fix(vu): drive needle from RMS-dB loudness instead of peak-of-bins

- Backend computes time-domain RMS, maps -60..-6 dB to 0..1, sends as
  `level` alongside the per-frame-normalized frequency bins.
- Frontend prefers `level` directly; drops the peak-of-bins fallback
  and the redundant volume-slider attenuation (loopback capture is
  already post-volume on Windows/macOS).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

media_server/services/audio_analyzer.py

@@ -284,7 +284,18 @@ class AudioAnalyzer:
                     counts = bin_ends - bin_starts
                     bins = (cumsum[bin_ends] - cumsum[bin_starts]) / counts
 
-                    # Normalize to 0-1
+                    # 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)
+                        # Map -60 dB..-6 dB to 0..1 (typical music range)
+                        level = max(0.0, min(1.0, (db + 60.0) / 54.0))
+                    else:
+                        level = 0.0
+
+                    # Normalize bins to 0-1 for spectrum display
                     max_val = bins.max()
                     if max_val > 0:
                         bins *= (1.0 / max_val)
@@ -295,9 +306,10 @@ class AudioAnalyzer:
                     # Round for compact JSON
                     frequencies = np.round(bins, 3).tolist()
                     bass = round(bass, 3)
+                    level = round(level, 3)
 
                     with self._lock:
-                        self._data = {"frequencies": frequencies, "bass": bass}
+                        self._data = {"frequencies": frequencies, "bass": bass, "level": level}
 
                     # Throttle to target FPS
                     elapsed = time.monotonic() - t0