Tie the README's version claim to the table it summarises

[sethbergman]

The README's Roadmap section said "Everything through v0.14 has shipped"
while v0.18 had. Four PRs each added a row to the Shipped table in
docs/roadmap.md and left the sentence alone — so the first claim a
reader meets, in the one section whose purpose is to say what is done and
what is not, was four releases stale.

Nothing caught it because nothing was looking. tests/docs-index skips
README.md by name, and markdownlint has opinions about the line rather
than the fact on it. This is the failure docs/README.md is generated to
avoid, arriving somewhere too small to generate — so it is asserted
instead.

What the check does

tests/lint/check_version_claim.py, wired into the existing lint suite
so no new CI job is needed:

1. The README's through vN.N has shipped claim must name the newest row
   of the roadmap's ## Shipped table.
2. That table must be no older than the newest git tag.

The table is the source of truth rather than the tags because of the
order releases are cut in here: the roadmap row lands in a PR and the tag
is pushed after it merges. Gating the README on tags would fail that PR
for being correct. So the tag half is one-directional — a table ahead of
the tags is the normal state, a table behind them is a release nobody
wrote down.

The lint job now checks out with fetch-depth: 0. Without the tags that
half would report a skip, and a skip in CI is indistinguishable from a
pass: decoration standing exactly where a reader assumes a guard, which
is the thing this check exists to stop. The only remaining skip path is a
checkout with no .git at all, which CI never has.

Mutations, each watched to fail

Mutation	Caught by
README one release behind (v0.17)	claim ≠ table
README reworded to "The latest release is v0.18"	no claim found — a reword cannot silently disable the check
README ahead of the table (v0.19)	claim ≠ table
v0.18 row deleted while v0.18 is tagged	the tag half; the README half stays green
## Shipped renamed to ## Releases	parse finds nothing, refuses to run rather than passing
git repo with no tags	fails with git fetch --tags rather than skipping

The fourth is the one that matters: it is caught by an assertion that
does not name the README at all. Numeric version sorting is confirmed by
the passing run itself — a lexicographic max over the table would pick
v0.9, not v0.18.

Also

• README.md corrected to v0.18.
• The lint suite's header and CI job name widened from "Shell invariants"
  to repo invariants; the scope is the invariant rather than the language
  it is written in.
• docs/roadmap.md gains a fifth entry under "What 'done' means here".
  It is framed as different in kind from the four above it: those are a
  test that existed and was weaker than it looked, this one had no
  assertion at all. The fix differs too — a weak assertion gets widened,
  an unchecked claim gets tied to whatever is already correct.
• CLAUDE.md records the invariant and why the job fetches tags.

Job count is unchanged at 38, so check-ci-coverage needs nothing.

🤖 Generated with Claude Code