fix: restore live-decode regressions guards

lib/src/live.dart

@@ -248,6 +248,12 @@ _Motion? _r10Motion(ByteData view, int len) {
   const activityFloor = 0.05;
   if (std < activityFloor || n < 24) return _Motion(activity, 0);
 
+  // Deliberately 9, and this branch no longer CHANGES it — main already reads
+  // 9, so there is nothing here to revert. Kept as a note on why not to widen
+  // it: no test (parity or otherwise) exercises steps_inc/activity from R10,
+  // and a ±25 window 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 behind it.
   const w = 9;
   final x = List<double>.filled(n, 0);
   for (int i = 0; i < n; i++) {
@@ -337,6 +343,25 @@ DecodedSample? decodeRecord(String hex) {
     return DecodedSample(
       ts: d.tsEpoch,
       hr: d.hr,
+      // On `wristOn` below: skinContact is contact QUALITY, not wear. 28 real
+      // v24 records in the parity fixture read hr 87-97 (clearly worn) with
+      // skinContact <= 50, so a contact threshold would call them off-wrist.
+      // `hr > 0` is kept for that reason.
+      //
+      // But the parity oracle cannot EVIDENCE it. `decode_parity_cases.json`'s
+      // `wrist_on` column was GENERATED by the same rule, verbatim:
+      //
+      //     wrist_on: d.hr > 0
+      //
+      // so the oracle agreeing is the expression read back, not confirmation.
+      // That line lived in `ts/live.ts`, which this commit removes; it is
+      // recorded here rather than by path so the provenance survives the
+      // deletion, and remains recoverable at protocol main 7edcb3e:
+      //
+      //     git show 7edcb3e:ts/live.ts | sed -n '229p'
+      //
+      // Confirming `hr > 0` needs captures with independently known wear
+      // state, not the fixture.
       // This record family carries no IMU stepping window at all — that is
       // "no usable IMU data", the same absence [DecodedSample.activity]'s
       // doc comment describes for a truncated R10, not a measured 0.0/0.