[kavemang]

Summary

Refs #2522.

• Take the daily resting-HR value from the longest sleep session's existing, gated five-minute floor instead of its whole-session sample mean, on Swift and Android.
• Preserve the fix(analytics): wire primary-session resting HR to fix nap floor dist… #2358 primary-session selection, so a short lower-HR nap cannot replace the main night. Preserve the Oura-only night is not scored: the ring's hypnogram never reaches the nightly scorer (motion-based staging can't run on a ring) #804 device-provided primary value when present.
• Remove the former cross-session .min() fallback: when the selected primary night has no HR samples, daily resting HR stays absent instead of borrowing a nap value. That can lower RHR baseline-night coverage; Android’s V5HealthSignals trust gate needs 14 nights with RHR when the HRV baseline is also sparse.
• Keep the whole-session mean in rhr_primary_session as a shadow metric, and update its comments to describe that role.

This changes a scored input and may move Charge, Effort, and other downstream values on re-score. It restores the previous floor definition within the primary night; it makes no claim of measured accuracy improvement. Oura's API documents a lowest five-minute HR and a separate average HR for each sleep session.

Verification

• New Swift/Kotlin parity fixtures cover a main-night floor below its mean, a still-lower short nap, a provided primary value, and a primary night with no HR. The Android fixture passed (PrimaryNightRestingHRDailyTest).
• Existing Android PrimarySessionRestingHRTest and AnalyticsEngineRefusedMainNightHrvTest passed.
• python3 Tools/doc_comment_lint.py and git diff --check passed.
• Swift and Apple app compilation require Xcode/Swift, unavailable on this Linux host; CI is needed for the Swift side.

Validation

• R1 Full preregistration against a participant reference set: unavailable. The synthetic test expectations were written before running the tests, but they establish behavior only.
• R2 Self-comparison audit: not applicable; no participant DB or stage reference was scored.
• R3 Regenerating commands and input: 2026-09-28, checked-in synthetic fixtures; cd android && JAVA_HOME=/usr/lib/jvm/java-17-openjdk ./gradlew :app:testFullDebugUnitTest --tests 'com.noop.analytics.PrimaryNightRestingHRDailyTest' --offline --no-daemon and the two existing suites above.
• R4 Instrument matches the question: unit tests establish the selected formula and precedence, not physiological accuracy.
• R5 Temporal held-out split: no participant reference set available; no held-out accuracy number is reported.
• R6 Mechanism: AnalyticsEngine previously selected PrimarySessionRestingHR.meanHR before the session floor; the new fixture reaches that exact path and distinguishes 55 bpm floor from >69 bpm mean.
• R7 Figures: the values above are synthetic bpm inputs/outputs from three checked-in cases (n=0 participants), measured by the Android unit test on 2026-09-28; no inherited accuracy statistic is used.
• R8 Accuracy on real nights is not measurable on the available data and is not claimed here.

Command(s): the Android commands above; Swift counterpart is cd Packages/StrandAnalytics && swift test --filter PrimaryNightRestingHRDailyTests on a Swift-equipped host.
Date run / input: 2026-09-28; checked-in synthetic fixtures.
Prediction (pre-registered in the tests): daily RHR is 55 bpm for the mixed main-night/nap fixture, 62 bpm for the provided-value fixture, and absent when the main night has no HR.
Outcome (confirmed): Android fixtures passed. Swift fixtures remain for CI.

[ryanbr]

@kavemang Reviewed. 17/17 green, including test (StrandAnalytics), which covers the Swift side you could not build.

Checked rather than assumed:

• The mechanism claim holds. sessionRestingHR really is the gated lowest five-minute bin (sessionRestingHR: any bin can win the night's floor, including a one-sample one #1943 gating, then ungated lowest bin, then all-sample mean), and the gating is identical on both platforms.
• Swift and Kotlin are the same edit structurally.
• The shadow metric survives: primarySessionRestingHRWithCoverage is still called from IntelligenceEngine.swift:1794 and IntelligenceEngine.kt:1438, so the old mean is not dead code.
• Test parity is exact. Three cases, same names, same expectations (55 / 62 / nil), and both assert sessionRestingHR(...) == 55 directly, so they pin the mechanism rather than the outcome.

One thing the summary understates. Besides mean to floor, the cross-session .min() fallback is gone:

old:  provided ?? primarySessionRestingHR(...) ?? physiologySessions.compactMap { $0.restingHR }.min()
new:  provided ?? primarySession?.restingHR

Deliberate and tested, and narrower than it looks, since sessionRestingHR is never-null-on-data: the new nil happens only when the primary session has no HR samples at all. Worth naming in the summary though, because it is a coverage change and not only a definition change.

One finding, not blocking. That nil still lowers count { it.restingHr != null }, and V5HealthSignals.kt:71 gates on it:

val rhrBaselineTrusted = baselineWindow.count { it.restingHr != null } >= MIN_BASELINE_NIGHTS  // 14

So a day that used to get a nap-derived RHR no longer counts toward that threshold. || hrvBaselineTrusted means it only bites when HRV is also sparse, and skipping a nap value is still the right call, but this reader counts emptiness rather than reading values and is not mentioned anywhere in the PR. ReadinessEngine uses compactMap and is fine.

Your base is 9 commits behind main. Rebase and I will take it.

Thanks for the R1 to R8 framing, and for marking R1 and R5 unavailable rather than reaching for a number.