fix(mcp): get_anomaly shows a quick Steward restart, not only a gap over 90 s (#5862)

A redeploy's new process ticks at once, so its once-a-minute samples never
broke for 90 s and get_anomaly showed none of the day's restarts. The first
self-monitor tick in a process now records restart_gap_secs (time since the
last process's last sample), and steward_gaps shows each one as a restart.

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

steward/core/formatting.py

@@ -298,6 +298,7 @@ METRICS: dict[str, tuple[str, str]] = {
     "ingest_slowest_ms": ("Slowest agent post", "the longest report in the last minute; agents give up at 10 s"),
     "db_roundtrip_ms": ("Database round trip", "a trivial query to Steward's database, connection included"),
     "unwatched_secs": ("Time not watching", "how long Steward wasn't recording before this minute; 0 when unbroken"),
+    "restart_gap_secs": ("Restarted after", "recorded when Steward starts: how long since it last recorded"),
     "cleanup_secs": ("Hourly cleanup took", "how long the last retention and rollup run took"),
     "backup_age_secs": ("Newest backup's age", "time since Steward last sent a backup to its bucket"),
     "backup_bytes": ("Backup size", "the size of the last backup sent"),

steward/core/self_monitor.py

@@ -25,6 +25,10 @@ source and resource:
   • ``db_roundtrip_ms`` — a timed ``SELECT 1``, connection checkout included
   • ``unwatched_secs`` — how long Steward wasn't recording before this tick (a
     deploy, an outage, a stalled loop), from its own last sample; 0 when unbroken
+  • ``restart_gap_secs`` — once, on a fresh process's first tick: how long since
+    the last process's last sample, however short. A quick redeploy never
+    breaks the minute long enough to show as unwatched; this marks it anyway,
+    so get_anomaly can show Steward's own restarts beside an anomaly (#5862)
 
 ``cleanup_secs`` is recorded by the hourly cleanup itself when it finishes.
 """
@@ -51,6 +55,7 @@ except ImportError:  # pragma: no cover - non-POSIX
 FD_WARN_PCT = 80.0
 
 _warned = False  # de-dupe the WARNING so a sustained leak doesn't spam the log
+_started = False  # this process has ticked: restart_gap_secs is recorded on the first tick only
 
 
 def count_open_fds() -> int | None:
@@ -98,7 +103,7 @@ async def record_self_metrics(app) -> None:
     The open-file pair is skipped (logged at debug) where fd accounting isn't
     available, so it's safe to schedule unconditionally.
     """
-    global _warned
+    global _warned, _started
 
     from sqlalchemy import select, text
     from steward.models.metrics import PluginMetricLatest
@@ -129,11 +134,13 @@ async def record_self_metrics(app) -> None:
                     ("ingest_posts", tally.posts), ("ingest_failed", tally.failed),
                     ("ingest_slowest_ms", tally.slowest_ms), ("db_roundtrip_ms", roundtrip_ms),
                     ("unwatched_secs", unwatched_secs(last, now)),
+                    ("restart_gap_secs", None if _started or last is None else (now - last).total_seconds()),
                     ("open_fds", fds), ("open_fds_pct", pct)]:
                 if value is not None:
                     batch.add(source_module="steward", resource_name="process",
                               metric_name=metric_name, value=value, recorded_at=now)
             await batch.record(session)
+    _started = True
 
     if pct is not None and pct >= FD_WARN_PCT:
         if not _warned: