fix: skip disposable pictures during local seek preroll

This commit is contained in:
2026-10-18 20:11:39 +02:00
parent 2e61bd7864
commit ddf53ea606
5 changed files with 97 additions and 3 deletions
+3 -1
View File
@@ -154,7 +154,7 @@ Settings load once at startup. Menu changes are session-only and never write the
## Limits
- SDR output only; PQ/HLG HDR is rejected. No stream playlists, playlist-file import, embedded/YouTube captions, manual audio-track selection, rotation, or deinterlacing. The session-only local-file queue and external resolved-source WebVTT subtitles are supported.
- H.264, HEVC, and VP9 prefer VideoToolbox decoding with software fallback; other codecs, including AV1, use software. Backward frame stepping can be slow on long-GOP or unindexed media.
- H.264, HEVC, and VP9 prefer VideoToolbox decoding with software fallback; other codecs, including AV1, use software. Local seeks remain frame-accurate: disposable non-reference pictures strictly before the target are skipped, but reference pictures still require decoding. Seeking and backward frame stepping can remain slow on long-GOP or unindexed media.
- Network buffering is bounded and memory-only; live/non-seekable sources disable seeking and frame stepping. Expired stream URLs are not automatically refreshed.
- Audio-device changes during playback are not handled.
@@ -169,6 +169,8 @@ The full suite requires a graphical macOS session and a working audio device; it
The startup checks verify that playing video remains attached, unhidden and sized to its native drag surface, including with the queue open; playback-state checks alone do not establish renderer visibility.
`make check-local-seek` measures paused/playing local seek-to-picture latency and compares the resulting timestamps and pixels against full decoding, including B-frame targets. It covers a 1080p H.264 long-GOP file with audio and the six hardware/software decoder fixtures; timing is diagnostic, not a machine-dependent pass threshold.
Queue-open resize checks require positive native resize increments and exercise continuous horizontal/vertical resize-delegate sequences, including reversals, to preserve video aspect without snap-back. The driving axis is fixed for each live resize gesture so aspect corrections do not feed back into axis selection. Physical edge-drag resizing still needs manual verification.
`make benchmark-playback` measures cold source startup, cached/uncached seeks, and sustained native/Cast playback with repeatable loopback fixtures. It requires the same macOS session/audio device and fixture-generation FFmpeg; results exclude live extraction, WAN/TLS, and physical-receiver latency.