The harness compared these tools against the code they replaced — Python
imported from book-pipeline, podcast-pipeline and knowledge-mcp, and
JavaScript copied out of the podcast media server and indexer, each pinned to
a commit. All of it has since been deleted, so the live half cannot run and
the golden half pins answers against a source nobody can consult.
What was worth keeping was never the comparison. `KNOWN.md` argued, at
length, why a dozen of these tools deliberately do something a reader would
not expect: why pdfgutter measures the crop box throughout, why avplan finds
the encoder rather than reading one out of the environment, why mdscrub will
not manufacture a page boundary. Those arguments are about this code and
stand on their own, so they are now `docs/decisions.md`, rewritten to say
what each tool does and why rather than what it does differently.
Binding coverage is unaffected: the bindings carry 24 tests of their own, a
node smoke test against the built package, and a typecheck.
Cuts a reference list into entries — at the note numbers where there are any,
at the layout where there are not — and reads each as far as its punctuation
allows. Fields that would not parse are named in `unsettled` rather than
guessed at, and `--prompt` asks a model about exactly those and no others.
Over 202 real reference sections from the corpus: 8,566 references, 95% with
a title, 82% with a year, 46% settling every field with no model involved.
A title is refused where the entry was never divided, so no record reports
its own text back as the work it cites.
`ibid.` is flagged and left alone — it points at a work, and the entry does
not say which. A rule of dashes is filled in, because the style defines it as
the author above.
chunkmd gains `note_starts` and `cites_something`: recognising a page of
notes and splitting one are the same knowledge, and two copies of that
pattern would drift.
funkpower