Writing down what was only in someone's head

A pass over the durable records rather than the code, on the principle the repository already states: the decision records are the memory and the conversation is not.

docs/BACKLOG.md brought up to date. Two entries claimed problems that had been solved - including one asserting no emulator implements a write path, which was wrong rather than stale (the path was there; every import was typed STT_NOTYPE so no loader could match it). Added: the unbuilt NID cracker, the unrendered responsiveness records, the uncharacterised churn threshold, condition variables and why they were deferred, the 289 names still presence-only, and the absence of version control.

CLAUDE.md gained the procedure for adding a check. Five steps, three of which are places the build fails in a confusing way if skipped - a name cannot be in both the census and platform.h, mkmodule will not guess a library, and regenerating the census undoes step one. The fifth is "run it on the host first", which has now caught two wrong checks: a probe whose two "must differ" inputs did not, and a set of checks ordered before the capability they required.

Also recorded there: that anything which can block is written as the try form or not at all, and that a looping check reports progress. Both were learned by losing a run.

Two small things finished

pretty now renders the responsiveness map, naming the stubs rather than counting them - "nineteen stubs" is a statistic, "sqrt, fabs, floor, ceil are stubs" is a list someone can act on. It prints above the totals, because it changes how every failure below it reads.

OBS_THREAD_CHURN is overridable so the crash threshold can be bisected. The first attempt found nothing and that is recorded as a null result with the reason: twelve runs at a one-in-seven rate is not enough to see anything, and the next person should budget twenty per point.