Close the remaining findings from the full-catalog read

[cbrock84]

Follow-up to the two mechanical fixes already merged. Each item was verified against the tree before being acted on; two of the reported findings turned out to be wider than reported and one turned out to be a misreading.

A fifth skill had the same defect

Four skills had a second list of rules spliced into an existing ## Never block, landing inside a wrapped bullet and stranding its continuation on an unrelated rule. Those are fixed. A fifth carried the same splice in the same foreign voice — nothing was orphaned there, so a lint keyed on the split could not see it.

All eleven spliced bullets are now in the house voice, and scripts/check-never-blocks.py holds it.

Why this check and not the obvious one. Two bullet styles are in use across the catalog and both are deliberate: most blocks are bare imperatives ending in a period, while the chief-level blocks are Never … / Do not … lines without terminal punctuation. Either is fine. What is never fine is both inside one block, because that is exactly what a list spliced in from elsewhere looks like.

lint	hits on the tree before this change
bullet lacks terminal punctuation	59 — mostly the deliberate style
block mixes the two styles	5 — every affected file, nothing else

Red on the pre-fix tree, green after, and red on the tree as it stood before the earlier fix too.

The surface guard missed real violations

git status --porcelain wraps any path containing a space in literal quotes and octal-escapes non-ASCII bytes, so a changed file at skills/a b.md arrived as "skills/a b.md" and failed the startsWith() prefix test. The guard then reported no violation for a file squarely inside another agent's surface — the one thing it exists to catch.

Reproduced in a throwaway repository with three such paths, using finance as the surface under test:

 M "plugins/finance/skills/a b.md"
RM plugins/finance/skills/plain.md -> "plugins/finance/skills/re named.md"
 M "plugins/finance/skills/\303\274n\303\257/c.md"

before:  inside plugins/finance/ : 0 of 3
after:   inside plugins/finance/ : 3 of 3

Both commands now read NUL-delimited output, which also removes the need to guess whether a -> is a rename arrow or part of a filename. Latent in this tree — no such paths exist here — but live for anyone who installs the hierarchy skill and runs the guard against their own repository, which is the point of shipping it.

A skill referenced a file it does not ship

A skill pointed at a repository document that ships in no plugin. It resolves for a reader of this repository and dangles for everyone who installed the department. The point is now stated inline.

A blanket dead-path check would false-positive on the hierarchy skill, whose docs/ paths are instructions to create those files in the reader's own repository rather than references to files here. The rule is written into CONTRIBUTING with that exception named, rather than enforced by a checker that would cry wolf.

Two skills in one department gave different thresholds

One said queue utilization above roughly 80% is where wait times rise sharply; the other headed a section with ninety percent. Both statements are individually true, but a reader asking one department gets two numbers.

Aligned on the lower one — that is where the non-linearity actually begins, and it is the specialist skill that owns the topic. The chief-level skill now points at the specialist for which threshold applies to a given queue.

Two overlapping claims, resolved in opposite directions

One description promised affiliate coverage. The body never had it, and another department covers affiliates properly — mechanics, commercial terms, fraud control. The claim is dropped and the skill hands affiliates over explicitly, which also fixes the routing: a question about affiliates should not load the skill that says least about them.

A second pair already had a boundary note, but only on one side. The reciprocal is now stated in the other skill's description and body, so the split is visible whichever one a reader reaches first. Both survive — they are genuinely different jobs, one about a format across platforms and one about a single platform in depth.

Recorded rather than changed

docs/DECISION-LOG.md gains D35 for the one open convention question — whether description shape should be constrained beyond the length and promise-versus-coverage checks that already exist. Recommendation is to leave it unconstrained; the two existing checks catch the failures that cost something, and shape is not one of them.

The same entry records the one reported inconsistency that is a misreading, so the next reader does not raise it again: the agent count in the hierarchy skill describes the implementation the method was extracted from, not this roster. Both instances say "extracted from". Changing the number would make it wrong.

Two indexes were stale

CONTRIBUTING's check table listed four of the ten checks CI runs. The decision log's status table stopped at D16 while eighteen later entries existed — an index that omits over half its entries is how an open decision goes unnoticed. Both are current.

Verification

./scripts/check-all.sh — all ten green: 172 skills valid, provenance clean, README / social card / org chart current, 277 skill references resolve, 0 British spellings, 172 ## Never blocks internally consistent, manifests parse.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FaQP2WCqeFYH7s9pi1CJ5u