Tune cluster liveness polling cadence

[AlexCheema]

Why

Snapshot/state reconciliation gets workers caught up correctly, but stale topology still affects placement and dashboard freshness. The original branch reduced the time it takes to notice dead nodes and connection changes. This PR keeps that behavioral tuning isolated from the event-sourcing changes.

How

• Master now checks node inactivity every 1s instead of every 10s.
• Master considers a node inactive after 5s instead of 30s.
• Workers poll connection reachability every 2s instead of every 10s.
• The values are local variables in each loop so future tuning is explicit.

Tests

• uv run pytest
• uv run basedpyright
• uv run ruff check
• nix fmt