feat(metrics): resync seq-gap histogram, drop per-resync logs

[mikhail-dcl]

Problem

Resync fallback to STATE_FULL for subject … (lastKnownSeq=…, gap=…) fired once per served resync at Warning, with a matching Information line beside it — too noisy to leave on, and the only useful signal in it was a single number.

Change

That number is now a histogram: latestSeq − knownSeq recorded per served resync, bucketed dense around Peers:SnapshotHistoryCapacity (the ring-eviction cliff that decides targeted-delta vs STATE_FULL), with a dedicated 0 edge so "client was already current" is distinguishable from "one publish behind".

Split into delta / full outcomes, so the fallback rate the old LogWarning conveyed reads off the full series' own _count — no extra counter needed.

• Console dashboard: one merged Resync Seq Gap row in the Latency group with the usual P50/P95/P99 window+lifetime columns, sparkline on window P99. Per-outcome split is Grafana-only, same convention as Peer RTT.
• Prometheus: dcl_pulse_resync_seq_gap{outcome="delta|full"} native histogram.
• The sibling per-request log in ResyncRequestHandler drops to Debug — it fires 1:1 with the removed pair, so leaving it at Information would have kept the console just as noisy. Still available when chasing a single client.

Follows the delta_staleness / peer_rtt pattern end to end; ResyncOutcome/ResyncOutcomes mirrors the existing ConnectionClass/ConnectionClasses idiom.

Note for review

ResyncRequest.KnownSeq is unvalidated client input — FieldValidator has no rule for it. The first cut used a serial-number cast, (int)(latestSeq - lastKnownSeq), to clamp a baseline ahead of the latest publish. That only disambiguates within 2^31: a client sending known_seq = 3_000_000_000 against a small latestSeq recorded a 1.29-billion sample, swamping _sum and the +Inf bucket and faking the exact "sustained loss / stalled client" signal the docs tell operators to watch for.

Now the ordering is tested before subtracting. Seq doesn't wrap within a session — it's server-assigned from 0 and reset to the sentinel on disconnect, the same assumption every other Seq comparison in PeerSimulation already makes with plain </>.

Testing

660 tests pass. New coverage: gap magnitude on both outcomes, outcome routing in the collector, Prometheus label/bucket shape, the already-current case, and four TestCases across the ahead-baseline range (two of which fail against the pre-fix arithmetic).

🤖 Generated with Claude Code