# Evolution 81 — Mission Durability Report

## Implemented

Durable mission state, atomic checkpoints (SHA-256 chain), replay-safe resumability, idempotency
keys, duplicate-side-effect prevention, transactional transitions, leases, fencing, scheduling,
ownership transfer, cancellation, expiration, versioned replanning, dependency recovery,
dead-letter handling, bounded queue/backpressure, graceful worker replacement, recovery
verification and mission lineage.

## Evidence

| check | result | source |
|---|---|---|
| process SIGKILL + replacement | `True`, leases recovered 1 | `recovery_drill.json` |
| dead worker late commit | `STALE` | `recovery_drill.json` |
| duplicate effects | 0 | `recovery_drill.json` |
| checkpoint chain | `True` | `recovery_drill.json` |
| coordinator restart sees state | `{'DONE': 6}` | `recovery_drill.json` |
| reference mission | `COMPLETED_WITH_UNVERIFIED_ITEMS`; tasks `{'DONE': 6}` | `mission.json` |
| cross-host lost acknowledgement | duplicate suppressed `True`, effects for lost task 1 | `cross_host.json` |
| host reboot | **NOT RUN** | — |

## Why this is not a host-restart proof

A host reboot is a different failure domain: the kernel, the disk cache, the container runtime and
the process supervisor all restart. This evolution killed a worker process and restarted the
coordinator inside one host. E72 rebooted a development replica; that evidence is preserved but was
not re-run here. G07 is **PARTIAL** by construction.

## Long-horizon support

The design supports long missions: state is durable, checkpoints are chained, and a resumed process
reuses completed work. A 1-hour real campaign was run (see `EVOLUTION_81_ENDURANCE_REPORT.md`).
Thousands of hours are **not** claimed; only the infrastructure exists.
