#817 DB failover breaks durability: duplicate side effects (exactly-once) + run failures under postgres restart (#agentic)

Chaos harness with --db-restarts 3 (postgres restarts mid-flight) on 20 canary runs: 3 runs FAILED and 26 DUPLICATE transactional side-effect markers (tasks whose commit was in-doubt when the connection dropped were re-executed -> double insert). No state corruption. Two gaps: (1) a task re-run after a commit-in-doubt (connection lost during/after COMMIT) does not dedup on the already-committed checkpoint -> duplicate side effect; (2) some DB-connection errors fail the run instead of being retried durably. Worker-crash chaos (no DB faults) passes cleanly; this is specific to DB connection loss. Found by tests/chaos --db-restarts.