
- updateUI: status-payload fingerprint short-circuits the redundant
  status_update broadcasts that fire during a track change
- swapArtworkSrc: only force layout reflow when keyframe is in flight;
  drop the ?_=Date.now() cache-buster so identical artwork URLs reuse
  the decoded bitmap; mini/glow imgs only re-set src when changed
- drop the fullscreen MutationObserver — fs-bloom-art is mirrored
  directly from the artwork-swap path, eliminating the second blur paint
- updateProgress: skip text writes when the rounded second hasn't moved;
  POSITION_INTERPOLATION_MS 100 → 250
- background.js: lift resizeBackgroundCanvas out of the rAF body, cache
  step, accept new int-scaled wire format

CSS:
- spectrum bars use transform: scaleY(var(--bar-h-scale)) + transition
  on transform; will-change updated to transform
- album-art-glow and fs-bloom-art switched to small-source-blur trick
  (render at 20-25% size, scale 4-6×, lower blur radius) — visually
  equivalent, ~10-25× cheaper repaint on track change
- drop unused transition: filter on .vinyl-stage #album-art

Backend (audio_analyzer.py, websocket_manager.py):
- pre-allocate windowed and cumsum buffers; replace
  np.concatenate(([0.0], np.cumsum(...))) with cumsum[0]=0 +
  np.cumsum(out=cumsum[1:]); float32 hanning window
- RMS via np.dot(mono, mono) — no astype copy, no ** temp
- int16 wire format (scale=1000) — smaller JSON, no Python float boxing
- versioned data + threading.Event so _audio_broadcast_loop is event-
  driven (ev.wait + monotonic seq dedup) instead of polling on a timer
  with the always-false `data is _last_data` identity check

ruff clean, pytest 7 passed / 3 numpy-skipped, esbuild bundle 113.6 kB.

media_server/services/audio_analyzer.py

@@ -1,6 +1,7 @@
 """Audio spectrum analyzer service using system loopback capture."""
 
 import logging
+import math
 import platform
 import threading
 import time
@@ -71,6 +72,14 @@ class AudioAnalyzer:
         self._lifecycle_lock = threading.Lock()
         self._data: dict | None = None
         self._current_device_name: str | None = None
+        # Generation counter — bumped each time _data is refreshed.
+        # Lets the broadcast loop dedupe without comparing dict identity
+        # (which is fragile because we always allocate a new dict).
+        self._data_seq = 0
+        # Threading.Event signaled when new frame data is available.
+        # The broadcast loop awaits this instead of polling on a timer,
+        # so it wakes up exactly once per produced frame.
+        self._data_event = threading.Event()
         # Slow AGC envelope so the spectrum reflects real dynamics
         # instead of being renormalized to peak=1.0 every frame.
         # A loud transient (e.g. notification beep) lifts the reference
@@ -128,17 +137,30 @@ class AudioAnalyzer:
         """Stop audio capture and cleanup."""
         with self._lifecycle_lock:
             self._running = False
+            # Wake any waiter so it can observe _running and exit cleanly.
+            self._data_event.set()
             if self._thread:
                 self._thread.join(timeout=3.0)
                 self._thread = None
             with self._lock:
                 self._data = None
+            self._data_event.clear()
 
     def get_frequency_data(self) -> dict | None:
         """Return latest frequency data (thread-safe). None if not running."""
         with self._lock:
             return self._data
 
+    def get_frequency_data_versioned(self) -> tuple[dict | None, int]:
+        """Return (data, seq) so callers can dedupe without identity tricks."""
+        with self._lock:
+            return self._data, self._data_seq
+
+    @property
+    def data_event(self) -> threading.Event:
+        """Event signaled when a fresh frame is ready. Caller must clear()."""
+        return self._data_event
+
     @staticmethod
     def list_loopback_devices() -> list[dict[str, str]]:
         """List all available loopback audio devices."""
@@ -250,12 +272,24 @@ class AudioAnalyzer:
             return
 
         interval = 1.0 / self.target_fps
-        window = np.hanning(self.chunk_size)
+        # Float32 window — matches soundcard's typical buffer dtype and
+        # halves FFT memory traffic vs. the default float64.
+        window = np.hanning(self.chunk_size).astype(np.float32)
 
         # Pre-compute bin edge pairs for vectorized grouping
         edges = self._bin_edges
         bin_starts = np.array([edges[i] for i in range(self.num_bins)], dtype=np.intp)
         bin_ends = np.array([max(edges[i + 1], edges[i] + 1) for i in range(self.num_bins)], dtype=np.intp)
+        # Counts are constant — compute once.
+        bin_counts = (bin_ends - bin_starts).astype(np.float32)
+
+        # Pre-allocate working buffers so the per-frame allocator churn
+        # on the capture thread (which runs at target_fps Hz, hours on
+        # end) drops to zero copies for these arrays.
+        fft_size = self.chunk_size // 2 + 1
+        windowed = np.empty(self.chunk_size, dtype=np.float32)
+        cumsum = np.empty(fft_size + 1, dtype=np.float32)
+        cumsum[0] = 0.0
 
         try:
             with device.recorder(
@@ -284,21 +318,23 @@ class AudioAnalyzer:
                         time.sleep(interval)
                         continue
 
-                    # Apply window and compute FFT
-                    windowed = mono[:self.chunk_size] * window
+                    # Apply window in-place into the pre-allocated buffer.
+                    np.multiply(mono[:self.chunk_size], window, out=windowed)
                     fft_mag = np.abs(np.fft.rfft(windowed))
 
-                    # Group into logarithmic bins (vectorized via cumsum)
-                    cumsum = np.concatenate(([0.0], np.cumsum(fft_mag)))
-                    counts = bin_ends - bin_starts
-                    bins = (cumsum[bin_ends] - cumsum[bin_starts]) / counts
+                    # Group into logarithmic bins (vectorized via cumsum).
+                    # Write into the pre-allocated [1:] slice so cumsum[0]
+                    # stays 0.0 and we never allocate a new array.
+                    np.cumsum(fft_mag, out=cumsum[1:])
+                    bins = (cumsum[bin_ends] - cumsum[bin_starts]) / bin_counts
 
-                    # True loudness from time-domain RMS, mapped via dB
-                    # so the VU needle reflects actual program level — not
-                    # the per-frame-normalized spectrum.
-                    rms = float(np.sqrt(np.mean(mono.astype(np.float64) ** 2)))
-                    if rms > 1e-6:
-                        db = 20.0 * np.log10(rms)
+                    # True loudness from time-domain RMS via single BLAS
+                    # dot — avoids astype() and ** allocations.
+                    mono32 = mono if mono.dtype == np.float32 else mono.astype(np.float32, copy=False)
+                    energy = float(np.dot(mono32, mono32))
+                    if energy > 1e-12:
+                        rms = (energy / mono32.size) ** 0.5
+                        db = 20.0 * math.log10(rms)
                         # Map -60 dB..-6 dB to 0..1 (typical music range)
                         level = max(0.0, min(1.0, (db + 60.0) / 54.0))
                     else:
@@ -314,18 +350,33 @@ class AudioAnalyzer:
                     else:
                         self._spectrum_ref += (current_peak - self._spectrum_ref) * 0.005
                     ref = max(self._spectrum_ref, 1e-4)
-                    bins = np.clip(bins / ref, 0.0, 1.5)
+                    np.divide(bins, ref, out=bins)
+                    np.clip(bins, 0.0, 1.5, out=bins)
 
                     # Bass energy: average of first 4 bins (~20-200Hz)
                     bass = float(bins[:4].mean()) if self.num_bins >= 4 else 0.0
 
-                    # Round for compact JSON
-                    frequencies = np.round(bins, 3).tolist()
-                    bass = round(bass, 3)
-                    level = round(level, 3)
+                    # Quantize to 0..1000 ints — same wire fidelity as
+                    # 3-decimal floats but smaller GC churn on both ends
+                    # (frontend smooths anyway, so quantization is
+                    # invisible). JSON encodes ints faster than floats.
+                    frequencies = (bins * 1000.0).astype(np.int16).tolist()
+                    bass_i = int(bass * 1000.0)
+                    level_i = int(level * 1000.0)
 
+                    new_data = {
+                        "frequencies": frequencies,
+                        "bass": bass_i,
+                        "level": level_i,
+                        # Wire-format flag: clients that see this know
+                        # values are 0..1000 ints, not 0..1 floats.
+                        "scale": 1000,
+                    }
                     with self._lock:
-                        self._data = {"frequencies": frequencies, "bass": bass, "level": level}
+                        self._data = new_data
+                        self._data_seq += 1
+                    # Wake any broadcast loop waiting on fresh data.
+                    self._data_event.set()
 
                     # Throttle to target FPS
                     elapsed = time.monotonic() - t0

media_server/services/websocket_manager.py

@@ -161,26 +161,48 @@ class ConnectionManager:
             self._audio_task = None
 
     async def _audio_broadcast_loop(self) -> None:
-        """Background loop: read frequency data from analyzer and broadcast to subscribers."""
-        from ..config import settings
-        interval = 1.0 / settings.visualizer_fps
+        """Background loop: read frequency data from analyzer and broadcast to subscribers.
 
-        _last_data = None
+        Event-driven: blocks on the analyzer's data_event so it wakes up
+        exactly once per produced frame, instead of polling on a timer.
+        Backstop sleep applies when capture is idle / has no subscribers.
+        """
+        from ..config import settings
+        idle_interval = 1.0 / max(1, settings.visualizer_fps)
+        # Bounded wait so we still notice subscribe/unsubscribe transitions.
+        wake_timeout = max(0.05, idle_interval)
+        loop = asyncio.get_event_loop()
+
+        last_seq = -1
 
         while True:
             try:
                 async with self._lock:
                     subscribers = list(self._visualizer_subscribers)
 
-                if not subscribers or not self._audio_analyzer or not self._audio_analyzer.running:
-                    await asyncio.sleep(interval)
+                analyzer = self._audio_analyzer
+                if not subscribers or not analyzer or not analyzer.running:
+                    await asyncio.sleep(idle_interval)
                     continue
 
-                data = self._audio_analyzer.get_frequency_data()
-                if data is None or data is _last_data:
-                    await asyncio.sleep(interval)
+                # Wait off-loop for a fresh frame. The capture thread sets
+                # data_event after each FFT update; we clear it before the
+                # next wait so we never burn a wake on stale data.
+                ev = analyzer.data_event
+
+                def _wait() -> bool:
+                    return ev.wait(wake_timeout)
+
+                got = await loop.run_in_executor(None, _wait)
+                if not got:
+                    # Timeout — loop around to re-check subscriber state.
                     continue
-                _last_data = data
+                ev.clear()
+
+                data, seq = analyzer.get_frequency_data_versioned()
+                if data is None or seq == last_seq:
+                    continue
+                last_seq = seq
 
                 # Pre-serialize once for all subscribers (avoids per-client JSON encoding)
                 text = json.dumps({"type": "audio_data", "data": data}, separators=(',', ':'))
@@ -198,13 +220,11 @@ class ConnectionManager:
                 for ws in failed:
                     await self.disconnect(ws)
 
-                await asyncio.sleep(interval)
-
             except asyncio.CancelledError:
                 break
             except Exception as e:
                 logger.error("Error in audio broadcast: %s", e)
-                await asyncio.sleep(interval)
+                await asyncio.sleep(idle_interval)
 
     def status_changed(
         self, old: dict[str, Any] | None, new: dict[str, Any]