native frontend: mount resample and mel stages cooperatively on Flashkern

[sydneyrenee]

Problem

AUDIO_PHASE_FRONTEND currently executes the entire resampler/frontend synchronously on lane zero. run_audio_program_stage returns immediately for every other lane, then calls prepare_audio_conformer, which runs resampling, frame geometry, and the full mel frontend serially.

Paired leaves exist, and the workspace aliases some dead planes correctly, but the execution is not a cooperative Flashkern program. Additional violations:

• non-Apple sgemm_rm is a scalar C++ triple loop;
• mounted frontend assembly calls per-element logf;
• DFT/power and mel/stat planes materialize more broadly than the register-FIFO contract requires.

Required stage graph

Mount validated generations for:

1. polyphase resampling over output ranges;
2. preemphasis/window/gather over frame bands;
3. DFT through an explicit Accelerate/AMX seam on Apple or supported architecture leaf;
4. power conversion, aliasing dead STFT-real storage where valid;
5. mel projection through an explicit dense seam;
6. log guard;
7. per-feature statistics and ddof=1 normalization;
8. exact BF16 publication directly into the Conformer consumer layout.

Lanes claim disjoint frame/bin tiles. Lane-zero work is limited to bounded phase transitions and plan-state publication.

Materialization law

A plane is allowed only when data must survive:

• a team barrier;
• DFT/mel opaque matrix seam;
• fan-out to multiple later consumers;
• continuation suspension;
• published Conformer input.

A C++ function or reference-module boundary is not sufficient. Scratch aliases immediately after the last consumer.

Work

• Split AUDIO_PHASE_FRONTEND into retained program phases.
• Bind all immutable FFT/window/mel tables at plan construction.
• Remove scalar product fallback; unsupported ISA fails readiness.
• Replace per-element libm calls with faithful paired SIMD/assembly math under frozen traces.
• Write normalized valid rows directly in the layout Conformer consumes.
• Preserve the implemented padding, frame-count, normalization, log-guard, and rounding semantics. Do not invent a dither stage: this frontend has no production dither source.

Acceptance

• Frontend uses all eligible lanes for representative utterances.
• No production scalar matrix fallback or hot libm call remains.
• No payload transpose/crop copy exists between frontend and Conformer.
• Plane lifetimes and legal barriers are documented and asserted.
• Frozen resample/mel traces pass across tails, rates, silence, padding, normalization, and rounding.
• AArch64 and x86_64/Rosetta pass.
• Real-checkpoint first and retained-context audio turns pass.

Related

• Numerical umbrella: native inference: remove residual C++ numerical fallbacks and materialized scratch #158
• Faithfulness traces: child of native inference: remove residual C++ numerical fallbacks and materialized scratch #158
• Conformer cooperative program: separate child issue

[sydneyrenee]

Structure verified: AUDIO_PHASE_FRONTEND runs entirely on lane 0 — run_audio_program_stage returns for every other lane (flashkern_engine.cpp:3888-3894), then prepare_audio_conformer executes resample → frame geometry → full mel synchronously. The 8 paired .S leaves are real but invoked serially inside that call.

Two precision notes:

1. The scalar fallback is doc-admitted debt, so the docs must move with the code. spec 05:150-156 explicitly admits "Accelerate SGEMM on Apple and the ordered scalar reference elsewhere" for the parity tranche. sgemm_rm (lfm_frontend.cpp:111-118) is that admitted reference. Removing it is correct per the final target and ABI rule 8, but the same tranche must amend spec 05, or the doc will keep sanctioning what the code no longer provides (overlaps docs: reconcile native PCM, Sesame, kcoro, and numerical ownership after the Rust cutover #161).

2. "Preserve exact dither" is confusing — this frontend has no dither: the ladder is resample → pad/preemphasis → windowed DFT → power → Slaney mel → log guard → ddof=1 normalize → RNE BF16. Nothing stochastic exists to preserve. Likely a paste error; suggest deleting it from Acceptance before an implementer goes hunting for a noise source that was never there.

— K3

[sydneyrenee]

Implementation handoff — frontend is a cooperative signal program, not one callback

The frontend is currently native but mounted as one lane-zero call. Do not confuse “uses assembly leaves” with “runs as Flashkern.” The work is to expose its real stage lifetimes to the team without changing the trained preprocessing contract.

Important deviations from common audio pipelines

• There is no production dither source in this implementation. Do not add one because another model config mentions dither.
• The Conformer consumes the frontend’s exact normalized BF16 layout and valid length. Do not publish a convenient tensor and transpose later.
• DFT and mel GEMM are legitimate opaque dense seams. Materialization around them is allowed only where their input/output contract requires it.
• Silence and padding are data semantics, not scheduler timing.
• No scalar portability fallback: unsupported ISA/platform fails readiness.
• The callback is not the place to run this work; it publishes PCM and resumes the retained route.

Before coding

1. Land/verify native inference: restore authoritative full-value faithfulness gates before FIFO fusion #170 resample and mel traces.
2. Draw a liveness table for signal, frames, DFT real/imag, power, mel, stats, normalized output.
3. Mark every actual barrier and AMX/Accelerate seam.
4. Define tile ownership for frame/bin/channel tails.
5. Bind the final destination stride to the Conformer input contract.
6. Only then split AUDIO_PHASE_FRONTEND.

Keep the window/filter/FFT tables immutable and plan-owned. Do not repack weights/tables for alignment. Replace hot per-element logf under the reference trace; setup-time math is a separate category.

Completion proof

Report lane utilization, exact stage boundaries, arena high-water mark, aliases after last consumer, tail handling, and full trace results. A speedup without identical valid-length and normalized BF16 evidence is not completion.