fix(core): validate replay operation identity

[zhongkechen]

Summary

• validate checkpoint type, subtype, and name before operation-specific replay handling
• fail mismatches with NonDeterministicExecutionError instead of consuming another operation's checkpoint
• cover type, subtype, and name drift through unit and end-to-end regression tests

Testing

• 1,587 core non-e2e tests passed
• 46 core e2e tests passed
• hatch run dev-core:typecheck
• hatch fmt --check

Closes #692

[github-actions]

Codex AI review · Finding arf_v1_dgw6ii4qxklkvjhea4ieegv2i6

[P1] Preserve fatal errors reported after early completion

Once a completion policy ends the coordinator loop, the queue is drained only once and the pool shuts down with wait=False. A running branch reaching this catch afterward leaves an unconsumed fatal event; ExecutionState.close() joins the pool without inspecting its futures, so map/parallel can checkpoint success despite nondeterministic history. Retain fatal state that is checked before the parent terminal checkpoint, and add a barrier-based early-completion regression test.

[github-actions]

Codex AI review · Finding arf_v1_7pptyptccbf3r2g5kacbf3pjp6

[P1] Reject NESTED branch checkpoints when replaying in FLAT mode

This validation runs only for non-virtual branches. When a STARTED or ReplayChildren parent is replayed after changing NESTED to FLAT, the old successful branch checkpoint has the same id, type, subtype, name, and parent; child_handler consumes its cached result without executing the virtual branch, so the changed inner hierarchy is never validated. In FLAT mode, reject any existing branch-container checkpoint before calling child_handler, and add a real map/parallel NESTED→FLAT replay test.

[github-actions]

Codex AI review

Two P1 replay-safety gaps remain in concurrent map/parallel handling. Tests do not cover early-completion races or actual NESTED→FLAT replay.

Reviewed commit d8f9c9846eb9aef8367324e72bf72f55f170293b. Workflow run