[ayiskakov]

Phase 2 fix for W3's heart-rate-only sleep path, used when a strap banks no motion (ryanbr#1801): W03-043 and W03-052 (S3), the follow-up W03-024's V2 asked for. One commit per fix, plus a docs commit from the independent review. Each row in docs/review/workstreams/W03-analytics.md holds its V0, fix, V1 and V2.

This can change stored HR-only nights on re-score. Nights detected from gravity are untouched.

The defect

• W03-043: the HR-only path anchored its sleep band on the 10th percentile of heart rate over the whole read. Since W03-024 that read is 66 h and holds the next night.
  • Lowering only the next night's heart rate removed the scored night.
  • A read ending 12 h after the night found a night the full read did not.
• W03-052: the no-motion gate skipped a day's HR-only night whenever the owner had any stored session that started anywhere in the read. analyzeDay keeps only sessions that end on the day.
  • A stored night on the day before or the day after (an imported WHOOP export, a ring's hypnogram) therefore blocked a day that had none of its own.
  • That day then scored no night at all.

The fix

• W03-043: SleepStager.hrOnlyLocalAnchors. Each epoch is held to the 10th percentile of the heart rate in the 24 h around it (radius localBaselineRadiusS, as on the gravity path).
  • At either end of the data the window slides inward rather than being cut, so every anchor stays on 24 h of heart rate, the span hrOnlyBandMult was measured on.
  • A cut window failed the field-shaped day tests, whose read ends with the night, as this morning's read does.
  • The trace prints the anchor and band as lo..hi.
• W03-052: the gate is IntelligenceEngine.noMotionNight. Only a stored session whose end falls on the day, under the pass's offset, decides it. That is analyzeDay's rule, and only those sessions are counted in the gate line's stored=.

Evidence

• V0/V1:
  • SleepHrOnlyWindowInvarianceTests fail on main and pass here.
  • NoMotionNightGateTests fail on the verbatim extraction of the old gate and pass here.
  • With the read-wide anchor restored, both invariance tests fail.
• Real data (the 2026-09-30 backup with gravity withheld, 12 stored nights of 3 h or more):
  • The night's span is the same whether the read is cut 12 h after it or not on 12 of 12 nights (old anchor: 3).
  • Nights found: 12 (old 11).
  • 2,578 of 7,269 minutes fall inside the gravity-path nights (old 2,639), and 2 minutes fall outside them (old 0).
  • The same at UTC+3 and UTC+5.
• Independent V2: holds, no S1 to S3 findings.
  • hrOnlyLocalAnchors matches an independent brute force on 43,944 cases with 0 mismatches. The cases cover repeated timestamps, gaps up to a day, out-of-range bpm and every window edge.
  • On reads under 24 h the result equals the old anchor on 27 of 27.
  • Stages and resting HR, not just bounds, are identical across the next night's heart rate and the read's reach.
  • The gate's day membership agrees with analyzeDay on 3,910 cases over ten offsets from −12 h to +14 h.
  • Cost: 0.07 s for the anchors and 0.88 s for the whole spine on 66 h at 1 Hz (debug).
  • The three doc claims it flagged are corrected in the last commit. W03-053 … W03-055 (S4) are recorded.
• Builds and tests, run locally on this branch rebased on main:
  • swift test in StrandAnalytics: 2,222 tests, 0 failures.
  • xcodebuild test, scheme Strand (macOS): 2,151 tests, 0 failures, 2 skipped.
  • NOOPiOS builds. That build was done without the watch app, which needs a watchOS runtime this machine lacks.

Not changed

• W03-054: a stored nap ending on the day still blocks that day's HR-only night, as before. Recorded.
• W03-053: the out-of-range side list can go quadratic, but only on a read made almost entirely of out-of-range bpm. Recorded.

Refs ryanbr#1801