The five catalog rows and their two fixtures are removed, along with the pinned
baseline constant that existed only for them.
The skill prose converged three rounds ago; every review round since has been a
defect in these cells, four of them mine: two independent needles asserting
co-occurrence rather than a relation, a write-mode grade left inert by a no-write
task, a tightened grade that dropped the verb it was meant to bind, and a fold
grade that would fail correct behavior because the fixture surfaces a new term
where the lifecycle defines folding for an existing entry losing its heading.
The behavioral evidence for this change does not rest on them. It came from
ad-hoc cells run against a real glossary and a real repository subtree, recorded
in the PR description with the transcripts read: the retirement A/B, the
neighborhood-trigger root cause, and the end-to-end run on real source.
Repeatable cells are still worth having and will land separately, designed from
the correct fold semantics and grading the artifact rather than narration.
Claude-Session: https://claude.ai/code/session_01EyiWKr4SRzWJWo6n14HSUH