fix: ground threshold tuning in ambient evidence

[rogerchappel]

Summary

• separate below-speech ambient peaks from average speech evidence when recommending thresholds
• preserve zero-distance semantics when speech exactly matches a preset
• add bounded regression expectations and document the reported evidence fields

Commits

• 1322ba8 fix: separate noise and speech tuning evidence
• 9a6daca docs: explain tuning evidence fields

Verification

• npm ci
• npm run release:check (41 tests, typecheck, build, fixture smoke, packed-package install/import checks)
• git diff --check

With the checked-in keyboard-noise and long-utterance fixtures, recommendations are now speechThreshold: 0.59 and noiseFloorThreshold: 0.27 instead of allowing speech peaks to drive the noise floor to its 0.5 clamp.

[rogerchappel]

Automated merge note

Triage class: auto-merge

Summary: Separates ambient/noise evidence from speech threshold tuning and preserves exact preset-distance semantics, with docs and regression tests.

Checks run: GitHub CI verify completed SUCCESS; npm ci, npm run release:check (41 tests, typecheck, build, fixture smoke, packed-package install/import), and git diff --check.

Rebased/CI-repaired: No.

Verified head SHA: 9a6daca7d68d7e2cd96ce57ccbcb1959ac7a93b0.