fix(alerts): a silence ends when Steward hears of it, once (#5720)

An agent's flushed backlog brings samples hours old. The resume was stamped
at the first old sample's time, inside the silence it ended, and silence was
still judged by the backlog's sample times, so the series was marked silent
again between posts and resumed a second time. A resumed series now counts as
heard at ingest time, its events from that batch are stamped no earlier, and
the rest of the backlog reads as late samples.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

steward/core/alerts.py

@@ -250,9 +250,14 @@ async def record_metrics(session: AsyncSession, rows: list[PluginMetric], phases
                          key=lambda r: r.recorded_at)
         muted = _muted_by(windows, rule)
         resumed = _track_samples(alert_state, samples, now)
+        # A silence ends when Steward hears of it: an agent's flushed backlog
+        # brings samples hours old, and stamping their events at their own
+        # times put the resume inside the silence it ends (#5720).
+        back = resumed is not None
         for sample in samples:
+            at = max(sample.recorded_at, now) if back else sample.recorded_at
             notifications = await _evaluate_rule(session, rule, alert_state, sample.value,
-                                                 sample.recorded_at, note=muted, resumed=resumed)
+                                                 at, note=muted, resumed=resumed)
             resumed = None  # said once, with the first sample back
             if muted:
                 continue
@@ -319,7 +324,13 @@ def _track_samples(state: AlertState, samples: list[PluginMetric], now: datetime
     series had been marked no_data, clear the mark and return the note saying
     so ("data resumed after 15m silent"); the first sample's evaluation
     records it, in one event with whatever that sample does to the rule
-    (#5442)."""
+    (#5442).
+
+    A resumed series counts as last heard `now`, when Steward heard it, not at
+    its newest sample's time: a backlog flushed a few samples a post is hours
+    old, and judging silence by it marked the series silent again between
+    posts, so one silence was resumed twice (#5720). The rest of the backlog
+    then reads as late samples, which teach the cadence nothing."""
     for sample in samples:
         if state.last_sample_at is not None:
             if sample.recorded_at <= state.last_sample_at:
@@ -331,6 +342,7 @@ def _track_samples(state: AlertState, samples: list[PluginMetric], now: datetime
         return None
     note = f"data resumed after {human_duration((now - state.no_data_since).total_seconds())} silent"
     state.no_data_since = None
+    state.last_sample_at = max(state.last_sample_at or now, now)
     return note