[inureyes]

Summary

• Port the pinned Phi-4 Multimodal SpeechLib feature extractor, Cascades Conformer encoder, speech/vision audio projections, and complete modality LoRA weights with fail-closed shape and completeness checks.
• Normalize and expand numbered audio/image placeholders, replace the exact media embedding rows, and keep language, speech, vision, and mixed adapter modes request-scoped through decode.
• Wire audio-only and mixed media through the CLI and OpenAI-compatible streaming server, including multiple server audio clips, bounded streaming-WAV sentinel handling, and the correct dense external KV layout for single-slot Phi4MM serving.
• Add pinned revision documentation, deterministic unit fixtures, and a gated real-checkpoint CUDA parity test for features, encoder/projection values, merged prefill, KV state, and greedy tokens.

Validation

• cargo fmt --all -- --check
• cargo check --all-targets
• cargo clippy -p mlxcel-core --lib --tests -- -D warnings
• cargo clippy -p mlxcel --lib --tests -- -D warnings
• Focused LoRA, streaming WAV, scheduler gate, Phi3, and Phi4MM tests: 48 passed
• Release CUDA parity against microsoft/Phi-4-multimodal-instruct revision 93f923e1a7727d1c4f446756212d9d3e8fcc5d81: feature/intermediate/KV checks passed and greedy tokens matched [13347, 382] (what is)
• CUDA CLI smoke: audio-only what is, image-only A solid, mixed image/audio A solid; prompt lengths matched 53/553/599
• OpenAI-compatible streaming server smoke: default warmup passed, audio streamed what is, and mixed media streamed A solid

Closes #874