Replay from a key frame after VideoToolbox send fallback

[Blackspirits]

Summary

The new VideoToolbox decoder fallback handles two classes of hardware failure differently.

When avcodec_receive_frame or av_hwframe_transfer_data fails, the player reopens the decoder in software and calls Seek(Position), which gives the demuxer a new serial, seeks backward, flushes packet/frame queues and replays from a key frame.

When avcodec_send_packet fails, however, the code reopens the decoder in software, sets only the video loop's local serial = -1, frees the rejected packet and continues with the next already-demuxed packet. A fresh software decoder can therefore start in the middle of a GOP without the reference frames required by that packet.

This change makes the send-packet failure path use the same recovery contract as the other VideoToolbox failures:

• reopen in software;
• invalidate the local decoder serial;
• call Seek(Position) so demux replays from a key frame.

Validation

• based on upstream f67b07f5b143fa47c87255ca1c3ed7adb3bd7142
• diff is limited to FfmpegPlayer.cs
• control flow verified against PerformSeek: it calls av_seek_frame(..., AVSEEK_FLAG_BACKWARD), advances the serial and flushes video/audio/frame queues
• the existing receive/transfer fallback already uses the same replay sequence
• macOS VideoToolbox failure injection is not available in the Linux CI environment
• full solution restore passed
• full solution build passed
• full solution test suite passed on the first run
• UI suite: 5,150 passed, 9 skipped, 0 failed (5,159 total)
• retry step was not used
• validation run: #34771533047

AI assistance: ChatGPT was used to adversarially review the newly merged VideoToolbox fallback paths and identify the missing demux replay after a hardware avcodec_send_packet failure.

[Blackspirits]

Independent adversarial re-check: the send-packet hardware-failure path had diverged from the receive/transfer fallback after the rejected packet was already freed. The final diff restores the same replay contract by seeking from the current position after reopening software decoding. Full CI #34771533047 passed restore, build and the complete suite on the first run; UITests reported 5,150 passed, 9 skipped and 0 failed. macOS VideoToolbox failure injection is not exercised by Linux CI, but the control-flow defect is deterministic from the packet lifetime and existing fallback contract. No blocker identified. Keep draft; no merge performed.

[Blackspirits]

Superseded by current-base draft PR #93. Upstream generalized the hardware path to D3D11VA/DXVA2 after this draft; #93 reapplies the same reviewed replay fix on c77c4b4 and has full green CI. Closing the older VideoToolbox-base draft to avoid competing fixes; no merge performed.