A patch, and both entries are the same act: take something this repository says about itself, ask what would fail if it stopped being true, and find out by emptying it. Ten claims and six pages had the same answer — nothing.
Added#
Six more pages could have been deleted in silence. The probe that found
docs/doctrine.md unguarded — empty a source and re-run the suites — was run
over every prose page in the repository. Six broke nothing at all:
docs/ci.md, docs/running.md, docs/accounts.md, docs/authoring-rules.md,
SECURITY.md and VERSIONING.md. The last two are the ones worth
naming: one tells somebody how to report a vulnerability, and the other defines
what this project's three version numbers mean, which every release depends on.
Six bespoke guards would have left the seventh page unguarded, so
packages/core/test/every-page.test.js is derived from the filesystem. It holds
every page to the three things a page goes wrong about quietly: it says
something (a page gutted to a heading fails), it links only to files that are
there, and it shows only commands this CLI dispatches — the last derived
from COMMAND_FLAGS, so the docs and USAGE cannot disagree with the product or
with each other. 30 pages, 219 relative links and 50-odd documented invocations,
all checked.
What that proves is narrow and worth stating: nothing mechanical can check that what a page says is true. But the failure the probe found was not a subtle mischaracterisation — it was six pages nothing was holding at all.
And the guard caught prose on its first run. Executable trazum not found,
quoted in a changelog entry, came back as a command called not. Exempting the
file would have been the wrong fix — a guard that needs an allowlist to stay
quiet is a guard somebody deletes — so the pattern was tightened to what an
invocation actually is: something that begins its command, at a line start,
inside a code span, or after a shell prompt. Both cases are kept as tests.
Four of the five promises docs/json-output.md opens with were prose. Only
absence is null and never zero was enforced, on five fields of one document.
The other four held — measured, not assumed, over 414 dollar figures and 296
token counts across the profile, the roll-up and the prompt draft — which is
exactly the state in which a promise quietly stops being true, because nothing
would say so.
packages/core/test/format-promises.test.js runs all of them over every
document this package can build, and takes the bullets out of the section
first: deleting a promise from the page fails here, so the guard cannot
outlive the claim or the claim the guard.
Rounding is the awkward one, and it is asserted backwards. The absence of
rounding cannot be proved from one document, so what is checked is the shape of
the failure: a document whose dollars had been through toFixed(2) would carry
none with more than two decimals, and this one has to carry a majority.
Every check was proved by breaking the product, not the test. Rounding every
dollar on the way out of profileUsage, returning one as a string, adding a
half to a token count, and putting a session key into the document — each fails.
And rewording the promise section without changing a promise must not, which it
does not.