feat(imu): unlock WHOOP 5/MG 100 Hz capture and labeled recording sessions

Packages/NoopLocalAccess/Sources/NoopLocalAccessCore/LocalAccessCore.swift

@@ -332,7 +332,7 @@ public final class ReadonlyNoopStore {
             let decodedTables = [
                 "hrSample", "rrInterval", "event", "battery", "spo2Sample",
                 "skinTempSample", "respSample", "gravitySample", "ppgHrSample", "stepSample",
-                "sleepStateSample", "ppgWaveformSample", "rawImuSample", "v18AuxSample",
+                "sleepStateSample", "ppgWaveformSample", "v18AuxSample",
             ]
             var decodedRows = 0
             for table in decodedTables where tableNames.contains(table) {

Packages/WhoopProtocol/Sources/WhoopProtocol/Whoop5RawImu.swift

@@ -8,9 +8,9 @@ import Foundation
 // app captures (documented in Asherlc/dofek `docs/whoop-ble-protocol.md` as the type-0x2B raw packet)
 // and the same columnar convention as the 4.0's 1917 REALTIME_RAW_DATA variant.
 //
-// IMPORTANT: the 5/MG's *live* raw-IMU stream is firmware-refused (TOGGLE_IMU_MODE / cmd 106 acks but
-// never streams), but the OFFLOAD buffer carries full accel AND gyro — so 100 Hz 6-axis IMU IS
-// obtainable on the 5.0 via the historical path.
+// Hardware validation on a WHOOP 5/MG established that live raw IMU requires START_RAW_DATA (81)
+// followed by TOGGLE_IMU_MODE (106) with payload [1,1]. Opcode 106 on its own acknowledges without
+// starting the producer. The same 1244-byte shape is also banked and later returned by history sync.
 //
 // Frame layout (the reassembled BLE frame = 8-byte puffin envelope + payload; offsets are FRAME-absolute):
 //   @15  u32 LE   strap unix seconds for this 1-second frame

Packages/WhoopStore/Sources/WhoopStore/Database.swift

@@ -572,16 +572,8 @@ extension WhoopStore {
         // the six wire columns (ax…az,gx…gz). Twin of the Android `rawImuSample` table (MIGRATION_20_21);
         // same column order + PK so a `.noopbak` round-trips byte-for-byte.
         //
-        // CONSUMER STATUS — deliberately none on this platform, in the same shape as the `ppgWaveformSample`
-        // (v27) and `v18AuxSample` (v31) notes. The writer (`Collector.storeRawImu`) is instrument-first: it
-        // fires only with raw capture enabled AND a 5/MG deep-data unlock, and is bounded to
-        // `WhoopStore.rawImuRetentionRows` (3600 one-second buffers, a rolling window — not a corpus). No
-        // analytic, score, gate, UI or export reads a row: the only SQL is the retention DELETE (`StreamStore`)
-        // plus a COUNT in the storage-stats readout (`LocalAccessCore`); `ImuFeatureExtractor` takes the
-        // protocol struct, never this table, so it is not a consumer either. Swift has NO reader yet, on
-        // purpose — a reader lands WITH a validated consumer, not before ("artifact, not one match", CLAUDE.md).
-        // Android keeps a dormant `rawImuSamples` reader (zero callers) for the eventual cross-check; do NOT
-        // "clean up" either side as dead code — the retained rows are the deliverable. Audit: #978.
+        // Historical rolling cache. Its opt-in writer retained at most 3,600 one-second rows, but no analytics,
+        // UI or export consumed them. v41 retires those legacy rows after file-backed capture replaces the cache.
         migrator.registerMigration("v28-raw-imu") { db in
             try db.create(table: "rawImuSample") { t in
                 t.column("deviceId", .text).notNull()
@@ -887,6 +879,10 @@ extension WhoopStore {
                 t.add(column: "skinTempC", .double)
             }
         }
+        // Retire the bounded, write-only legacy cache. Session-owned IMU now lives in the file-backed store.
+        migrator.registerMigration("v41-drop-raw-imu-sample") { db in
+            try db.drop(table: "rawImuSample")
+        }
         return migrator
     }
 }

Packages/WhoopStore/Sources/WhoopStore/DeviceRegistryStore.swift

@@ -134,7 +134,6 @@ public struct DeviceRegistryStore: Sendable {
         "ppgWaveformSample",
         // v28-raw-imu (#423): the opt-in 5/MG raw-IMU offload capture is deviceId-keyed too — "delete all
         // of this device's data" must clear it, or the raw inertial samples survive deletion (same defect).
-        "rawImuSample",
         // v31-deep-capture-channels: the banked 5/MG v18 auxiliary fields are deviceId-keyed per-second
         // rows like every stream above, so a "delete all of this device's data" must clear them too.
         "v18AuxSample",

Packages/WhoopStore/Sources/WhoopStore/Reads.swift

@@ -397,7 +397,7 @@ extension WhoopStore {
             let rawTables = ["hrSample", "rrInterval", "event", "battery",
                              "spo2Sample", "skinTempSample", "respSample", "gravitySample",
                              "stepSample", "ppgHrSample", "sleepStateSample", "ppgWaveformSample",
-                             "rawImuSample", "v18AuxSample"]
+                             "v18AuxSample"]
             var decoded = 0
             for t in rawTables {
                 decoded += try Int.fetchOne(db, sql: "SELECT COUNT(*) FROM \(t)") ?? 0

Packages/WhoopStore/Sources/WhoopStore/StreamStore.swift

@@ -58,14 +58,9 @@ extension WhoopStore {
         return out
     }
 
-    /// #423 rolling retention for the raw-IMU capture table (twin of Kotlin `RAW_IMU_RETENTION_ROWS`):
-    /// ~1 h at 1 row/strap-second (~4 MB) hard-caps the table during a multi-day offload replay.
-    public static let rawImuRetentionRows = 3600
-
     /// v31 rolling retention for the v18 aux-slot table (twin of Kotlin `V18_AUX_RETENTION_ROWS`).
     ///
-    /// `rawImuSample` is the closest precedent — raw instrumentation banked as a blob, capped rather than
-    /// unbounded — and the same reasoning applies here: nothing reads these rows yet, so a cap is far
+    /// Raw instrumentation must be capped rather than unbounded. Nothing reads these rows yet, so a cap is far
     /// cheaper to RELAX later than to impose once users have a year of history. Unbounded, this table is
     /// the one genuinely new source of row growth in v31 (the four named channels only WIDEN rows that
     /// were already being written: ~14 bytes on a `gravitySample`/`skinTempSample`/`sleepStateSample` row
@@ -102,27 +97,6 @@ extension WhoopStore {
         }
     }
 
-    /// #423: persist decoded 5/MG raw-IMU offload buffers (one row per strap-second, packed i16 BLOB),
-    /// then bound the table to the newest `retentionRows` for the device (rolling retention). Written from
-    /// the deep-buffer capture seam, not the normal stream path, so it inserts directly (idempotent by ts).
-    /// Twin of Kotlin `WhoopRepository.insertRawImu`.
-    public func insertRawImu(deviceId: String, rows: [(ts: Int, cols: [Int16])], retentionRows: Int) async throws {
-        guard !rows.isEmpty else { return }
-        try syncWrite { db in
-            let ins = try db.cachedStatement(sql: """
-                INSERT INTO rawImuSample (deviceId, ts, samples) VALUES (?, ?, ?)
-                ON CONFLICT(deviceId, ts) DO NOTHING
-                """)
-            // Pack the raw i16 columns to the LE BLOB HERE (packImuColumns is module-internal), so the
-            // caller passes plain [Int16] and never needs the packer. Mirrors how `insert` packs ppgWaveform.
-            for r in rows { try ins.execute(arguments: [deviceId, r.ts, WhoopStore.packImuColumns(r.cols)]) }
-            try db.execute(sql: """
-                DELETE FROM rawImuSample WHERE deviceId = ? AND ts < (
-                    SELECT MIN(ts) FROM (SELECT ts FROM rawImuSample WHERE deviceId = ? ORDER BY ts DESC LIMIT ?))
-                """, arguments: [deviceId, deviceId, retentionRows])
-        }
-    }
-
     /// Idempotent upsert of decoded streams by natural key. Returns the number of rows
     /// ACTUALLY inserted per stream (0 for rows that already existed).
     ///
@@ -350,9 +324,8 @@ extension WhoopStore {
             return (hr, rr, ev, bat, spo2, skin, resp, grav)
         }
 
-        // Rolling retention (the `insertRawImu` shape, #423) but AMORTISED. The delete finds the
-        // Nth-newest row by rank, so it walks up to `v18AuxRetentionRows` index entries; `insertRawImu`
-        // keeps 3,600 so that is free, this keeps 604,800 and an offload inserts once per chunk. Swept
+        // Rolling retention is amortised. The delete finds the Nth-newest row by rank, so it walks up to
+        // `v18AuxRetentionRows` index entries. The 604,800-row cap is swept
         // once per `v18AuxPruneEveryRows` rows instead, which keeps newest-N-rows exactly (a time window
         // would not — a sporadically-worn strap's rows span far more than a week, and the census wants
         // that). Counter is per device because the delete is.

Packages/WhoopStore/Sources/WhoopStore/TimestampHeal.swift

@@ -55,7 +55,7 @@ extension WhoopStore {
             let rawTables = ["hrSample", "rrInterval", "event", "battery",
                              "spo2Sample", "skinTempSample", "respSample",
                              "gravitySample", "stepSample", "ppgHrSample",
-                             "sleepStateSample", "ppgWaveformSample", "rawImuSample", "v18AuxSample"]
+                             "sleepStateSample", "ppgWaveformSample", "v18AuxSample"]
             var rawDeleted = 0
             for table in rawTables {
                 try db.execute(sql: "DELETE FROM \(table) WHERE ts < ? OR ts > ?",

Packages/WhoopStore/Tests/WhoopStoreTests/DeepCaptureChannelsTests.swift

@@ -347,7 +347,6 @@ final class DeepCaptureChannelsTests: XCTestCase {
     /// The shipped cap is a real bound, and documented in row terms.
     func testShippedRetentionConstant() {
         XCTAssertEqual(WhoopStore.v18AuxRetentionRows, 604_800)   // 7 x 86_400 strap-seconds
-        XCTAssertGreaterThan(WhoopStore.v18AuxRetentionRows, WhoopStore.rawImuRetentionRows)
     }
 
     /// Deleting a device's data must clear the aux rows too — the same privacy rule every other

Packages/WhoopStore/Tests/WhoopStoreTests/PruneTests.swift

@@ -77,4 +77,5 @@ final class PruneTests: XCTestCase {
                                               maxUnsyncedBytes: 1_000_000)
         XCTAssertEqual(pruned, 0)
     }
+
 }

Packages/WhoopStore/Tests/WhoopStoreTests/ReadTests.swift

@@ -275,19 +275,16 @@ final class ReadTests: XCTestCase {
                     ppgWaveform: [PpgWaveformSample(ts: 400, samples: [1, 2, 3])],
                     v18Aux: [V18AuxSample(ts: 400, slotValues: [1, 2])]),
             deviceId: "dev1")
-        _ = try await store.insertRawImu(deviceId: "dev1",
-                                         rows: [(ts: 400, cols: [Int16(1), Int16(2)])],
-                                         retentionRows: 1000)
         try await store.enqueueRawBatch(
             RawBatchMeta(batchId: "b1", deviceId: "dev1",
                          clockRef: ClockRef(device: 0, wall: 0), capturedAt: 1,
                          startTs: 0, endTs: 0, frameCount: 1, byteSize: 4),
             frames: [[0xAA, 0x00, 0x01, 0x02]])
         let stats = try await store.storageStats()
         // dev1: 3 hr + 2 rr + 1 event + 1 battery + 1 spo2 + 1 skinTemp + 1 resp + 1 gravity
-        //       + 1 step + 1 sleepState + 1 ppgHr + 1 ppgWaveform + 1 v18Aux + 1 rawImu = 17
-        // other: 1 hr = 1 → 18 decoded rows across all 14 raw tables.
-        XCTAssertEqual(stats.decodedRows, 18)
+        //       + 1 step + 1 sleepState + 1 ppgHr + 1 ppgWaveform + 1 v18Aux = 16
+        // other: 1 hr = 1 → 17 decoded rows.
+        XCTAssertEqual(stats.decodedRows, 17)
         XCTAssertEqual(stats.rawBatches, 1)
         XCTAssertEqual(stats.rawBytes, 4)
     }

Packages/WhoopStore/Tests/WhoopStoreTests/Resources/schema_oracle.json

@@ -41,7 +41,8 @@
     "v37-daily-avg-sdnn",
     "v38-apple-step-hour",
     "v39-ppg-burst-index",
-    "v40-daily-skin-temp-absolute"
+    "v40-daily-skin-temp-absolute",
+    "v41-drop-raw-imu-sample"
   ],
   "divergenceReasons": {
     "android-orphan-synced-column": "REAL COLUMN DRIFT: `synced` exists on Android only. GRDB's v10 note is explicit that stepSample gets no `synced` column ('unused; see StreamStore'), and v12's ppgHrSample never had one either; the Room entities copied the flag from the older per-second entities. Nothing reads it on Android — the per-row upload flag belongs to an upload path that does not exist there — so it is dead width, not divergent data. Removing it needs a Room table rebuild.",
@@ -1259,34 +1260,6 @@
       ],
       "indices": []
     },
-    "rawImuSample": {
-      "platform": "both",
-      "columns": [
-        {
-          "name": "deviceId",
-          "affinity": "TEXT",
-          "notNull": true,
-          "default": null
-        },
-        {
-          "name": "ts",
-          "affinity": "INTEGER",
-          "notNull": true,
-          "default": null
-        },
-        {
-          "name": "samples",
-          "affinity": "BLOB",
-          "notNull": true,
-          "default": null
-        }
-      ],
-      "primaryKey": [
-        "deviceId",
-        "ts"
-      ],
-      "indices": []
-    },
     "respSample": {
       "platform": "both",
       "columns": [

README.md

@@ -30,6 +30,7 @@
   <a href="https://www.reddit.com/r/NoopBand/">👽&nbsp;Reddit</a> ·
   <a href="#features">Features</a> ·
   <a href="docs/PROTOCOL.md">Protocol</a> ·
+  <a href="docs/RAW_DATA_CAPTURE.md">Raw data capture</a> ·
 </p>
 
 <p align="center">