Three legs against the closed-loop blind spot the duplicate-style bug
exposed: package-lint.test.ts checks invariants the schemas cannot
express (unique style ids/names, dangling style/num/rel references,
balanced ranges, manifest completeness); lo-roundtrip gains a
soffice→fodt leg asserting the styles LibreOffice resolves (Heading
chain Arial/bold/sizes, Word defaults); lo-render-consistency.test.ts
renders the kitchen sink from both formats through LibreOffice and
compares pagination and pixels (ratchet thresholds). CI's LO job
gains poppler-utils; test:lo runs both LO files.
tests/kitchenSink.ts authors one deterministic document using every
feature; schema-validation.test.ts validates its .odt against the ODF 1.3
RelaxNG and its .docx parts (after MCE resolution, as Word reads them)
against the ECMA-376 transitional XSDs — the structural half of Word's
strict reader, which LibreOffice silently forgives. Schemas vendored under
tests/schemas/ (sources and offline patches in its README); self-skips
without xmllint, which CI now installs.
The leg self-skips wherever soffice is absent, so until now it only ran
when someone remembered to locally. A parallel job installs
libreoffice-writer + fonts-liberation and runs npm run test:lo.
tests/smoke/run.mjs boots the production build in Playwright Chromium and
checks what jsdom cannot: the bundle boots, typing and the bold shortcut
reach ProseMirror, autosave survives a reload, a corpus ODT imports and
paginates; any uncaught page error fails the run. Wired as npm run
test:smoke and a CI job (build + chromium install), playwright-core is
now a devDependency.
The parity findings named the six files the corpus is built from, so every
measurement carried a document title into the history. They move to a
gitignored FINDINGS.md beside the corpus itself; the README keeps the harness.
A test that needs a document builds one, so nothing under src/ or tests/
reads the gitignored corpus — CI greps for exactly that.
The root file is read at every session start and had grown to 87 KB, so every
session paid for knowledge it did not use. Move the per-directory detail into
nested CLAUDE.md files, which load when a file in that directory is touched,
and the five deep-dive topics into docs/. The root keeps the commands, the hard
rules and an index of where the rest lives.
Content moved verbatim — the probed Word/LibreOffice behaviour, sentinel order
and shared constants are what the code cannot say. CI fails the root over 120
lines, since a prose budget alone is what let it grow.
Steffen Becker