Motion-gate the intraday stress timeline: mask exertion, not stress (stacked on #463)

[vishk23]

Stacked on #463. The motion gate extends DaytimeStress.analyze, whose signature #463 (baseline-relative daytime stress) also changes — so this branch includes #463's commits underneath and the diff shows them until #463 lands, at which point I'll rebase and only the motion-gate commit remains. Review target = the single motion-gate commit on top. The gate itself works identically in the pre-existing .dayRelative mode; it does not depend on #463's baseline logic, only on the same file.

Motion-gate the intraday stress timeline: mask exertion, don't score it as stress

Cardiac signals cannot tell psychological stress from exertion — a brisk walk and an anxious meeting both raise HR and drop HRV. DaytimeStress.analyze never saw motion, so an ambulatory hour was scored as "stress." This is the intraday timeline's central correctness gap.

The gate

When the caller supplies the day's gravity (wrist accelerometer), an hour is masked (level == nil, HourPoint.maskedForActivity == true) instead of scored when it was ambulatory for at least activityMaskFraction (0.30) of its records. "Ambulatory" = a per-record activity intensity above WorkoutDetector.motionThreshold (0.20 L2-g — the codebase's already-calibrated walk floor: desk ≈ 0.05–0.10 g, walking ≈ 0.2–0.4 g). It reuses the exact WorkoutDetector.activitySeries that SedentaryDetector / AutoWorkoutDetector already trust, so there's no new motion math.

• Excluded from the calm reference. An exertion hour's elevated HR / suppressed HRV no longer pulls the day's own calm anchor up or inflates the spread the z-scores divide by.
• Post-exercise shadow. HR stays elevated after exertion, so the single hour immediately following a directly-ambulatory hour is also masked while its mean HR is still above the calm reference — self-limiting to genuine cardiac recovery, and deliberately one hour deep so a genuinely tense afternoon that happens to follow a workout isn't masked away.
• Honest coverage. Result.activityMaskedHours counts the masked hours so a caller can say "N hours excluded — you were moving" rather than showing a silently short timeline.

Additive and honest by construction

gravity defaults empty: with no motion signal nothing is masked and the read is byte-identical to before (we only gate when we can actually see motion; hardware/imports without accelerometer simply keep the prior behaviour). StressView reads gravity through the same store handle it already uses for R-R.

Validation (real 1 Hz wrist data)

Replayed the gate's exact math over a real waking day's raw gravity (61k samples):

• Only 1 of 11 waking hours masked — the one with 36 % sustained wrist motion. Desk/light hours (3–17 % active) were all preserved and scored.
• The 0.20 / 0.30 thresholds cleanly separate genuine ambulation from ambient movement, so a step-heavy day (this wearer logs 15–35 k steps) does not empty the timeline — most "steps" are ambient (low L2-delta), not sustained vigorous motion.

Notes

• New tests pin masking, a still-HR control (identical HR but low motion still scores — proving motion, not HR, drives the mask), additive empty-gravity, and both shadow directions. Full StrandAnalytics suite green.
• This unblocks finer 5-min stress epochs (the roadmap gates finer grain behind this — without it every active epoch reads as stress).
• Surfacing maskedForActivity / activityMaskedHours in the Stress UI ("you were moving") is a small follow-up; this PR is the analytics correctness fix.
• Swift-first; the Kotlin twin isn't included here — happy to add it or leave as a follow-up.