Plan 4: the prose kit, corrected by executing it

[robertblust]

Adds docs/superpowers/plans/2026-08-31-prose-kit.md, alongside plans 1, 2 and 3. Kept out of the
three adoption pull requests for the same reason as before: those are exactly the files each site
needed, and this document describes all three sites and the tiers still to come.

The implementation is in robertblust.github.io#40, companygraph.github.io#42 and
guestgraph.github.io#24, with the package released as
v0.4.0. All green.

It corrected the spec before writing a line

the spec said	what measuring found
the kit is "54 lines common to all three sites" across 16 pages	the 54 was measured across the three privacy pages. Across all sixteen, eight rules are byte-identical
its content is .card .cols .lede .note .rules h1	.card, .cols and .note are a privacy-page pattern. They live in the tail of 141 single-page rules
one file, prose.css	the shared CSS is not contiguous — a reset at the top, a footer near the bottom, per-page material between. Two fences, not one

Five corrections made while running it

the plan said	what is true
a page that loses a fence is invisible to everything	design:check stays green; verify does not. Deleting the language fence gives eight ReferenceError: langFromUrl is not defined, a mobile menu that will not open, and a failed language handoff — naming none of them a fence. The value is the diagnosis, not the detection
Tasks 8–10: "no change beyond the marker lines"	all ten credit pages carry a four-line prose comment the package folds into its header. The gate would have read the intended outcome as drift and stopped three tasks
npm install --package-lock-only	moves the lockfile and leaves node_modules on the old release, so the sync writes the previous version's bytes while the lockfile claims the new one
two of the test specifications	both buggy — a blank-line regex spanning the header comment, whose paragraph breaks are blank lines in every block, and a substring filter missing two wrapped continuation lines. The rewrites are stricter than what I wrote
ruling 5: the brand mark lost its accent, and restoring .brand svg{color} fixed it	the mark consumes no currentColor. It is <rect class="plate"> + <text class="rb">, filled directly. It never lost anything, the declaration is inert on that site — and the check written to prove otherwise measured getComputedStyle('.brand svg').color, the property the edit had just set

That last one is the one worth reading twice. The plan's first three tasks exist to hunt exactly
that shape of check — one that cannot fail — and I wrote one three tasks later, in the step meant to
verify the only visible change I believed the plan was making. It was caught because the final
reviewer rendered the mark instead of reading the diff.