[Robolightning]

After avcodec_open2 rejects an invalid audio configuration, FFmpeg clears AVCodecContext.codec during cleanup. PyAV subsequently reads that pointer for codec identity and retry, so a second encode/open can crash the interpreter instead of raising an exception.

Keep the selected const AVCodec pointer independently of the mutable FFmpeg context and use it for open, static codec metadata/capabilities, parser selection and profile guards. The selected implementation is retained, including libopus, rather than looking up a default codec by ID.

The subprocess regression performs three actual AAC open failures, checks codec identity, restores the full valid audio configuration (FFmpeg also clears ch_layout), then encodes, muxes and decodes on the same context. A native crash or hang fails the parent test.

Validation: clean CPython 3.15t Windows/Linux builds of all extensions against the official FFmpeg 9.0.2-2 shared SDK; canonical suite 543 passed, 31 SDK capability skips on each platform, plus real libopus/Ogg retries and recovery, every native module import and lossless encoding across native threads with natural GIL OFF. Ruff, configured isort and mypy passed. These local native results apply to the matching runtime-source change on the StableKite branch; this PR preserves newer upstream changes including boolean property results and GIL release when freeing a codec context. Upstream CI is the separate validation of this PR on current master.

[WyattBlue]

Thanks for digging into this — the diagnosis here is spot on. I've opened #2506, which fixes the same root cause with a narrower approach: instead of a shadow _codec field read at each call site, it captures the codec pointer in open() and restores ptr.codec on failure, so every existing reader (codec, name, is_encoder, parse, _decode, …) stays correct without new guards. AVCodec structs are static globals, so holding the pointer is safe.

Since #2506 supersedes this, I'll close it in its favor — but credit to you for the root-cause analysis and repro. Happy to hear if you spot anything the pointer-restore approach misses.