Mecatl snapshot-failure experiment
Prepared 3 October 2026 for quietbybar.com

Prerequisites: Go 1.27.1 (tested), Git, Python 3, POSIX shell.
Unzip into an empty working directory, then run:
  sh reproduce.sh

The script downloads the public repository at the exact pinned commit and Go
dependencies on first execution. The tests do not call an LLM or external API,
require no credentials, and create no Kubernetes resources. It refuses to
overwrite an existing ./mecatl directory. Test code is injected with Go's overlay
facility. Production source files are not modified.

Interpretation:
- Faulted snapshot writes: the engine can emit end_turn after executing the tool.
- Reload and deliberate resubmission with a scripted model: the counter reaches 2.
- Healthy control: completed state and 4 messages reload after one execution.

The second faulted run is explicitly requested by this test. It is NOT an
assertion that Mecatl automatically retries or that real models always duplicate
an operation. There is no process kill, Kubernetes, Redis, server event log,
network timeout, or concurrent-writer test here. The control proves persistence,
not end-to-end deduplication. The counter is only a simulated external effect.

The test file depends on package-local helpers from the pinned Mecatl test suite.
The repository's own license applies to that upstream code. The two supplemental
research tests and reproduction script in this archive are provided under MIT;
see LICENSE.txt. No upstream production source is redistributed in this archive.
