feat: durable missions with typed waits and explicit side-effect reconciliation #3518

Capability gap

A workflow can persist a status, but long-running work still needs a durable execution boundary: wait for a typed human/external signal, restart after a process stops, and avoid repeating an uncertain external side effect.

Proposed first slice

• Persist a small dependency graph of agent, signal, and external-action steps with an inspectable transition log.
• Dispatch ready agent steps and expose typed signal waits through MCP tools.
• Issue external action intents with stable idempotency keys and require a receipt to continue.
• If an agent call or external action may have happened before a crash/timeout, block automatic replay until explicit reconciliation.
• Write mission state atomically and restrict local storage permissions.

Acceptance

A focused test creates a mission, runs an agent step, waits for approval, resumes after persisted state reload, and finishes without rerunning completed steps. Timeout/crash simulations leave the uncertain step blocked. External actions need a matching key and receipt before downstream work runs. Concurrent callers cannot both claim the same pending agent step.

Limits for this slice

This is a local-file runtime, not a distributed exactly-once system. A provider's idempotency key only suppresses duplicates if that provider honors it; otherwise the operator must reconcile against the provider before retry.