fix: [sc-89444] Derive per-cycle context so in-flight commands cancel on disconnect

[mlataza]

Summary

• A single root context was created once at service startup and reused across every MQTT reconnect cycle. Hung commands could not be cancelled on disconnect, leaving worker goroutines blocked and consuming slots after reconnection.
• runCycle now derives a fresh cycleCtx from the parent context on each invocation. The cycle context is cancelled when the cycle ends (disconnect or stop), which propagates through processMessage → executor.Execute → exec.CommandContext to terminate any in-flight command.
• Deferred cleanup ordering is updated to MQTT teardown → cycleCancel() → close(msgQueue) → wg.Wait(). Cancelling before wg.Wait is required so a hung command cannot block worker drain.
• Commands started in a later cycle bind to that cycle's own context and are unaffected by the previous cycle's cancellation.

Test plan

• go build ./...
• go test ./cmd/agent_smith/... (existing suite green, including the worker/queue/lost-channel regression tests)
• Manual: run a 60s sleep command, force a network drop, verify the command process exits and the worker slot is reclaimed before the next command arrives on reconnection

🤖 Generated with Claude Code