# E86 Atomic Kill Fence

## The E85 defect

E85's 500-worker stage recorded **328 commits after the engage request**, p50 stop latency **2.56 s**, and one stage whose effect/done reconciliation was false. A worker could pass the kill check, then commit after another process engaged: a check-then-commit window.

## E86 semantics

`FencedLedger` puts the epoch read and the effect insert in **one `BEGIN IMMEDIATE` transaction**. `engage()` increments the epoch in the same kind of transaction. A commit that observes a newer epoch is refused and writes nothing.

- **Stop-new-work:** after the engage transaction commits, every new commit is refused.
- **In-flight:** a commit already holding the write lock may complete **before** the engage transaction acquires it. It is reported separately (`in_flight_effects_before_effective`), not counted as a post-stop effect.
- **Evidence:** every engage returns `requested_at` and the effective `at`; effect timestamps are compared to the effective time.

## Measured

`artifacts/e86/killfence/lab.json`: 50 real worker processes, 2,000 commits, **109 committed before the effective stop, 1,891 refused after, 0 effects after the effective engage**, 3.9 s wall clock.

**Limits:** one host, one SQLite file, synthetic effects, no network partition, not the E85 MissionStore itself, no independent witness. The E85 500-worker result remains PARTIAL; E86 proves the corrected reference boundary at 50 processes and should be integrated into the mission store next.
