Reduce chromaprint peak-scan cost without changing fingerprint output

Constraint: the live FMA smoke is still running, so the optimization had to preserve existing hash semantics rather than adopt the faster non-equivalent peak picker
Rejected: maximum_filter-based peak picking | changed peak/hash outputs despite much larger speedup
Confidence: high
Scope-risk: narrow
Directive: Keep future chromaprint optimizations hash-equivalent unless evaluation baselines are intentionally regenerated
Tested: compared old vs new peaks and hashes on fma_00000.mp3, measured 2.02x speedup, py_compile passed, rechecked live FMA smoke still in build-index
Not-tested: full build-index completion on the live 8000-reference FMA run has not finished yet

@@ -9,6 +9,7 @@ Implements landmark-based audio fingerprinting:
 
 import numpy as np
 import librosa
+from numpy.lib.stride_tricks import sliding_window_view
 from collections import defaultdict
 from typing import Dict, List, Tuple, Optional
 import pickle
@@ -53,13 +54,18 @@ class ChromaprintMatcher:
         return S
 
     def _find_peaks(self, S: np.ndarray) -> List[Tuple[int, int, float]]:
-        peaks = []
-        for t in range(0, S.shape[1] - self.peak_neighborhood):
-            for f in range(0, S.shape[0] - self.peak_neighborhood):
-                region = S[f:f + self.peak_neighborhood, t:t + self.peak_neighborhood]
-                center = S[f, t]
-                if center == np.max(region) and center > self.min_peak_energy:
-                    peaks.append((t, f, center))
+        if S.shape[0] <= self.peak_neighborhood or S.shape[1] <= self.peak_neighborhood:
+            return []
+
+        windows = sliding_window_view(S, (self.peak_neighborhood, self.peak_neighborhood))[:-1, :-1]
+        region_max = windows.max(axis=(-1, -2))
+        centers = S[: S.shape[0] - self.peak_neighborhood, : S.shape[1] - self.peak_neighborhood]
+        mask = (centers == region_max) & (centers > self.min_peak_energy)
+
+        peaks = [
+            (int(t), int(f), float(centers[f, t]))
+            for f, t in np.argwhere(mask)
+        ]
         peaks.sort(key=lambda x: x[2], reverse=True)
         return peaks[:200]
 

+## 2026-06-02 chromaprint peak scan exact-safe optimization checkpoint
+
+完成项：
+- 将 `acr-engine/src/engines/chromaprint_matcher.py` 的 `_find_peaks()` 从 Python 双层循环改为 `sliding_window_view` 向量化窗口最大值实现。
+- 放弃了更激进但会改变 hash 结果的 `maximum_filter` 方案，改用严格保持旧语义的等价版本。
+
+验证结果：
+- 单曲样本 `/tmp/fma_real_smoke_stopcheck/fma/audio/fma_00000.mp3`：
+  - `old_sec=1.3305`
+  - `new_sec=0.6579`
+  - `speedup_x=2.02`
+- 等价性验证：
+  - `same_all200=true`
+  - `same_hashes=true`
+  - `old_hashes=609`
+  - `new_hashes=609`
+- `python -m py_compile src/engines/chromaprint_matcher.py` 通过
+- 长跑中的真实 FMA smoke 在 `2026-06-02 13:49:25 UTC` 仍处于 `build-index`，尚未产出 `chromaprint.pkl` / `reference_progress.json`
+
+结论：
+- 这次提交安全降低了 chromaprint 峰值扫描热点成本，且不改变当前 fingerprint/hash 行为。
+- 后续可继续观察真实 FMA 全量 smoke 是否更快进入 `chromaprint.pkl` 或 embedding checkpoint。
+
 ## 2026-06-02 真实 FMA smoke build-index 13:36 UTC delivery checkpoint
 
 完成项：