fix(health): rewrite Apple Health sleep as one in-bed night (Core + stale samples) - #271

sequenceDiagram
  participant HealthExporter
  participant HealthKitSleepSessionExporter
  participant MethodChannelHealthKitSleepSessionWriter
  participant NativeHealthKit
  HealthExporter->>HealthKitSleepSessionExporter: replace sleep session
  HealthKitSleepSessionExporter->>MethodChannelHealthKitSleepSessionWriter: send cleanup bounds and session
  MethodChannelHealthKitSleepSessionWriter->>NativeHealthKit: invoke replacement operation
  NativeHealthKit-->>MethodChannelHealthKitSleepSessionWriter: return result or error
  MethodChannelHealthKitSleepSessionWriter-->>HealthKitSleepSessionExporter: propagate result or error
  HealthKitSleepSessionExporter-->>HealthExporter: complete sleep export
  HealthExporter->>HealthExporter: process remaining generic export types

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

_{Comment @coderabbitai help to get the list of available commands.}

Same night on device (20 Aug 2026). Edge has the full hypnogram; Health does not.

Edge — 5h 41m asleep, 12:50 AM–7:17 AM, 6h 28m in bed. Light (Core) is present.

Apple Health — 2 hr 36 min Time Asleep. Core is empty. Only REM/Deep slivers survived.

the #225 half of this is real and i had it wrong — i assumed #258 fixed it and it didn't. main's window starts at dayStart when onset moves later, so 23:00 leftovers are never reached, and the plugin delete used .strictStartDate so it wouldn't have caught them from an overlapping window either. noon-to-noon plus overlap semantics fixes both. that part i want.

but it doesn't build, and there are two things i can't take as written.

1. the ios app doesn't compile. HealthKitSleepWriter.swift:126-128 uses .asleepREM / .asleepCore / .asleepDeep, all ios 16+, and the runner target is IPHONEOS_DEPLOYMENT_TARGET = 15.0 (project.pbxproj:774, and platform :ios, '15.0' in the podfile).

'asleepREM' is only available in iOS 16.0 or newer   :126:24
'asleepCore' is only available in iOS 16.0 or newer  :127:26
'asleepDeep' is only available in iOS 16.0 or newer  :128:25

flutter build ios --debug --no-codesign succeeds on main and fails on this head. fix is an availability guard falling back to .asleep on 15 (that's what pre-16 healthkit has), or raise the target — your call, but a guard is the smaller change. worth knowing analyze is clean and all 2833 tests pass on this branch: nothing in dart-land can see a swift compile error, which is why i asked for a real xcodebuild.

2. unobserved is exported as awake, and that fabricates a measurement. health_sleep_session.dart:193 plus the gap-fill in normalizeHealthSleepSession. onehz_pipeline.dart:1720 says it outright: "unobserved is not a stage and not wake — it is a second the band did not watch, and it belongs to neither side". seconds we didn't watch get published to apple health and health connect as measured wake, and other apps read that as fact.

and it doesn't buy the thing it's justified by — apple's time asleep is asleepCore + asleepDeep + asleepREM, so awake adds zero to it. the 2h36m → full night comes entirely from writing core at all. drop the 'unobserved' case and the append(..., awake) fill and you lose nothing the pr claims.

that's also an android regression, since normalizeHealthSleepSession is shared and HealthConnectSleepWriter.buildStage maps awake → STAGE_TYPE_AWAKE. your own test change shows it — the android assertion moved awake: 32 → 42 because a 10 min unlabelled hole is now filled as wake.

3. the cleanup window can delete a previous night, and it's not recoverable. health_sleep_session.dart:242-252 unions noon-to-noon with the whole calendar day. android only does noon-to-noon. for a night that wakes after local noon:

• d−1 night 02:00→13:30, samples at d−1 02:00–13:30
• d night 01:00→09:00 → cleanup [d−1 12:00, d 12:00) ∪ [d 00:00, d+1 00:00) = [d−1 12:00, d+1 00:00)
• the swift predicate is predicateForSamples(withStart:end:options: []) — no .strictStartDate, so it matches on overlap and deletes the whole d−1 envelope sample, not a clipped part
• d−1 is behind health_export_through, so it never gets rewritten

now that native owns the delete, just use sleepSessionCleanupRange alone like android does and drop the day union.

4. dst. health_sleep_session.dart:236 — add(const Duration(days: 1)) on a 25 hour fall-back day doesn't advance the date, so calculatedStart lands a day early and widens the window into the previous night (which feeds 3). health_export.dart:729 has a comment specifically about this trap, and the kotlin you ported from uses plusDays(1). one token: DateTime(localEnd.year, localEnd.month, localEnd.day + 1).

5. the epoch reset isn't gated on apple. health_export.dart:103-113, called at :488 — it clears health_export_through on every platform, so android replays up to 400 day bundles on the first sync after update, re-running type deletes, hourly energy writes and the minute-hr batch. nothing in the health connect writer changed in this pr. gate it on isApple. everything else about the cursor checks out — idempotent, can't loop, can't duplicate, and a partial failure leaves it sane.

on the premise: bullet 1 of your description says core and in-bed still don't land after the plugin bump. #258 merged 7h before this opened and the screenshots are of the night of 19→20 aug, so that night was slept on a build without it. i went through health 12.2.1 and couldn't find a mechanism — _alignValue maps SLEEP_LIGHT → 3 and the swift writer builds the category sample straight off that, which is .asleepCore. can you reproduce "no core" on a build that actually contains #258? bullets 2 and 3 stand either way.

things i checked and they're fine, so you don't have to defend them: the delete predicate is correctly scoped — HKSource.default() is this bundle only, not the watch, not manual health entries, not another tracker, and healthkit refuses to delete objects an app didn't save anyway. the pbxproj is 4 clean additions with no version strings touched. no derived number moves. and removing every SLEEP_* from healthDeleteTypes closes a real hazard — the plugin's delete never calls result() on an unsupported type, so that await used to hang rather than fail. every path in your swift replace() calls completion exactly once.

last thing, minor: most of the 356 changed lines in the test file are dart format churn, and it hides the three real assertion changes. we don't run it in this repo — the old style and the sdk's tall-style formatter reflow thousands of lines against each other. worth reverting the reformat so the diff is just what changed.