docs(spec): §3.1 preprocessing pipeline steps 1–4 still describe pre-M14 implementation

[shaypal5]

Surfaced as PR #82 review thread COPILOT-2 (resolved as out-of-scope for #78). The complaint is correct and pre-existing — should have been synced as part of M14 (PR #48, 2026-05-01) and was not.

Drift between spec and implementation

docs/spec.md §3.1, lines 110–113:

1. **Resample** — convert to 16,000 Hz (SoX `rate` with VHQ quality, or `torchaudio.functional.resample`)
2. **Downmix** — stereo → mono (average channels)
3. **Spectral filter** — low-pass at 7,500 Hz to remove irrelevant high-frequency noise from budget sensors (Butterworth order 4)
4. **Denoising** — spectral subtraction (Wiener filtering) to remove electrical hum; parameterize noise profile from silent leading segment

Actual implementation in synthbanshee/augment/preprocessing.py (post-M14):

step	spec (stale)	code (current)
1	SoX or torchaudio	scipy.signal.resample_poly (CLAUDE.md forbids torchaudio)
2	average channels	data.mean(axis=1) if multichannel — match
3	7.5 kHz low-pass Butterworth-4	80 Hz high-pass Butterworth-2 (M14 removed the LPF entirely; LPF at 7.5 kHz on 16 kHz audio is just below Nyquist and was removing real signal)
4	Wiener spectral subtraction (default on)	Optional Wiener via PreprocessingConfig.wiener_denoise, default False (M14 changed default; on-by-default Wiener over-smoothed clean TTS)

Why this matters

Anyone reading spec.md §3.1 to understand the pipeline gets pre-M14 information. Step 5 (loudness, just rewritten in #82) is in sync; steps 1–4 are not. This is exactly the paper-vs-reality drift that lets future audits over-trust either source.

Suggested fix

Rewrite §3.1 steps 1–4 to match current preprocessing.py:

1. **Resample** — convert to 16 kHz with `scipy.signal.resample_poly`
   (polyphase filter; `torchaudio` is forbidden in this repo, see AGENTS.md).
2. **Downmix** — stereo → mono via channel averaging.
3. **High-pass filter at 80 Hz** (Butterworth order 2, sos form) to remove
   DC and sub-bass rumble that small phone microphones cannot capture.
   Note: M14 (PR #48) removed the legacy 7.5 kHz low-pass filter — at
   16 kHz Nyquist it was destroying sibilants and breathiness cues.
4. **Wiener denoising** — *optional*, controlled by
   `PreprocessingConfig.wiener_denoise` (default `False`).  M14 changed the
   default because Wiener on clean TTS output over-smooths high-frequency
   transients.  Enable only for clips with real added noise (Tier B/C
   after acoustic augmentation).

Plus a sentence at the top of §3.1 noting that authoritative pipeline order lives in synthbanshee/augment/preprocessing.py:preprocess() so future drift is harder to introduce silently.

References

• PR feat(m14): fix muffled audio, click artifacts, and voice identity shifts #48 (M14, 2026-05-01) — introduced the drift; should have synced spec.
• PR fix(preprocessing): #78 define loudness contract + metadata trail (does NOT recover Whisper — see #83) #82 review thread COPILOT-2 — surfaced this; resolved as out-of-scope.
• synthbanshee/augment/preprocessing.py:preprocess() — authoritative source.