Five surfaces refuse compressed media for the same missing demuxer, and no row owns the codec brick mm-serving.md already proposes

Row: -

Five surfaces across this tree refuse compressed media for the same reason -- nothing here
vendors a demuxer -- and each carries its own named residual. This issue proposes one shared
codec brick to discharge them, and records the evidence so the next reader does not
re-derive it.

The five

.agents/specs/mm-serving.md already proposes the right shape for this in prose: "a small
codec brick." Nothing has owned it.

Why it is not a model row's job

The surface is unbounded. libsndfile 1.2.2 reports 26 container formats, and upstream's
own fallback is PyAV/ffmpeg, which adds hundreds of demuxers
(vllm/multimodal/media/audio.py:216-258 @ 9035151d6: soundfile first, PyAV on
ImportError or a LibsndfileError whose code is in _BAD_SF_CODES).

A verified inventory of what this tree has: zero audio dependencies of any kind. No
libsndfile, no libav*/ffmpeg linkage, no libswresample, no dr_wav, stb_vorbis, minimp3,
libFLAC, opus, ogg, soxr, libsamplerate, speexdsp. third_party/ holds six entries -- blake3,
doctest, httplib, minja, nlohmann/json, Vulkan headers -- none of them audio. ffmpeg appears
only as an external process whose argv the library composes and never spawns
(src/vllm/model_executor/models/minimax_h3_mux.cpp:12).

The decision this needs

It turns on the archive contract's no-vendor rule, which is project policy rather than a
row decision. Note the precedent already exists for a deliberate exception:
docker/Dockerfile:209-215 installs the ffmpeg BINARY as a documented divergence from that
rule.

Options, roughly in ascending cost: vendor a narrow decoder set (dr_wav / minimp3 / stb_vorbis
are single-header and permissively licensed, but that is three formats, not 26); link
libsndfile optionally and refuse by name when absent; keep refusing permanently and document
raw PCM/RGB as the contract.

One record repair that should ride along

.agents/specs/ltx25-retake.md's ## Owed points at issue #924 for "A container demuxer.
Without it the audio half of retake cannot exist" -- and #924 does not resolve. That
tracking number is dangling and needs reconciling whichever way this issue is decided.

Start with .agents/specs/mm-serving.md and compare the five residuals in server_main.cpp, ltx2_audio_input.cpp, ltx25-a2v-audio-input.md, speech_api.cpp, and dots3-note.md. Review the cited audio fallback and Dockerfile precedent before choosing a policy option. Done means the codec brick has an owner and the dangling reference to issue #924 is reconciled.