fix: restore live-decode regressions guards

lib/src/control.dart

@@ -190,10 +190,14 @@ RealtimeHr? parseRealtimeHr(Uint8List inner) {
   final n = inner.length > 9 ? inner[9] : 0;
   if (n > 0 && inner.length >= 12) {
     final v = u16(inner, 10);
+    // 200-2500ms = 24-300bpm; was relaxed to `v>0` without comment, which lets
+    // corrupted/out-of-range timing noise straight into live RR — reverted.
     if (v >= 200 && v <= 2500) rr.add(v);
   }
   if (n > 1 && inner.length >= 14) {
     final v = u16(inner, 12);
+    // 200-2500ms = 24-300bpm; was relaxed to `v>0` without comment, which lets
+    // corrupted/out-of-range timing noise straight into live RR — reverted.
     if (v >= 200 && v <= 2500) rr.add(v);
   }
   final wearing = inner.length > 18 ? inner[18] == 1 : true;

lib/src/live.dart

@@ -206,6 +206,11 @@ _Motion _r10Motion(ByteData view, int len) {
   const activityFloor = 0.05;
   if (std < activityFloor || n < 24) return _Motion(activity, 0);
 
+  // NOTE: was bumped to 25 without comment/test coverage; reverted — no test
+  // (parity or otherwise) exercises steps_inc/activity from R10, and ±25
+  // approaches the 7-40 sample autocorrelation lag range this detrend feeds,
+  // which risks attenuating genuine step periodicity at slower cadences.
+  // Re-widen only with a real-capture regression test backing it.
   const w = 9;
   final x = List<double>.filled(n, 0);
   for (int i = 0; i < n; i++) {
@@ -292,6 +297,11 @@ DecodedSample? decodeRecord(String hex) {
       hr: d.hr,
       activity: 0,
       stepsInc: 0,
+      // skinContact is contact QUALITY, not wear (see protocol notes) — the
+      // parity oracle proves it: 28 real v24 records have hr=87-97 (clearly
+      // worn) with skinContact<=50. hr>0 is the only signal in this fixture
+      // that agrees with ground truth on every case; keep it even though it's
+      // always false for v25 (hr deliberately undecoded there — see parseR24).
       wristOn: d.hr > 0,
       recType: recType,
     );