Workload old-generation stop recovery
This bounded model isolates the durable effect around stopping the old application deployment during M5. It requires full target health and an observed mixed route before reserving the stop effect. It covers a stop accepted by Nomad whose HTTP response is lost, recovery by the intent marker, allocation draining, and completion only after every pinned old allocation is terminal.
The model was rerun with the repository's statechart skill on 2026-11-08; TLC checked 10 states and 15 steps. It checks the target-health/mixed-route prerequisite, one stop write per intent, marker uniqueness, receipt binding, recovery without another stop write, terminal-allocation completion, and deadlock freedom. The deterministic Go fixture additionally closes the HTTP response after Nomad accepts the stop, then recovers the same durable effect and verifies there was only one CAS registration.
This is still a bounded abstraction. It does not model PostgreSQL/etcd effect-store crashes, signature cryptography, route publication, runtime supervision, or live Nomad behavior. The worker now dispatches signed app.workload-release operations, binds them to the app lock, and joins target-health readback, route publication, old-stop/readback, and completion receipt. Focused tests cover each durable seam, and a driver-level lifecycle fixture now injects lost responses after stop, drain, and other durable boundaries before retrying the same reconciler. A production-driver process/lease crash matrix and live qualification remain open.
Regenerate the chart and report with the repository's statechart skill:
bun /Users/arti/.agents/skills/statechart/scripts/statechart.ts \
"$PWD/WorkloadStopRecovery.tla" \
--config "$PWD/WorkloadStopRecovery.cfg" \
--view "$PWD/view.ts" \
--out "$PWD/reproduced-output"The output directory must not already exist.