prove-the-setting-workslisted

# Prove the Setting Works ## Overview A config key's NAME is not evidence of its BEHAVIOR. Neither is its docstring, nor a rule you wrote in memory six months ago. Two symmetrical failures keep recurring: - **False positive:** "Set `enabled: false` and the row disappears." It loads without error, looks applied, and silently does nothing. - **False negative:** "That can't be done with config." Said after reading one code path, when a different lever would have worked. Both are cured by the same three-minute move: **mutate a throwaway copy, run the real code path, read the real output.** The rule: _never describe a config change's effect you have not observed._ ## When to Use - Before promising a config-only fix ("I can hide that with a setting"). - Before declaring a limit ("that needs an upstream change"). - Before removing a config block that something else might reference. - When a standing rule in memory would block an approach — test whether the rule is actually as broad as it was written. - After any config edit, to confirm the observable behavior changed. ## The sandbox probe Never edit the live config to find out what a key does. 1. **Copy the config into a throwaway home.** ```python TMP = pathlib.Path(tempfile.mkdtemp(prefix="probe_")) shutil.copy(SRC / "config.yaml", TMP / "config.yaml") ``` 2. **Mutate exactly one lever** with a round-tripping parser so unrelated formatting is preserved: ```python y = ruamel.yaml.YAML(); y.p