Race: concurrent discovery can overwrite fresher registry for same context

[ndenev]

Summary

Concurrent resource discovery for the same context can race and overwrite fresher registry data with older results.

Affected code

• src/kubernetes/client.rs (discover_resources_for_context)

Root cause

discover_resources_for_context has no per-context synchronization.
Multiple in-flight discoveries for same context can run concurrently and each performs a blind final write:

• registries.insert(context, CachedRegistry::new(registry))

The last finisher wins, regardless of freshness/force intent.

Impact

• Potential stale/older registry replacing fresher one.
• Non-deterministic table metadata under concurrent operations.
• Correctness/freshness issue (higher priority).

Proposed fix

1. Serialize discovery per context using per-context async mutex/lock.
2. Or apply generation/epoch guard so older discovery results cannot overwrite newer ones.
3. Keep cross-context discovery parallel (only serialize within each context key).

Tests to add

1. Simulate concurrent force and non-force discovery for same context; ensure final registry corresponds to freshest/force result.
2. Regression test for deterministic final registry under concurrent discovery calls.