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/static/js/app.js

@@ -180,8 +180,10 @@ window.addEventListener('DOMContentLoaded', async () => {
         const frag = document.createDocumentFragment();
         for (let i = 0; i < SPECTRUM_BARS; i++) {
             const s = document.createElement('span');
-            // Pseudo-random heights for the synthetic CSS animation phase
-            s.style.setProperty('--bar-h', (25 + Math.abs(Math.sin(i * 0.7)) * 70).toFixed(0) + '%');
+            // Pseudo-random initial scaleY for the synthetic CSS-only
+            // animation (used while no real audio is flowing).
+            const scale = (0.25 + Math.abs(Math.sin(i * 0.7)) * 0.70).toFixed(2);
+            s.style.setProperty('--bar-h-scale', scale);
             s.style.setProperty('--bar-delay', (-Math.random() * 1.1).toFixed(2) + 's');
             frag.appendChild(s);
         }