[inureyes]

Summary

• emit the pinned Phi4MM cascaded Conformer and speech/vision projection as a static IREE audio.main module
• implement complete per-row language/speech/vision LoRA selection in prefill and decode without global mutable adapter state
• validate complete LoRA A/B pairs, rank, scale, shapes, projection lengths, placeholders, and artifact identity before native execution
• preserve adapter mode through ragged batching, cancellation, and slot reuse
• integrate bounded audio-only, image-only, and approved mixed image/audio preparation into the shared CLI and streaming-server XLA path
• keep only the filtered text embedding and qualified vision producer on MLX; the full decoder and audio encoder remain IREE-owned
• fix the production audio-only CLI path so an empty image list does not enter image decoding

Validation

• cargo fmt --all -- --check
• git diff --check
• cargo check --features xla-iree --lib --tests
• cargo clippy --features xla-iree --lib --tests --no-deps
• Phi4MM XLA focused tests: 12 passed, 5 diagnostic tests intentionally ignored
• root Phi4MM focused tests: 32 passed
• mixed scheduler admission covers simultaneous audio-only, image-only, and image+audio requests with Speech/Vision/Vision per-slot modes and one-token completion
• cancellation and slot-reuse tests prove adapter state is cleared and cannot leak
• official revision 93f923e1a7727d1c4f446756212d9d3e8fcc5d81: IREE projection produced 44 rows and passed the feat(vlm): add a token-exact Phi4MM audio reference path #874 eight-probe gate at max abs 0.01486496 < 0.025
• actual production XLA CLI with the official WAV generated token IDs [13347, 382] (what is), exactly matching the pinned oracle
• artifact-hardening tests: 9 passed, 4 environment-only diagnostics ignored; the cache identity includes the compiler executable's raw SHA-256 and decoded audio weights reject NaN/Inf before native upload

Notes

• full-vector projection diagnostics retain the documented reduction outliers, while the qualified feat(vlm): add a token-exact Phi4MM audio reference path #874 probes and final token gate pass
• the branch was rebased onto PR feat(xla): execute LLaVA vision through IREE #913 and now uses its concurrent, crash-recoverable auxiliary artifact cache
• the production mixed request prepared one image plus the official 351-frame/44-token WAV into 592 prompt tokens; its local-task CLI generation was stopped at the 5-minute bound, and the CUDA binary could not be linked on this AArch64 host because GNU ld overflowed R_AARCH64_CALL26. Mixed correctness is therefore covered compositionally by the feat(vlm): add a token-exact Phi4MM audio reference path #874 pinned MLX mixed oracle (A solid), feat(xla): execute LLaVA vision through IREE #913's production IREE vision gate, the production IREE audio token gate above, and the mixed per-slot scheduler regression rather than by claiming an unavailable direct mixed token run.

Closes #877