feat: use lossless FFV1 and PCM for analysis clips

extract.py

@@ -23,6 +23,22 @@ for input_file in inputs:
 DURATION   = 90  # 1.5 minutes in seconds
 OUT_DIR    = os.path.join(os.path.dirname(os.path.abspath(__file__)), "extracted")
 
+# ─── Analysis-friendly codecs (NOT GPU) ───
+# Video: FFV1 (lossless, intra-only) in MKV. Every frame stands alone, no
+#   P/B-frame prediction, no blocking artifacts -> exact frames for similarity
+#   (SSIM/pHash/features) + fast random access. Decoded with -pix_fmt yuv444p
+#   to avoid 4:2:0 chroma-subsampling noise.
+# Audio: PCM float32 WAV, 48 kHz mono. Bit-exact waveform preserves phase for
+#   cross-correlation / GCC-PHAT. Never MP3/AAC/Opus here (MDCT + padding
+#   smears transients and shifts correlation peaks). FLAC is the only
+#   acceptable compressed alternative (lossless).
+# GPU (RX 7900 XTX via VAAPI/AMF: h264_vaapi/hevc_vaapi/h264_amf): deliberately
+#   NOT used. Those encoders are lossy delivery codecs, lower quality per bit,
+#   inter-frame only, and FFV1/PCM/ProRes/DNxHR have no GPU implementation.
+#   GPU only makes sense for preview proxies, not for measurement. CPU encode
+#   of 90 s clips is fast enough and bit-exact.
+VIDEO_CODEC = ["-c:v", "ffv1", "-level", "3", "-pix_fmt", "yuv444p"]
+AUDIO_CODEC = ["-c:a", "pcm_f32le", "-ar", "48000", "-ac", "1"]
 # ─── Check ffmpeg ───
 if shutil.which(f"ffmpeg") is None:
     sys.exit(f"ERROR: 'ffmpeg' not found on PATH. Install it and try again.")
@@ -34,22 +50,22 @@ for INPUT_FILE in INPUT_FILES:
 
     # ─── Derive output filenames ───
     stem = os.path.splitext(os.path.basename(INPUT_FILE))[0]
-    video_out = os.path.join(OUT_DIR, f"{stem}_video.mkv")
-    audio_out = os.path.join(OUT_DIR, f"{stem}_audio.mp3")
+    video_out = os.path.join(OUT_DIR, f"{stem}_video.mkv")  # FFV1
+    audio_out = os.path.join(OUT_DIR, f"{stem}_audio.wav")  # PCM f32
 
     print(f"Input : {INPUT_FILE}")
     print(f"Video → {video_out}")
     print(f"Audio → {audio_out}")
     print(f"-" * 50)
 
-    # ─── Extract video-only (first 90s, no audio) ───
+    # ─── Extract video-only (first 90s, no audio, CPU FFV1 re-encode) ───
     cmd_video = [
         "ffmpeg",
         "-y",               # overwrite output
         "-i", INPUT_FILE,
         "-t", str(DURATION),
         "-an",              # no audio
-        "-c:v", "copy",     # copy video stream (no re-encode = fast + lossless)
+        *VIDEO_CODEC,       # ffv1 lossless intra-only, CPU (no GPU equiv.)
         video_out
     ]
     print(f"[video: {stem}] running ffmpeg…")
@@ -58,15 +74,14 @@ for INPUT_FILE in INPUT_FILES:
         sys.exit(f"ffmpeg (video) failed:\n{r.stderr}")
     print(f"[video: {stem}] done ✔")
 
-    # ─── Extract audio-only (first 90s, no video) ───
+    # ─── Extract audio-only (first 90s, no video, CPU PCM re-encode) ───
     cmd_audio = [
         "ffmpeg",
         "-y",
         "-i", INPUT_FILE,
         "-t", str(DURATION),
         "-vn",              # no video
-        "-c:a", "libmp3lame",
-        "-b:a", "192k",     # 192 kbps
+        *AUDIO_CODEC,       # pcm_f32le 48kHz mono, CPU (no GPU equiv.)
         audio_out
     ]
     print(f"[audio: {stem}] running ffmpeg…")