fix(sandbox-cache): broadcast state transition event from StartRemoving

[AdaAibaby]

Summary

StartRemoving atomically writes a new sandbox state to Redis via a Lua script (e.g. Running → Killing), but never called publishSandboxEvent. Every allocation's in-process sandbox cache (introduced in #3593) continued to serve the old Running state until the sandbox was eventually deleted and a remove event arrived.

Symptoms:

• TeamItems(states: [Running]) returned sandboxes already mid-removal
• Metrics and user-facing sandbox list showed inflated Running counts during high-eviction periods

Root Cause

Three code locations form the gap:

1. StartRemoving — Lua write succeeds, no event published

// state_change.go
written, err := startTransitionScript.Run(ctx, s.redisClient, ...)
// Redis now has State=Killing — but no publishSandboxEvent call here
return updated, false, s.createCallback(...), nil

2. createCallback — publishes routing key, not a sandboxEvent

s.publisher.Publish(cbCtx, getTransitionRoutingKey(...))
// payload: "lock:sandbox:storage:...:transition:<uuid>"

3. dispatch — routing key never reaches cache.apply

if isSandboxEvent(payload) { // strings.HasPrefix("{") → false for routing keys
    m.cache.apply(evt)        // never reached
}

Fix

After startTransitionScript.Run() succeeds, publish a sandboxEvent{update}:

s.publisher.publishSandboxEvent(ctx, sandboxEvent{Op: sandboxEventOpUpdate, Sandbox: &updated})

dispatch() routes it to cache.apply, so all allocations immediately reflect the intermediate state.

createCallback needs no extra publish:

• Terminal transitions: Remove() broadcasts a remove event
• Transient transitions: restoreToRunning calls Update() which broadcasts the restored Running state

Tests

Two new integration tests (real Redis via testcontainers):

• TestStartRemoving_CacheBroadcastsTransitionState — Kill transition is visible as Killing in cache before Remove is called
• TestStartRemoving_CacheBroadcastsTransientTransition — Snapshot transition visible as Snapshotting, restored to Running after callback

Related

• Fixes sandbox cache: StartRemoving state transition not broadcast, all allocations see stale Running state #3595
• Depends on perf(api): add per-allocation sandbox cache to eliminate TeamItems MGET bandwidth #3594 (per-allocation sandbox cache)