Repository navigation
P1: Per-day reuse cache serves HRV-less night scans permanently when R-R lands after HR; force refresh does not bypass it #29
Description
Found by an independent review of the integrated fork state (2026-08-18). The code arrived with
the upstream range d4b7cc8..f2271cd (ryanbr#1005 / ryanbr#1395 / ryanbr#1396 per-day reuse cache); affects
upstream equally. Filed here first — forwarding upstream is a separate decision.
Problem
The per-day reuse cache in Strand/Data/IntelligenceEngine.swift:780 (same design in the
Kotlin twin, IntelligenceEngine.kt:617-643) keys a day's scan on owner + HR fingerprint
(count, maxTs) + skin anchor + config signature only. R-R / respiratory / SpO2 rows arriving
with no HR-row delta produce the identical key, so a stale cached DayScan is served.
Force-refresh does not bypass the day cache — force=true only bypasses the whole-pass
watermark gate.
Failure scenario
WHOOP history offload delivers channels independently, and offloaded HR rows duplicating live
rows are dropped by ON CONFLICT DO NOTHING. A banked night whose R-R packets land after its
HR is scored once with rmssd=nil and cached; every subsequent pass — including user-initiated
force refreshes — recomputes the same key and re-serves the HRV-less scan. Night HRV /
respiratory rate stay blank until an unrelated HR row, a config change, or an app restart
changes the key.
Possible directions
Fold the other stream fingerprints (R-R / resp / SpO2 counts + maxTs) into the day key, or let
force=true bypass the day cache as users already expect it to.
Activity
[-]Per-day reuse cache serves HRV-less night scans permanently when R-R lands after HR; force refresh does not bypass it[/-][+]P1: Per-day reuse cache serves HRV-less night scans permanently when R-R lands after HR; force refresh does not bypass it[/+]bhelm commentedon Sep 7, 2026
Largely fixed upstream by ryanbr#1907 (merged), which folds a device- and window-scoped
count/max witness for every other consumed stream — band sleep state included — into the day-cache
key, so a later, richer scan no longer gets served the earlier HRV-less result.
One part was deliberately not taken upstream:
force=truestill does not bypass the daycache, on the grounds that doing so "would effectively disable normal reuse". The cache-key fix
addresses the correctness half of this report; the force-refresh expectation is a UX question. If
you still want an explicit user-triggered refresh that ignores the cache, please reopen (or file a
fresh issue) framed as a UX request, so it can be decided on its own merits. Closing for now.