test(cluster): fix flaky sync hang via lossless watch; adopt store update

[benitogf]

Summary

Fixes intermittent cluster-sync misses under heavy load by adopting the store's lossless real-time delivery, and makes the cluster sync test deterministic. Closes #73.

What changed

• Adopt lossless real-time delivery. The cluster sync test harness opts into the store's lossless watch mode, so every committed write reaches every live subscriber instead of occasionally being dropped when a consumer stalls under load. A guard fails the test immediately if any event is ever dropped. The sync test's own synchronization is unchanged — the dropped delivery, not the test, was the flake (it manifested as a ~1/300 hang under high-iteration -race).
• Stop a background watch-worker leak. The instance ties its storage watchers to its lifetime, so they stop on shutdown instead of lingering across runs.
• Adapt to the dependency's removed request gate. The store replaced its built-in request gate with standard router middleware. Permissive test hooks drop to the new default (open), and the one hardened-node probe test moves to middleware.

Verification

• /flaky-check: 300 cumulative full-suite -race iterations clean against the released dependency (and 540 against the same fix before release), zero dropped events. Full suite green under -race.

Compatibility

• Picks up a store release that removes the built-in request gate (now middleware-based). Only test code referenced it in this repo.

---

🤖 Generated with Claude Code

[CBosch101]

Adopts the lossless-watch ooo release to kill the ~1/300 cluster-sync flake and ports the removed Server.Audit gate to router middleware. Mostly mechanical test adaptation; the one production change is sound. Approving.

Verified

• go build ./... + go vet ./... clean against the bumped ooo; TestClusterSync* and TestPingNodeProbesUnauthedEndpoint pass under -race.
• instance.go:399 threads i.ctx into WatchWithCallback — ctx is set in Setup and cancelled in Shutdown, so attached storages' watch workers now stop on shutdown instead of leaking.
• Bumped ooo actually provides the API this PR relies on: LosslessWatch/OnDroppedEvent fields and the ctx-first WatchWithCallback signature; Server.Audit is genuinely removed (the dropped allow-all test hooks and the nodehealth_pingnode_test.go Router.Use port match that).
• Issue #73 acceptance: lossless delivery + OnDroppedEvent guard that fails on any drop, dependency updated, suite builds/passes adapting to the removed gate — all addressed.

The fix is structural (block-instead-of-drop + a guard that fails the test if any event is ever dropped), not statistical, so it targets the documented root cause directly. I ran the cluster/ping tests once under -race, not the full 300-iteration sweep.