[andre-fu]

What this adds

A run or ride recorded in another app without an HR sensor exports as a GPX carrying position and times but no heart rate. This adds a Data Sources card — Add heart rate to a workout file — that takes such a file, splices the strap's stored per-second HR into each timed trackpoint as Garmin TrackPointExtension (gpxtpx:hr), and hands the merged file straight back out through the share sheet / save panel. Uploading it to the originating service (Strava included) shows the route and heart rate / zones there.

This is the #2679 option-1 direction: automating the capability that already exists, with no third-party credential, no network, and no scope change. Refs #2679.

How it works

• GpxHrInjector (Packages/StrandImport, pure, Linux-safe): deliberate TEXT SURGERY on the original document, not a parse-and-reserialize — position, elevation, times and everything else survive byte-for-byte; the only changes are the inserted extension blocks and (when missing) the xmlns:gpxtpx declaration on the root tag. HR matches by nearest stored sample within 5 s (ties take the earlier sample).
• Read-only on the store: the file's own time window selects repo.hrSamples(from:to:) — the existing union/dedup funnel — and nothing is imported or persisted.
• Every refusal is loud and carries its numbers (not GPX / no timestamps / already has HR / nothing matched with the window and counts / too many points). A silent unchanged re-export is the one outcome the flow cannot produce.

Cross-platform parity

• GpxHrInject (android com.noop.ingest) is the byte-identical Kotlin twin, pinned by oracle per AGENTS.md: the Swift implementation's stdout over shared fixtures is pasted verbatim as the Kotlin test's expected literals (GpxHrInjectTest), covering the Strava-shaped file, the existing-<extensions> case, and the tolerance boundary, plus the same refusal cases.
• Deliberately not in this PR: the Android Compose entry point. The byte-contract layer ships on both platforms together here; the Android card + string resources are a mechanical follow-up and are called out rather than bundled (UI parity is feature-level, and keeping this PR to the contract layer + one platform's UI keeps it reviewable).

Verification

• Packages/StrandImport: full suite passes locally (263 tests incl. the 14 new — round-trip through ActivityFileImporter, byte-preservation by stripping the inserted blocks, tolerance/tie rules, every refusal).
• Android: GpxHrInjectTest (oracle + behaviour pins) passes under testFullDebugUnitTest.
• macOS app target builds (xcodebuild -scheme Strand … build) and StrandTests run locally; iOS compiles via this PR's app-build.yml run (local Xcode here has no iOS 26 SDK).
• Tools/i18n_audit.py --ci main green: the 12 new catalog keys are translated for all nine shipped locales.
• Tools/doc_comment_lint.py green.
• No BLE, migration, or scoring path is touched.

🤖 Generated with Claude Code

[andre-fu]

Withdrawing this for now — needs more local testing first. Will resubmit when it's ready.