[GavinDarkglider]

Summary

Three changes to the built-in media player (cores/libretro-ffmpeg):

1. Gaps in the video track no longer freeze RetroArch. A recording that dropped frames, or any variable-frame-rate file, has stretches with no video. Across a gap of more than about two seconds, the main thread waited for a frame. Meanwhile the decode thread, which only decodes video once audio has caught up, filled the audio queue and blocked too. The core now holds the frame on screen and keeps playing audio until the next frame arrives.

2. Decoder detection instead of first-registered. avcodec_find_decoder() returns whichever decoder the build registered first. A platform's hardware decoder (h264_nvv4l2, v4l2m2m, rkmpp, cuvid, mediacodec, ...) was used only if the build reordered its list, as the L4T FFmpeg patches do. A default decoder that can't work there failed the file even when another one was built in. The core now tries:

   • platform hardware decoders, behind a new option (on by default);
   • the default decoder with the existing hwaccel path (libdav1d still first for AV1, as in Support More Video Files #19583);
   • the other software decoders.

   It keeps the first one that decodes a frame this core can use.

3. More formats, only when the build can open them. It adds 31 extensions (.ogv, .opus, .mka, .aac, .ac3, .aiff, .wv, .ape, .mts, .ivf, ...), each claimed only if this FFmpeg build has the demuxer and, where the extension implies a codec, the decoder. RetroArch's platforms ship very different FFmpeg builds, so this is checked at runtime.

Commits and the issues they address

#	Commit	Issues
1	ffmpeg core: don't hang on gaps in the video track	no existing report; found playing back recordings with dropped frames (the companion recording PR)
2	ffmpeg core: detect working decoders instead of taking the first	#18702 (generalises the AV1 fallback from #19583 to every codec and platform); hardware decoders without patching FFmpeg's registration order
3	ffmpeg core: claim more media formats, but only those the build can open	.ogv and others previously failed with "path is not set"

Testing

• Gap hang regression test. samples/cores/ffmpeg_sweep gains a clip with a 4 s hole in its video track. It checks that audio and video both carry on past the hole.
  • On master the test hangs until the watchdog kills it.
  • With this PR, all 7.5 s of audio play and 30 new frames follow the gap.
• Gap hang in a full RetroArch build (gl, Xvfb), playing a 12 s file with a 5 s hole:
  • master: froze at the hole and had to be killed;
  • this PR: played through, and the 600-frame run exited normally after 11 s.
• Decoder detection on x86 with FFmpeg 6.1 and platform hardware decoders enabled:
  • h264 tries h264_v4l2m2m and h264_cuvid, both fail to open, then plays with h264;
  • mpeg4 does the same through mpeg4_v4l2m2m and mpeg4_cuvid;
  • AV1 plays with libdav1d.
  • Not yet run on a Switch: there it should pick h264_nvv4l2 by detection, without the reordered decoder list. I'll confirm before merge.
• CI lanes (Linux-samples-cores) all pass locally:
  • ffmpeg_sweep (plain, ASan + UBSan, TSan);
  • the standalone core link;
  • ffmpeg_finished_wait (ASan, TSan);
  • ffmpeg_audio_pipe (ASan, TSan).
• C89: the core stays C89. It has no new library dependencies.

Notes

• New core option: ffmpeg_hw_wrappers ("Use platform hardware decoders", default enabled). It applies to video only; the existing ffmpeg_hw_decoder (hwaccel) option is unchanged.
• Probing decodes from the start until the first frame, then rewinds. Subtitles, unseekable inputs and codecs with a single decoder take the first decoder that opens, as before.