HR-led wake detection over-calls WAKE (and mis-times onset/wake) when overnight HR is elevated — should be motion-weighted

[vishk23]

Problem

The sleep/wake detector leans heavily on heart rate to separate wake from sleep. When overnight HR is elevated for a NON-wake reason — stimulants/supplements, illness/fever, alcohol, a hot room, stress — the detector reads the elevated HR as wakefulness: it over-calls WAKE mid-night and mis-times onset + final wake, fragmenting one night into several short, low-efficiency sessions.

Observed on real WHOOP-era nights where a supplement held overnight HR ~55–60 (vs a normal ~45–50): nights that were genuinely ~6–7 h of sleep detected as fragmented 2–5 h pieces (efficiency 0.48–0.77), needing 2–3 manual re-stages. The mis-scored short sessions then propagate into Rest, sleep debt, and the next day's overnight-quiet baseline.

Proposal

Weight motion more heavily (and HR less) in the wake decision — a still body with elevated HR is far more likely asleep-with-high-HR than awake. Options: require accelerometer/gravity agreement before an HR-driven wake call; scale the HR wake threshold by the person's OWN recent overnight-HR baseline (so a globally-elevated night doesn't trip it); or suppress HR-led wake during sustained stillness.

Context

Recurring on my own device; surfaced again while validating Wave 0 sleep scoring (#456) on real history.

[ryanbr]

Closing as a duplicate of #462, resolved on main by #465 (merged).

This is the same HR-led-wake-over-call you described in more detail in #462 (filed ~an hour later with the two evidence nights). #465 ships the motion-corroboration at both layers, on by default, reading the gravity posture/jerk signal that's always present:

• "require gravity agreement before an HR-driven wake call" / "suppress HR-led wake during sustained stillness" → SleepStagerV2.motionQuiescent clamps the AWAKE cardiac term to ≤ 0 on a motionless epoch (keeps wake-suppressing low-HR evidence, drops the wake-promoting half), and SleepStager.confirmSleepWithHR widens the sleep band ×1.05→×1.30 on a deeply motion-quiescent run (with a floor, so a genuinely-awake still run is still rejected). Byte-identical on normal nights — the frozen goldens are unchanged.
• "scale the HR wake threshold by the person's own recent overnight baseline" → shipped as the floor-protected adaptiveOvernightHRBaseline helper; its live cross-night wiring in IntelligenceEngine is the one remaining follow-up, tracked on Sleep scoring over-calls WAKE when resting HR is elevated but the wrist is motionless (supplement / fever / alcohol nights) #462.

Kotlin twin landed in the same PR, so parity holds. Thanks for the clear write-up on both.