feat: add hardware-validated YCBT R10M support

[DBozhinovski]

Summary

• add the YCBT / SmartHealth protocol stack and LittleMeatball R10M pairing support
• implement capability-gated startup, history sync, live status, spot HR/SpO2/BP measurements, battery, activity, sleep, and REM
• harden Android GATT subscription, serialization, teardown, timeout, and reconnect behavior without changing existing ring-family contracts
• make history persistence reconnect-safe, preserve revised sleep sessions, and add a Room v7 to v8 duplicate-measurement cleanup
• keep measurement settings and pairing claims aligned with firmware-declared capabilities

Hardware validation

Validated on R10M FCF4, firmware 2.32:

• pairing, handshake, reconnect, and day rollover
• activity and history sync
• HR, SpO2, and BP history and spot measurements
• battery, sleep stages, and REM rendering

The implementation intentionally does not advertise or query capabilities this firmware does not declare, including temperature, glucose, HRV, stress, fatigue, Find Device, and dedicated SpO2 history.

Verification

• TZ=UTC ./gradlew --no-daemon testDebugUnitTest --rerun-tasks
• TZ=Europe/Skopje ./gradlew --no-daemon testDebugUnitTest
• ./gradlew --no-daemon :app:assembleDebug
• manual SQLite validation of the v7 to v8 deduplication and unique-index statements
• hardware testing on R10M firmware 2.32

Debug APK: app/build/outputs/apk/debug/app-universal-debug.apk

[foureight84]

We just had a massive update. Could you please resolve the conflicts?

[DBozhinovski]

Absolutely, on it.

[foureight84]

Review notes

A few things I'd want addressed or answered before merge, grouped by severity.

High

1. handleTerminal count-mismatch silently stalls the transfer instead of retrying — YCBTHistoryTransfer.kt

if (packets != expectedPackets || bytes != expectedBytes || bytes != buffer.size) {
    return emptyList()
}

On a metadata/byte-count mismatch this returns with no block ACK and no advance(). Since the ring won't release the next type until it receives 05 80 {status}, the transfer now hangs until the 10–30 s watchdog fires and skips the whole type. The previous code fell through to the CRC path, which sent ACK_CRC_FAILURE and got exactly one retransmit — so a single dropped data frame used to cost one retry and now costs a stalled type plus a lost history category. Suggest: on a count mismatch, send ACK_CRC_FAILURE and go through retryOrSkip(type) rather than returning empty.

Medium

2. CONNECTED no longer clears real sleep for any family — EventPersistenceSubscriber.kt + DataRepairs.kt

The CONNECTED handler now calls clearDemo() on sleep sessions/blocks instead of clear(). The rationale (YCBT re-emits CONNECTED repeatedly + async history ⇒ don't flash-then-vanish a night) is sound for YCBT, but this changes behavior for the other ring families, which previously wiped-and-rebuilt sleep on every connect — and the retired DataRepairs comment shows that wipe was load-bearing for cleaning up midnight-split / mis-keyed legacy sessions. This is validated on R10M only. Please confirm the other families don't accumulate stale/duplicate sleep sessions now, or gate the "preserve" behavior to RingDeviceType.YCBT.

3. MIGRATION_12_13 does a silent destructive dedup + unique index — PulseLoopDatabase.kt

DELETE FROM measurements WHERE rowid NOT IN (SELECT MAX(rowid) ... GROUP BY kindRaw, timestamp, sourceRaw)

then CREATE UNIQUE INDEX ... (kindRaw, timestamp, sourceRaw). timestamp is epoch millis, so this is usually safe, but it permanently drops any pre-existing rows that legitimately share (kind, ms, source). Worth confirming no existing family ever wrote two valid rows at the same kind+ms+source (combined-packet bursts, live sampling). Related: this PR relabels live HRV/temperature sourceRaw from "colmi" → "live", so already-persisted "colmi" rows and new "live" rows for the same kind+ms will both survive the unique index (parallel lineages). Probably cosmetic, but worth calling out.

4. forget() / factoryReset() dropped the stop() call — RingSyncCoordinator.kt

Both paths went from client.forget(); stop() to just client.forget(). stop() cancelled streamJob/startupJob. Leaving them running means the coordinator keeps observing (and a startupJob may still be mid-flight) against a device being torn down. Intentional (to survive a re-pair)? If so a one-line comment helps; otherwise it's a late-write/leak risk.

5. RESPIRATORY_RATE / VO2MAX are decoded then unconditionally filtered out — YCBTDriver.isSupported vs MetricsService.kt

combinedVitals/bodyData decode RESPIRATORY_RATE and VO2MAX, but YCBTDriver.isSupported maps both to false, so events.filter(::isSupported) drops them and nothing is ever persisted. Meanwhile MetricsService.isSupported was just changed to gate them on db.measurementDao().latest(kind) != null — which can now never be non-null for YCBT. So it's dead decode feeding an unreachable UI branch. Pick one: stop decoding them, or let them through.

Low / polish

6. RingBLEClient.onCharacteristicChanged now holds forgetLock across the entire driver.ingest() (full history-buffer CRC + decode) and every publishBlocking. The GATT callback thread is already serial, so the lock really only guards against finalizeForget() on another thread — consider narrowing its scope so it isn't held through a large history decode.

7. PulseEventBus.publishBlocking now does check(pending.trySend(event).isSuccess). The channel is UNLIMITED and never closed, so this is practically unreachable — but a check() throwing IllegalStateException from a BLE callback thread is a sharp edge; prefer logging over throwing there.

8. measureBloodPressure() hand-rolls the poll loop (for step in 0 until BP_MEASURE_SECONDS*2 { delay(500) }) while measureHR/measureSpO2/measureHRV all use pollForValue(...). Reuse pollForValue for consistency.

9. SPOT_MEASURE_SECONDS = HR + SpO2 + BP(40) + HRV(40) + 3 ≈ up to ~110 s, and measureSpot() runs the four legs sequentially with the optical sensor pinned. Early-exit mitigates the typical case, but the worst-case countdown is ~2 min — confirm that's the intended UX for rings that grant all four manual caps.

10. YCBTSyncEngine.writer was changed val → var but is never reassigned — revert to val.

11. decodeDeviceInfo now emits FirmwareVersion(major*100 + minor) (e.g. "1.05" → 105); the prior code deliberately avoided this int encoding as misleading. Fine if consumers treat it as opaque — just confirming.

12. In YCBTDriver.ingest, a MEASUREMENT_STATUS push carrying a real value clears the entire pendingMeasurementReplies FIFO. During a workout HR stream plus a concurrent spot measurement, that could drop an unrelated pending correlation. Edge case, but noting it.

[foureight84]

Protocol / wire-layer pass

Went through the framing, CRC16, opcode tables, history record offsets, and the support-function bitmap separately — the wire layer looks correct, so nothing blocking there. Two small things worth a look:

Unverified/possibly-wrong mode constants — YCBTMeasurementMode.BLOOD_KETONE = 0x07 and RESPIRATORY_RATE = 0x03 don't line up with the mode numbering the ring actually uses for those measurements, and BLOOD_KETONE in particular looks off. Both are inert today (nothing ever starts a ketone/respiratory-rate spot measurement), so there's no runtime impact — but worth double-checking before anything is ever wired to them.

Inferred conversions with no hard ground truth — the blood-sugar tenths-of-mmol → mg/dL conversion (bloodSugarMgdl, used by the 0x2f, 0x09, and 04/13 paths) and the stress/fatigue ×10 digit-concatenation in score() are interpretations rather than confirmed decodes; the code already carries UNVERIFIED notes on them. Just flagging that these are the decodes most likely to need a real hardware check, since a wrong scale would still land plausible-looking values.

Minor: the 06/00 live-status comment still labels distance/calories "UNVERIFIED (capture-inferred)" — those offsets are actually correct, so the comment is stale.

[DBozhinovski]

Addressed all review notes in 849131d and merged the latest main in b0b5fd9.

1. History terminal count mismatches now send ACK_CRC_FAILURE, retry once, then skip/advance. Added coverage for the mismatch/retry path.
2. Sleep preservation is limited to the YCBT protocol stack. YCBT, TK5, and COLMI_SMART_HEALTH all use YCBTDriver and share its repeated status packets and asynchronous history transfer; packet-based Colmi/Jring/CRP families still clear and rebuild sleep on connect.
3. Removed destructive measurement deduplication and the unique identity index. The migration now preserves legitimate same-millisecond collisions, normalizes the old HRV/temperature source lineage, and adopts deterministic primary keys only for replayed history rows.
4. Forget/factory-reset now stop coordinator jobs before unbind, await ACK-or-timeout teardown, then restart the process-long event collector so same-process re-pairing still works.
5. Respiratory-rate and VO2max history now pass capability filtering, plausibility validation, and persistence.
6. forgetLock now protects only short generation/state checks. Decode, event publication, sync-engine handling, GATT teardown, preferences, and state updates run outside it.
7. Event-bus send rejection logs instead of throwing from a BLE callback.
8. Blood-pressure measurement now uses the shared polling helper.
9. Confirmed: 173 seconds is the intentional worst-case UI countdown for sequential HR + SpO2 + BP + HRV. Capability gating and stable-reading exits shorten normal runs.
10. Restored YCBTSyncEngine.writer to val.
11. Removed the misleading integer firmware event from the YCBT device-info path; firmware remains the protocol-formatted string.
12. Measurement pushes now remove only the pending correlation for the matching mode.

Wire-layer follow-ups:

• Removed the inert, unverified blood-ketone and respiratory-rate spot-mode constants.
• Restored explicit UNVERIFIED notes on blood-sugar scaling and stress/fatigue score composition.
• Removed the stale distance/calorie uncertainty comment.

Additional lifecycle hardening found during device testing:

• Intentional background disconnects no longer get undone by the 15-second watchdog.
• Foreground and short-lived background clients share a process-wide GATT ownership gate, and workers yield promptly when the app foregrounds.
• Worker-owned clients always destroy their watchdog/GATT resources.
• Graceful teardown persists DISCONNECTED rather than leaving stale state.

Verification:

• testDebugUnitTest passed.
• :app:assembleDebug passed after merging current main.
• git diff --check passed.
• In-place Pixel 7 testing preserved app data and validated connect, history sync, foreground-worker exclusion, background teardown without watchdog reconnect, and foreground reconnect.
• Migrated database is v14 with PRAGMA integrity_check = ok; ring identity, measurements, and sleep sessions were retained.