[checkpoint] Declare and test LongMT restart-only actor pools

Parent

Child of #3223 and #3024.

Shared state

resources_servers/longmt_eval/app.py owns a lazily initialized SegALE/Ray evaluator actor pool, initialization flag, lock, and reconstructible caches rather than per-rollout semantic state.

Proposed capability

Advertise restart-only dependency recreation, rebuild and health-check actors before resume, and keep in-flight verification outside a published checkpoint.

Do not serialize futures, locks, credentials, waiters, or raw external handles. Capability reporting must distinguish coordinated-global, drain-only, and restart-only semantics from ordinary per-session export/restore.

Acceptance criteria

• Equivalent verification succeeds after cold actor recreation; stale actor references and initialization failure surface clearly without claiming export/restore.
• Prepare accounts for every accepted shared operation and fails closed at its deadline.
• Process replacement leaves no stale waiter, batch member, metric interval, or actor reference.
• Duplicate prepare/restore/resume is idempotent or rejected deterministically.
• Cold restart and failure cleanup are covered.

Start in resources_servers/longmt_eval/app.py and trace the evaluator actor pool, initialization flag, lock, and reconstructible caches. Read the parent issues #3223 and #3024 for surrounding design, then run the existing evaluation checks before defining coverage for cold actor recreation, deadlines, process replacement, duplicate operations, and failure cleanup.