fix: coordinate event-driven mesh refreshes

[LiranCohen]

Summary

• add one single-flight refresh coordinator that durably coalesces subscription, endpoint, manual, retry, startup, and anti-entropy invalidations
• use authoritative stream health to switch between a 30-second fallback cadence and a 5-minute healthy cadence, with bounded debounce, rate-limit-aware jittered retry, and subscribe-before-repair fencing
• preserve the immutable last-good local snapshot, derive peer liveness and expiry locally, and expose scheduler/stream health through the daemon status API
• remove the redundant second startup DWN load by replaying an isolated local map after the engine event bus is ready
• keep role-only topology streams on explicit polling fallback until the scalable observer stream in feat: add scalable MessagesSubscribe mesh invalidation stream #255 lands

Closes #258.

Validation

• go build ./...
• go vet ./...
• go test ./... -count=1 -race
• git diff --check

All pass locally.