D085 - Before believing a setting works, set it to something that must visibly break

Status: adopted from the orbistoun side, and applied.

They found their stub policy had been wired to nothing for months: the default and every per-symbol override were consulted nowhere between the configuration file and the guest. Documented-and-true and implemented-and-false at once, invisible to reading the code, and caught by setting the default to a value that must change behaviour and observing that nothing changed.

That is a better version of a lesson this project keeps relearning. verify.sh could not fail for its entire existence; guards.py was written because a rule in CLAUDE.md was enforced nowhere; the multipass stderr hazard was documented in one script and absent from another. Every one of those is the same shape: a mechanism believed to work because it was written down.

Applied to this project's own knobs, none of which had ever been tested for effect:

knob	control	result
EXCLUDE	exclude a check that otherwise passes	skipped, with the right reason
CHURN	build at 3, read the reported count	0x28 becomes 0x3

Both wired. Nothing was found, and that is worth recording rather than skipping: the thread-churn bisect (BACKLOG §6c) rested on CHURN doing something and had never confirmed it, so a negative result there retires a real risk to a conclusion already drawn.