feat(xla): add bounded audio preprocessing and serving plumbing

[inureyes]

Parent and dependency

• Parent epic: epic: OpenXLA multimodal / VLM architecture support (Qwen2-VL, Gemma3n, Phi4MM, Molmo, and more) #566
• Blocked by feat(xla): admit image requests in CLI and continuous-batch serving #861
• Existing XLA rejection point: src/server/batch/xla_worker.rs
• Existing audio request/decode surfaces: src/commands/generate.rs, src/commands/generate_vlm.rs, src/server/model_worker.rs, src/server/chat_template.rs, and src/audio/

Problem

The image prepared-prefill contract is not sufficient for audio ingestion. Audio arrives as encoded bytes/URLs or waveform data, requires bounded decode/downmix/resample and family-specific feature extraction, has duration/frame masks rather than pixel grids, and can expand to a model-specific number of placeholder tokens. The XLA worker currently rejects audio before any of this state exists.

Family issues must not each invent their own unbounded decoding, request cancellation, sample-rate, or accounting policy.

Goal

Add a backend-neutral, bounded audio preprocessing and serving seam that family-specific Phi4MM and Gemma3n preprocessors can use to produce the existing owned prepared-prefill contract.

Canonical request contract

• Represent each decoded clip as owned mono f32 samples plus explicit sample rate, source duration, and source metadata needed for diagnostics. Samples must be finite and normalized to [-1, 1] after conversion.
• Keep encoded bytes and decoder handles out of mlxcel-xla and out of the prepared-prefill DTO.
• Define a family preprocessor interface that consumes decoded clips plus logical prompt tokens and returns family-specific feature/mask tensors and finally an owned PreparedMultimodalPrefill.
• Preserve per-clip boundaries, ordering, placeholder mapping, effective expanded token count, and internal audio-frame/token counts. Do not flatten multiple clips without lengths.

Required implementation

• Reuse the existing CLI/server media acquisition, URL/data handling, byte limits, and audio decoder. Centralize downmix/resample/normalization so MLX and XLA families receive identical waveforms.
• Make target sample rate, channel policy, maximum duration, feature dtype, and placeholder expansion model configuration, not global guesses. The first families use their pinned official processor settings.
• Perform decode/resample/feature extraction in a bounded cancellable stage outside the XLA scheduler loop. Bound concurrent jobs, encoded bytes, decoded samples, total duration, feature frames, clips per request, and queued host memory.
• Reject NaN/Inf samples, unsupported codecs, empty audio, excessive duration, arithmetic overflow, placeholder mismatch, and post-expansion context overflow with per-request typed errors.
• Extend CLI and OpenAI-compatible request preparation to preserve audio parts in prompt order. Do not move placeholders outside their rendered user turn.
• Carry cancellation through acquisition, decode, feature extraction, and queue submission. Large buffers must be released exactly once on every exit.
• Add separate metrics for source seconds, decoded samples, feature frames, effective prefill tokens, preprocessing latency, and rejection reason. Preserve public token-usage semantics.
• Keep XLA multimodal/audio capability false until a loaded family preprocessor and runtime bundle complete the path; foundation availability alone is not capability.
• Continue explicit rejection for models/backends without audio support and for video unless separately implemented.

Non-goals

• Implementing a Phi4MM or Gemma3n audio feature extractor/encoder.
• Selecting one universal duration limit for all checkpoints.
• Streaming audio ingestion or speech output.

Validation

• Unit-test mono/stereo downmix, sample-rate conversion, integer/float normalization, empty/corrupt/nonfinite input, byte/duration/frame/count limits, multi-clip ordering, and placeholder mismatch.
• Add deterministic resampling golden vectors so CLI and server produce byte/numerically identical waveforms.
• Test cancellation at download, decode, resample, feature extraction, and pending-queue stages; verify host-buffer release and worker health.
• Stress bounded concurrency with oversized requests beside valid text/image requests and prove scheduler progress and memory limits.
• Add capability tests proving no request can be accepted and then have audio silently dropped.

Acceptance criteria

• CLI/server share one bounded waveform normalization path and identical model inputs.
• Family-specific sample rate/duration/feature policy is explicit configuration.
• Multi-clip boundaries, prompt order, and placeholder mapping survive to the preprocessor.
• Decode/preprocessing is bounded, cancellable, observable, and outside the scheduler loop.
• All failure paths release buffers and affect only the failed request.
• Public capability stays false until a complete family implementation is loaded.
• Existing text/image request behavior is unchanged.