Codebase review: stability, performance, usability, and i18n fixes

Stability:
- Fix race condition: set _is_running before create_task in target processors
- Await probe task after cancel in wled_target_processor
- Replace raw fetch() with fetchWithAuth() across devices, kc-targets, pattern-templates
- Add try/catch to showTestTemplateModal in streams.js
- Wrap blocking I/O in asyncio.to_thread (picture_targets, system restore)
- Fix dashboardStopAll to filter only running targets with ok guard

Performance:
- Vectorize fire effect spark loop with numpy in effect_stream
- Vectorize FFT band binning with cumulative sum in analysis.py
- Rewrite pixel_processor with vectorized numpy (accept ndarray or list)
- Add httpx.AsyncClient connection pooling with lock in wled_provider
- Optimize _send_pixels_http to avoid np.hstack allocation in wled_client
- Mutate chart arrays in-place in dashboard, perf-charts, targets
- Merge dashboard 2-batch fetch into single Promise.all
- Hoist frame_time outside loop in mapped_stream

Usability:
- Fix health check interval load/save in device settings
- Swap confirm modal button classes (No=secondary, Yes=danger)
- Add aria-modal to audio/value source editors, fix close button aria-labels
- Add modal footer close button to settings modal
- Add dedicated calibration LED count validation error keys

i18n:
- Replace ~50 hardcoded English strings with t() calls across 12 JS files
- Add 50 new keys to en.json, ru.json, zh.json
- Localize inline toasts in index.html with window.t fallback
- Add data-i18n to command palette footer
- Add localization policy to CLAUDE.md

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

server/src/wled_controller/core/audio/analysis.py

@@ -99,6 +99,12 @@ class AudioAnalyzer:
         self._spectrum_buf_right = np.zeros(NUM_BANDS, dtype=np.float32)
         self._sq_buf = np.empty(chunk_size, dtype=np.float32)
 
+        # Pre-compute band start/end arrays and widths for vectorized binning
+        self._band_starts = np.array([s for s, _ in self._bands], dtype=np.intp)
+        self._band_ends = np.array([e for _, e in self._bands], dtype=np.intp)
+        self._band_widths = (self._band_ends - self._band_starts).astype(np.float32)
+        self._band_widths[self._band_widths == 0] = 1.0  # avoid divide-by-zero
+
         # Pre-allocated channel buffers for stereo
         self._left_buf = np.empty(chunk_size, dtype=np.float32)
         self._right_buf = np.empty(chunk_size, dtype=np.float32)
@@ -205,11 +211,15 @@ class AudioAnalyzer:
         fft_mag = np.abs(np.fft.rfft(self._fft_windowed))
         fft_mag *= (1.0 / chunk_size)
         fft_len = len(fft_mag)
-        for b, (s, e) in enumerate(self._bands):
-            if s < fft_len and e <= fft_len:
-                buf[b] = float(np.mean(fft_mag[s:e]))
-            else:
-                buf[b] = 0.0
+        # Vectorized band binning using cumulative sum
+        valid = (self._band_starts < fft_len) & (self._band_ends <= fft_len) & (self._band_ends > 0)
+        buf[:] = 0.0
+        if valid.any():
+            cumsum = np.cumsum(fft_mag)
+            band_sums = cumsum[self._band_ends[valid] - 1] - np.where(
+                self._band_starts[valid] > 0, cumsum[self._band_starts[valid] - 1], 0.0
+            )
+            buf[valid] = band_sums / self._band_widths[valid]
         spec_max = float(np.max(buf))
         if spec_max > 1e-6:
             buf *= (1.0 / spec_max)