D069 - Volatile facts are generated; static facts are written. The distinction is the whole fix

Status: decided, on the observation that most of a review's findings were drift.

A count in prose has no mechanism to stay right. The README said 79 checks when there were 106, 326 censused symbols when there were 312, and screen.c said "fifteen rows, which fits with room to spare" while drawing twenty-one rows off the bottom of the framebuffer. None of it was carelessness - every number was correct when written.

The fix is not proofreading. It is deciding, per fact, whether it moves:

Things that change every working day - check counts, census size, the provenance split

• belong to the code. scripts/counts.py renders them into marked regions and --check fails when they have drifted. Two rendering sites, one source of truth, and no way to forget.

Things that are settled - the hash suffix, the dynamic tag values, why -fno-plt was removed - stay hand-written, and should. Generating those would be machinery for nothing.

The same split resolves the duplication question underneath it: the numbers appeared in both README.md and BACKLOG.md, which is two places to update and two places to drift. They now render from one computation.