A patch, and nothing installable changed. The arcs are done as far as they can go — six of nine, with the other three open and named — so this is the honest shape of the work that follows a plan: a guard that was checking the wrong thing, the document it was supposed to be checking, and the rule that would have caught both.
Changed#
A rule joined the doctrine, and it is the other half of one already there. Prove a guard by breaking it has been on that list for arcs. The half nobody writes down is and prove it does not fire on anything else — a guard that fails on its defect and on things that are not its defect gets deleted after enough false alarms, and by then nobody remembers whether it was ever right.
Two instances on this project's own record, the same shape both times — a proxy
that correlated with the property until it did not. A docs/releasing.md guard
that matched every quantity word near "manifest" and failed two correct
sentences, and the padding-versus-format guard fixed below. The question that
catches it is not does this fail on the bug but what else does this fail
on, and the cheap way to answer is to run the finished check against the real
thing, whole.
Fixed#
A guard that asserted padding and called it format.
transcript-format.test.js checks that the README's trazum doctor transcript
writes its money column the way the command does. It did that by taking the
~ prefix off the first money line of a live run and requiring every transcript
line to use the same one.
That space is right-alignment, not format. A live run prints ~ $10.59,
~ $8.82 and ~$0.5300 in the same column, because $0.5300 is two characters
wider than $8.82. So the guard agreed or disagreed depending on how wide this
repository's own figures happened to be — and it broke the day a config changed
them, on a change that had nothing to do with the README.
It now measures what the column actually promises: the text starts at the same offset on every row, priced or not, on both sides — the command's block is the yardstick and the transcript is compared against it. Proved by handing it the defect it was written for.
And that defect was real. The transcript's two unpriced rows were indented eleven columns where the command indents twelve, so Below the cacheable minimum sat one space left of the priced rows above it. Found by measuring both, not by reading either.