delta-sync poison-node: follow-ups from PR #86 review

Non-blocking observations from sage's review of #86 (poison-node resilience). All deferred as nice-to-haves.

1. Poison-cluster visibility — adjacent poison nodes each trigger independent bisection (~7 fetches each; MAX_PAGE_ITERATIONS caps pathological cases). Add an escalated log line when poisonNodesSkipped > 3 in a single cycle to surface unexpected Tana API degradation early.

2. Watermark-after-poison-only-cycle trade-off — advancing to Date.now() when only poison nodes were skipped (zero real nodes found) means a node changed between the old watermark and now won't reappear in the next delta unless it changes again. Low risk (poison is rare + full sync index recovers). Add a code comment acknowledging the trade-off for future readers.

3. getFailureState() test — add a lightweight poller-test assertion that consecutiveFailures resets to 0 after a successful cycle.

Source: PR #86 review by sage (cortex).

Start with PR #86 and its poison-node resilience changes, then locate the delta-sync cycle, MAX_PAGE_ITERATIONS, poisonNodesSkipped, and getFailureState(). Add the escalated visibility, document the watermark trade-off, and assert the consecutiveFailures reset in the poller test; verify the relevant tests pass and the sync index recovery behavior remains available.