Wide-path consolidation: offline runs the shipped wide stage, M/S stereo (Stage 14)

[robmorgan]

ROADMAP Stage 14, behavior half. The dead-code/doc sweep follows in a separate PR.

Offline wide = live wide

stretch() beyond ±20% (within the engine's 0.25–4.0 rate range) now renders through the shipped WideKeylockStage, via the same host-emulation driver as the keylock branch — the profile is simply parameterized. The direct batch PV survives only beyond 4× either way (outside any deck use).

• Determinism extends to wide rates: streaming (WideKeylock profile) and offline renders are sample-identical at 0.5× and 1.5× (streaming_offline_determinism.rs).
• The Stage 13 sub-bass trade dissolves: through the live stage the PV runs at unity with the resampler transposing, and the two-tone balance measures at the ideal (~0.53–0.56 amplitude ratio) at every wide ratio — the rigid sub-bass attenuation band pinned two days ago is gone. Sweep bands re-pinned at ideal-centered values.

Mid/side stereo for the wide corrected path

Centered content lives entirely in M under one vocoder: identical channels stay bit-identical, and center-to-side leak is structurally impossible (S = 0 encodes to exact zeros through the whole chain). The raw/toggle arm stays L/R.

Honest measurements, including the fixture archaeology:

• Identical-channel and steady-side-tone fixtures measured identical pre/post — deterministic pipelines can't diverge on identical input, and a strong center is always its own locked peak. Both are documented as non-discriminating.
• The discriminating fixture (independent noise beds per channel, shifting peak landscape around a center tone): per-channel leak 64.6 dB rejection → M/S 70.9 dB. Modest — the gates in tests/wide_stereo_coherence.rs are regression tripwires, and the audible width verdict belongs to the owner listen. Note for that listen: Stage 11 heard ours as "a bit crowded" vs R3's "wider, more open," and M/S touches exactly that quality.

Peak-set unification

fill_spectral_peaks now applies the same MIN_PEAK_MAGNITUDE floor as the vocoder's own peak pass (constant moved to phase_locking.rs, shared), so locking, gradient integration, and IF-blend operate on one peak set — noise-floor ripple no longer counts as peaks in quiet passages (review finding).

Verification

A/B matrix, keylock/transient QA gates, pv_null, wide-rate determinism, stereo coherence gates, and both full suites (debug + release) green.

Remaining stage items

Dead-code/doc sweep (next PR): hop_synthesis, the no-op envelope block, the set_stretch_ratio seam-case chain, legacy LR8Crossover, duplicate bessel_i0s, stale Makefile/qa/README.md/RESEARCH.md references. Then the owner ±30/±50 listen vs the pre-consolidation renders.

🤖 Generated with Claude Code