MOSS-Music stage extraction: expose valid-frame metadata in stable C ABI

[jajmangold]

Context

The moss_audio backend already exposes encoder, DeepStack tap, and adapter stage helpers in src/moss_audio.h / src/moss_audio.cpp. A downstream MOSS-Music-8B-Thinking feature pipeline needs a stable way to obtain the final 1280-D encoder stream, taps 8/16/24, and the 4096-D adapted stream together with observed valid-frame metadata, without the 3000-frame mel padding being mistaken for real content.

This request is only for a stable source-level C ABI. Downstream HTTP, CAS publishing, authentication, and rights policy remain outside CrispASR.

Acceptance criteria

• Add new exported C symbols without changing existing moss_audio_* signatures. A metadata-returning encoder entrypoint returns the existing tensor and out_T_enc / out_d, plus caller-allocated per-chunk valid-frame counts, out_num_chunks, the pre-pad out_T_mel_actual, and out_total_valid.
• Reuse the existing encoder chunk loop and its computed valid_lens; do not derive valid frame counts from a global floor formula.
• Preserve the existing final 1280-D encoder output and DeepStack taps 8/16/24. Report the adapter output dimension from GGUF metadata/weights rather than hardcoding it.
• Correct the stale header comment: moss_audio.llm.hidden_size is 2560 for MOSS-Audio-4B and 4096 for MOSS-Music-8B-Thinking.
• Fail closed at load time if llm_hidden disagrees with the adapter output-row shape; do not silently accept a 2560 fallback for a 4096-D checkpoint.
• Document ownership and threading: one context has one in-flight stage run; callbacks, if added, run synchronously on the caller thread and must not re-enter the same context; caller owns returned allocations/copies.
• Preserve cached-encoder-graph invalidation/rebuild behavior on the new entrypoint.
• Add a differential test proving the sum of per-chunk valid frames equals out_total_valid and that a sub-30-second input does not expose zero-content frames created only by the 3000-frame mel pad.

Suggested smallest patch

• src/moss_audio.h: declare a metadata-returning encoder API and document dimensions, ownership, and single-inflight behavior.
• src/moss_audio.cpp: expose the existing per-chunk valid_lens, retain the pre-pad mel length, validate adapter/LLM dimensions, and keep existing symbols untouched.

An optional chunk-local callback may follow the existing moss_audio_process_cb pattern, but it is not required if the metadata-returning entrypoint can return final/tap/adapter tensors without duplicate model execution.

[CrispStrobe]

Resolved and validated.

Implementation and ABI changes merged via #349, with the live-test harness fix merged via #351.

Validation:

• Full Release build succeeded.
• Complete unit gate: 1,598/1,598 tests passed.
• Hermetic metadata tests: 220 assertions across 4 cases passed.
• Kaggle GPU/model-backed run against commit 41ef3c3f: 3 live test cases, 40 assertions, return code 0. The MOSS-Audio GGUF loaded successfully (encoder d=1280, adapter hidden=8192, LLM hidden=2560).
• Formatting and diff checks passed.

The initial Kaggle failure was traced to the live test using a relative samples/jfk.wav path while Kaggle runs the executable outside the repository. #351 makes the sample path absolute via CMake.