diagnostics: REM funnel names the stager it replays vs the one that staged the screen

Packages/StrandAnalytics/Sources/StrandAnalytics/SleepStager.swift

@@ -2656,12 +2656,21 @@ public enum SleepStager {
         return .noRespFallbackBar                          // resp never measured and the no-resp bar unmet
     }
 
-    /// Read-only REM-funnel triage for ONE in-bed window [start, end] (#688). Re-runs the SAME Stage-0→3
-    /// staging seam `stageSession` uses (epoch grid → Cole–Kripke → features → classify → smooth →
-    /// re-impose), but instead of emitting a hypnogram it COUNTS where REM was lost. Changes NOTHING:
-    /// no label, no score, no session. Returns nil only when the window has too little gravity to grid
-    /// (mirroring `stageSession`'s degenerate fallback, which carries no REM to explain). The caller
-    /// logs `.summary`; tests assert the counts. Pure + deterministic. (#688)
+    /// Read-only REM-funnel triage for ONE in-bed window [start, end] (#688). Re-runs THIS type's
+    /// Stage-0→3 seam (epoch grid → Cole–Kripke → features → classify → smooth → re-impose), but
+    /// instead of emitting a hypnogram it COUNTS where REM was lost. Changes NOTHING: no label, no
+    /// score, no session. Returns nil only when the window has too little gravity to grid (mirroring
+    /// `stageSession`'s degenerate fallback, which carries no REM to explain). The caller logs
+    /// `.summary`; tests assert the counts. Pure + deterministic. (#688)
+    ///
+    /// WHICH HYPNOGRAM THIS EXPLAINS. V1's, always — this function is `SleepStager`'s own seam. It used
+    /// to say it explained "the SAME hypnogram" as the screen, and that has been false since V2 became
+    /// the default: the shipped hypnogram is staged by `SleepStagerV2` whenever
+    /// `experimentalSleepV2Enabled` says so, which is by default. On a 5/MG the two can be far apart,
+    /// because V1's primary REM gate needs the raw respiratory channel the hardware never emits while V2
+    /// recovers respiration regularity from R-R: one field pair reported ~46 min REM here against hours
+    /// on the screen for the same night (#2365). The caller's line names both stagers for that reason
+    /// (#2366); do not restore the claim that they are one. 
     public static func remFunnelDiagnostic(start: Int, end: Int, grav: [GravitySample],
                                            hr: [HRSample], rr: [RRInterval],
                                            resp: [RespSample]) -> REMFunnelDiagnostic? {

Strand/System/DebugDataDiagnostics.swift

@@ -309,7 +309,17 @@ enum DebugDataDiagnostics {
             return lines
         }
         if let rem = SleepStager.remFunnelDiagnostic(start: cs.startTs, end: cs.endTs, grav: grav, hr: hr, rr: rr, resp: resp) {
-            lines.append(rem.summary)
+            // The funnel replays the V1 classifier, but the shipped hypnogram is staged by V2 whenever
+            // the default-on flag says so — name both, or the two totals read as one fact disagreeing.
+            // On a 5/MG the gap is maximal: V1's primary REM gate needs the raw resp channel that
+            // hardware never emits, while V2 recovers respiration from R-R, so the funnel can report
+            // ~46min REM against a 231min screen for the same night.
+            let screenStager = PuffinExperiment.experimentalSleepV2Enabled ? "V2" : "V1"
+            var summary = rem.summary + " · funnel replays V1; screen staged by \(screenStager)"
+            if screenStager != "V1" {
+                summary += " — totals can differ"
+            }
+            lines.append(summary)
         } else {
             lines.append("REM funnel: insufficient motion data (<2 gravity samples)")
         }

android/app/src/main/java/com/noop/analytics/SleepStager.kt

@@ -2955,12 +2955,21 @@ object SleepStager {
     }
 
     /**
-     * Read-only REM-funnel triage for ONE in-bed window [start, end] (#688). Re-runs the SAME
-     * Stage-0→3 staging seam [stageSession] uses (epoch grid → Cole–Kripke → features → classify →
-     * smooth → re-impose), but instead of emitting a hypnogram it COUNTS where REM was lost. Changes
-     * NOTHING: no label, no score, no session. Returns null only when the window has too little gravity
-     * to grid (mirroring [stageSession]'s degenerate fallback, which carries no REM to explain). The
-     * caller logs `.summary`; tests assert the counts. Pure + deterministic. Mirrors Swift. (#688)
+     * Read-only REM-funnel triage for ONE in-bed window [start, end] (#688). Re-runs THIS object's
+     * Stage-0→3 seam (epoch grid → Cole–Kripke → features → classify → smooth → re-impose), but
+     * instead of emitting a hypnogram it COUNTS where REM was lost. Changes NOTHING: no label, no
+     * score, no session. Returns null only when the window has too little gravity to grid (mirroring
+     * [stageSession]'s degenerate fallback, which carries no REM to explain). The caller logs
+     * `.summary`; tests assert the counts. Pure + deterministic. Mirrors Swift. (#688)
+     *
+     * WHICH HYPNOGRAM THIS EXPLAINS. V1's, always — this is [SleepStager]'s own seam. It used to say it
+     * explained "the SAME hypnogram" as the screen, and that has been false since V2 became the
+     * default: the shipped hypnogram is staged by `SleepStagerV2` whenever `experimentalSleepV2` says
+     * so, which is by default. On a 5/MG the two can be far apart, because V1's primary REM gate needs
+     * the raw respiratory channel the hardware never emits while V2 recovers respiration regularity
+     * from R-R: one field pair reported ~46 min REM here against hours on the screen for the same night
+     * (#2365). The caller's line names both stagers for that reason (#2366); do not restore the claim
+     * that they are one.
      */
     fun remFunnelDiagnostic(
         start: Long, end: Long, grav: List<GravitySample>,

android/app/src/main/java/com/noop/testcentre/AndroidDiagnostics.kt

@@ -414,7 +414,18 @@ object AndroidDiagnostics {
                 return@runCatching
             }
             com.noop.analytics.SleepStager.remFunnelDiagnostic(session.startTs, session.endTs, grav, hr, rr, resp)
-                ?.let { add(it.summary) } ?: add("REM funnel: insufficient motion data (<2 gravity samples)")
+                ?.let {
+                    // The funnel replays the V1 classifier, but the shipped hypnogram is staged by V2
+                    // whenever the default-on flag says so — name both, or the two totals read as one
+                    // fact disagreeing. On a 5/MG the gap is maximal: V1's primary REM gate needs the
+                    // raw resp channel that hardware never emits, while V2 recovers respiration from
+                    // R-R. Suffix byte-identical to the Swift twin in DebugDataDiagnostics.
+                    val screenStager =
+                        if (com.noop.ble.PuffinExperiment.from(context).experimentalSleepV2) "V2" else "V1"
+                    var summary = it.summary + " · funnel replays V1; screen staged by $screenStager"
+                    if (screenStager != "V1") summary += " — totals can differ"
+                    add(summary)
+                } ?: add("REM funnel: insufficient motion data (<2 gravity samples)")
             val det = com.noop.analytics.DetectedSleep(
                 start = session.startTs, end = session.endTs,
                 efficiency = session.efficiency ?: 0.0, stages = emptyList(),