Sleep scoring over-calls WAKE when resting HR is elevated but the wrist is motionless (supplement / fever / alcohol nights) #462

Summary

Both stagers (SleepStager V1 and the default SleepStagerV2) call wake primarily off HR and
HR-variability, and the session-detection HR gate (SleepStager.confirmSleepWithHR) rejects a still
in-bed run whose median HR sits above the sleep band. Neither path consults motion or posture
before making that call. On a night whose resting HR is held elevated without the wearer getting up
— a supplement protocol, a fever, a hot room, alcohol — the HR-led logic reads hot-but-motionless
sleep as wake. This over-calls WAKE (tanking efficiency, Rest, and the debt ledger) and mis-places
sleep onset, and it has required manual relabeling on multiple confirmed nights.

The motion-aware wake refinement in #402 (merged 2026-07-13) does not prevent this, for three
compounding reasons detailed below.

Evidence — two confirmed nights (both manually corrected)

Night A (2026-07-13). A supplement protocol held overnight resting HR at ~55–60 (vs a natural
~47–51) and skin temp +0.4…+1.2 °C. The scorer produced 194 min WAKE across a ~6.4 h in-bed span
(efficiency ≈ 0.50). Minute-level review of the largest WAKE blocks found zero walking cadence:
each was a brief turn-over burst (a single-minute posture-variance spike) bracketed by minutes of
complete stillness (per-minute posture variance < 0.05, no step ticks), recurring on a ~90-min rhythm
consistent with hot-but-atonic REM being scored as wake. Correct WAKE was ~67 min.

Night B (2026-07-14, PDT). Onset was scored 1:41 AM vs a sensor truth of ~1:29 (wrist still
from 1:29:00, HR at 50 by 1:31), and a 44-min solid WAKE block 1:41–2:27 was emitted where motion
shows stillness with occasional turn-overs and HR 55–65. Sleep end (7:32:38) was correct.

The discriminating pattern, in both nights:

• Sleeping: per-minute posture variance < 0.01–0.05 with zero walk ticks; HR flat 50–65.
• Genuinely awake / out of bed: sustained walking cadence (walk ticks > 0 held for minutes, e.g.
  35–53 steps/min) or sustained high posture variance with an HR jump (79–100 bpm, 46–109
  steps/min).

Elevated-but-flat HR with a motionless wrist is not in the "awake" column.

Why #402 does not prevent this

1. Default-OFF. The refinement ships behind PuffinExperiment.motionAwareWakeEnabled, default off.
   A wearer on a supplement protocol gets no protection unless they know to toggle an Experimental flag.
2. Wrong layer. It is a post-pass over an already-staged hypnogram — it only reclassifies WAKE
   stage segments inside an already-accepted session to light. It never participates in
   detectSleep / confirmSleepWithHR, so it cannot fix Night B's late onset or protect a session
   the HR gate rejected outright. And on Night A the over-called WAKE is produced by the stager's own
   HR-led emission model, upstream of any post-pass tuning.
3. Requires dense STEP data. Its density self-gate (isMotionDense) needs ≥80% of minutes to carry
   StepSample records with an activityClass. Its corroboration keys on walk-class step ticks.
   When that stream is sparse or the activity class is unattributed, the gate declines and the pass is a
   no-op — exactly on the nights that need it.

Proposed fix — motion-corroborated wake, on by default, at both layers

The rule: elevated HR alone is insufficient to call wake. An epoch or run at the night's quiescent
motion floor with unchanged posture cannot be scored WAKE on cardiac evidence alone. Corroborate
with the gravity posture/jerk signal both stagers already consume (always present), not step ticks.

• (a) Stage level (default SleepStagerV2). On a motion-quiescent epoch (no observed movement;
  peak per-second jerk at/below the night's own quiescent floor × the wake-gate multiplier), clamp the
  cardiac contribution to the AWAKE emission to ≤ 0 — keep any wake-suppressing (low, flat HR) evidence,
  discard the wake-promoting half. Genuine motion (the movement z-score and the night-relative jerk
  gate, which by construction cannot fire on a quiescent epoch) still drives wake. A still low-HR sleep
  epoch is byte-identical; only a still epoch whose elevated HR was about to push it awake is held.

• (b) Detection level (confirmSleepWithHR). When a run is deeply motion-quiescent (≥ ~90% of its
  dense-gravity minutes are posture-stable, per-minute variance < 0.05), widen the HR sleep band from
  ×1.05 to ×1.30, so a supplement-elevated but motionless overnight run is not rejected. The band still
  keeps a floor: a run whose median HR is above even the widened band is rejected even when still, so
  genuine all-night in-bed wakefulness is not scored asleep. With no gravity evidence the strict band
  stands (unchanged).

• (c) Personalised adaptive HR baseline. Derive the sleep band from the wearer's recent overnight
  medians (adaptiveOvernightHRBaseline) rather than a fixed day-median, so a sustained supplement /
  fitness era self-calibrates; keep a floor so a wakeful era can't collapse the band.

Out-of-bed detection is untouched: sustained walking cadence and sustained posture change with an HR jump
still break runs and score WAKE.

Notes

• Swift lands first (Packages/StrandAnalytics); the Kotlin analytics twin
  (com.noop.analytics) transcription is a deliberate follow-up to preserve the parity contract.
• All thresholds are named constants fixed a-priori from the two reference nights, not fit to labels,
  consistent with the repo's rule for a derived physiological signal.