MeshSync: periodic reconciliation to self-heal missed watch events (design)

[leecalcote]

Summary

Add a periodic reconcile that re-lists watched resources and publishes only genuine drift, to self-heal missed watch events. Today the informer resync period is 0 (pure watch); a watch gap silently drifts MeshSync's snapshot from cluster truth until an external resync is triggered.

Key decisions

• A dedicated LIST-vs-Store diff loop, not a non-zero informer resync period - a client-go resync only replays the local cache and cannot detect events missed entirely (including deletes). The loop diffs a fresh LIST against the informer's cache.Store: LIST minus Store = missed add, higher resourceVersion = missed update, Store minus LIST = missed delete (reconstruct the tombstone through the existing DeleteFunc path).
• Publishes only drift (reimplements resourceVersion suppression at the reconcile layer) so it never spams Server; steady state = zero publishes.
• User-tunable interval via the MeshSync CRD, default off, with an enforced floor; jittered and staggered across GVRs to avoid re-list storms.

Risks flagged

A partial/failed LIST must abort that GVR's pass for the tick (never diff a truncated list as a mass-delete). Fixes a pre-existing h.stores unguarded-map race that the loop's concurrent read is the first to expose.

Complementary to JetStream durability (companion blueprint).

Full design

docs/design/fd5-periodic-reconciliation.md (added in #576).