P1: add durable replay and evidence coordination store

[safal207]

Parent

#216

Goal

Replace process-local replay and workflow state with durable atomic coordination suitable for retries, restarts, and multiple agents.

Required operations

try_reserve
release_if_unattempted
mark_succeeded
mark_failed
get_state
append_evidence
read_current_checkpoint

State model

NEW → IN_PROGRESS → SUCCEEDED | FAILED
        ↓
       NEW  only before executor reachability

Acceptance criteria

• atomic compare-and-set semantics across processes;
• stable logical action and attempt identities;
• crash recovery for abandoned reservations;
• TTL/lease behavior documented and tested;
• terminal states cannot be silently reopened;
• pre-execution transient failures can become retryable;
• append-only evidence and state-transition receipts;
• stale writers cannot overwrite newer checkpoints;
• reference adapter for LiminalDB or a storage-neutral interface;
• deterministic tests for concurrency, crash recovery, replay, and lease expiry.

Dependencies

• P0: make guarded-merge reservation observation atomic #217 defines the in-process atomic contract;
• P1: add causal CI/CD lifecycle event envelope #219 defines lifecycle events.