android: fix head-gesture axis bias, add sensitivity slider

[ohuc]

What

Head gestures (nod = answer, shake = decline) gave seemingly random results. This reworks the detector and adds a user-facing sensitivity slider.

Review note: only the second commit (android: fix head-gesture axis bias...) is the actual change — 3 files. The first commit is a checkpoint mirroring the working copy's uncommitted M3 Expressive redesign state, so this fix applies on top of the code the app is actually built from.

Why it was 'randomly biased'

Four compounding bugs in GestureDetector.kt:

1. Cross-axis contamination — peakIntervals/movementSpeedIntervals were shared between the vertical and horizontal axes, so motion on one axis corrupted the other's rhythm score.
2. Index saturation — extremes were recorded with buffer.size - 1 as their index, which pins at 99 once the 100-sample ring fills; the chronological sort and peak/trough alternation check then ran on scrambled data (i.e. detection quality degraded the longer the phone rang).
3. No expiry — peaks/troughs accumulated forever, so a stray early movement (glancing at the ringing phone) permanently primed one axis; whichever axis got primed first usually won.
4. No baseline removal — the head-tracking accel pair carries a posture-dependent DC offset, so fixed absolute thresholds favoured whichever axis rested further from zero and broke sign-alternation scoring.

The rework

• Strictly per-axis state (extremes, rhythm, direction tracking) behind a single lock.
• 8-sample baseline warmup + slow EMA (α = 0.04) — all thresholds now apply to a centered signal, immune to posture offset/drift.
• Extremes outside a 2.6 s window are pruned; only the gesture happening now counts.
• Alternation checked on chronological peak/trough types instead of value signs.
• Same confidence blend as before (amplitude 0.4 / rhythm 0.2 / alternation 0.2 / isolation 0.2, fire at ≥ 0.7), same both-axes tie-break.

Sensitivity slider

New pref head_gesture_sensitivity (0..1, default 0.5 ≈ historical thresholds), surfaced on the Head Tracking screen. It interpolates the peak threshold (550→150), full-score amplitude (950→350) and audio-feedback threshold (900→300). Read at every startDetection(), so 'Test Head Gestures' picks changes up immediately.

Testing

• :app:compileFossDebugKotlin and full assembleFossDebug build green.
• No device attached in this environment — on-device: adjust the slider, then use Test Head Gestures on the Head Tracking screen; adb logcat -s GestureDetector shows per-axis confidence. The GCAP capture broadcast is still in place for offline tuning if the endpoints need refinement.

🤖 Generated with Claude Code