[digitalerdude]

Addresses the confirmed, narrowed part of #510: "the issue with the missing HR and calories in the workouts is still there. However, the values are only missing when I manually enter the workout afterward. When I start and end a workout through the app, the values are displayed correctly." (The Effort-score part of the original report turned out to be unrelated — wearing the strap on the bicep for a more accurate HR reading, per the reporter's own follow-up.)

Root cause

Confirmed directly against the reporter's own Test Centre → Workouts trace (attached to the issue):

[workouts] detectedBout verdict=droppedOverlap durMin=11 avgBpm=150 overlapSource=manual
[workouts] detectedBout verdict=droppedOverlap durMin=7 avgBpm=151 overlapSource=manual

IntelligenceEngine's auto-workout detector independently finds the same activity from raw HR + motion and computes a real avgHR/calories/maxHR/strain for it (avgBpm=150 right there in the trace) — then throws all of that away the instant it overlaps a real (manual/imported) row, purely to avoid persisting the same workout twice (#975).

Separately, WhoopRepository.fillWorkoutHrFromStrap (the display-time fill) reads HR strictly inside the row's own [startTs, endTs] window. A manual entry's user-typed boundaries rarely match the detector's own (tighter) boundaries for the same activity, so that raw-window read can come up under minSamples (60) even on a single-strap install with dense HR coverage — while the detector, using its own boundaries, found plenty. Net effect: real, already-computed HR/calories data existed for the exact activity and was silently discarded instead of being used.

Fix

When a detected bout collides with a real row, backfill only the fields that row doesn't already have — never touching anything typed by the user, imported, or previously filled — from the bout's own computed values, then upsert in place (same natural key (deviceId, startTs, sport), so this updates rather than duplicates).

Extracted the merge into a small, pure, tested helper on each platform so the logic is unit-testable without standing up the whole analyzeRecent pipeline:

• Android: IntelligenceEngine.backfillWorkoutFromDetectedBout
• Swift: WorkoutDetector.backfillWorkout (in Packages/StrandAnalytics, so it's covered by swift test rather than only the App-target test suite)

Added a new trace verdict droppedOverlapBackfilled (alongside the existing persisted / droppedOverlap) so the Workouts test-mode trace can distinguish "dropped and backfilled" from "dropped, nothing to add" — useful for the next report like this one.

Scope

Byte-parity: both platforms touched in this PR, same natural key / same field semantics, no analytics formula or stored-value shape changed — this only decides which of the fields a row already has get filled in from data the pipeline already computes.

Testing

• New unit tests, both platforms, exercising fill / never-override / no-op-when-already-full:
  • Android: IntelligenceEngineWorkoutBackfillTest (3 tests)
  • Swift: WorkoutDetectorBackfillTests (3 tests)
• ./gradlew compileFullDebugKotlin testFullDebugUnitTest --tests "com.noop.analytics.*" ✓ (JDK 17 via Homebrew)
• swift test (Packages/StrandAnalytics): 1110 tests, 0 failures ✓
• xcodegen generate && xcodebuild -project Strand.xcodeproj -scheme Strand -destination 'platform=macOS' CODE_SIGNING_ALLOWED=NO build — BUILD SUCCEEDED
• xcodebuild -project Strand.xcodeproj -scheme NOOPiOS -destination 'generic/platform=iOS Simulator' CODE_SIGNING_ALLOWED=NO build — BUILD SUCCEEDED
• Not tested on-device — this only changes what gets written to already-computed rows during the periodic analytics pass, no BLE path touched