Prose sweep on the in-flight UMS branch: 8 CI-blocking line breaks plus 23 style findings

[d-morrison]

A prose-standards sweep ran against an in-flight UMS branch (the session's
review learnings, touching shared/coding/least-flexible-tool.md,
shared/workflow/{adversarial-self-review,algorithmatize-checks,fixtures-are-not-evidence,self-review-fallback,claim-pr}.md
and self-review-fallback.cases.md). That branch has no PR open yet, so the
findings are recorded here rather than on a thread, and whoever picks the branch
up should apply them before opening it.

24 findings. The first is CI-blocking and was verified by fetching and
running the exact checker CI uses
(Morrison-Lab/gha/check-new-line-breaks@fdaeccc4, the SHA pinned at
.github/workflows/validate.yml:301) against git merge-base origin/main HEAD
--- exit 1 with exactly these eight errors, not predicted:

HIGH

1. Eight added lines fail new-line-breaks (fail: 'true', diff-scoped):
   least-flexible-tool.md:147,157; adversarial-self-review.md:93;
   algorithmatize-checks.md:750,755; fixtures-are-not-evidence.md:253;
   self-review-fallback.md:182,192. Each is >=80 stripped characters with a
   mid-line semicolon. least-flexible-tool.md:147 is 522 visible characters.
2. algorithmatize-checks.md:757 --- a Do/Don't block with only the Do.
   CLAUDE.md's "Record both the pattern and the anti-pattern" requires both,
   and requires deriving the missing half rather than omitting it. The retired
   behaviour is in the prose above and never labelled.
3. adversarial-self-review.md:77 --- in its words binds to the diff,
   five noun phrases from the intended referent (the reviewer). The wrong
   reading is fluent, which is ambiguous-reference.md's dangerous case.
4. least-flexible-tool.md:147 --- one 522-character sentence, three
   semicolon-joined independent clauses each with its own because. Past
   plain-prose.md's limit, and it is the same line failing CI in (1).

MEDIUM

5-6. self-review-fallback.cases.md:238/:242 state the same claim and the
same citation four lines apart; :213 restates :197-200 near-verbatim.
7. cases.md:216 --- "the primary" widens silently from claude-review to
"same-vendor readings", and the table at :222-225 contradicts the sentence.
8-9, 11, 13-14. Ambiguous pronouns: adversarial-self-review.md:83 (reading it), :79 (however hard it worked), fixtures-are-not-evidence.md:243
(what exposed it), least-flexible-tool.md:147 (two pronouns binding to
-c rather than to COMMENT_FLAG_RE), claim-pr.md:223 (the Don't above,
four candidates).
10. adversarial-self-review.md:90 --- a Do half too abstract to violate
("discount it by however much of its reasoning came from the artifact" names no
observable act), against CLAUDE.md's own requirement.
12. algorithmatize-checks.md:748 vs :750 --- the scope claim ("both under
outcomes above") contradicts the file's own taxonomy at :588-590.
15. 82 of 105 non-blank added lines exceed 80 characters, median 148, max 607,
against host files running 12/228 and 45/233 on main.

LOW

16-24: a forward reference (self-review-fallback.md:191), two more ambiguous
referents, a subject/verb mismatch with neither over eleven items
(:187), a clause-referring which with a noun in the way, one
"not merely X, they were Y" antithesis (ai-tells.md's named biggest tell) that
is also a comma splice, a mechanism restated within one file, an undated count,
and a quotation cited to the PR rather than to the comment permalink.

Checked and clean

All back-references in the added text resolve; only one directional-word hit,
and it is a back-reference; the adversarial-self-review.md and
self-review-fallback.md additions are genuine increments over the sections
they sit under rather than restatements.

Note

Filed rather than fixed because the session that produced the branch ran out of
quota. Finding 1 is the one that must be done before the PR opens; the rest are
ordinary prose-review work.

[d-morrison]

A separate verification pass on the same branch (at f083e125, after merging origin/main d735dcc4) found one defect beyond the prose findings above, and it is a falsified corpus-state claim rather than a style issue.

shared/coding/least-flexible-tool.md:148 cites COMMENT_FLAG_RE in hooks/require-agent-disclosure.py. No such constant exists in the repo. That prose line is the only occurrence of the string COMMENT_FLAG anywhere. The hook on main is PR #2131's state (c3da9e1b), whose constants are POST_RE, API_CMD_RE, API_BODY_FIELD_RE, API_COMMENT_TARGET_RE, REVIEW_ONLY_RE, ANY_BODY_FLAG_RE, MARKER_RE, HAS_INLINE_BODY_RE, BOT_COMMAND_RE, BOT_BODY_RE.

COMMENT_FLAG_RE and CLOSE_REOPEN_RE were introduced by 10909ff1 --- a #2185 branch commit, ancestor of neither this branch nor main. So the entry describes a constant that lives only in an unmerged PR.

Not merge-caused: git show f083e125^1:hooks/require-agent-disclosure.py | grep -c COMMENT_FLAG_RE returns 0, and so does the same against ^2. The claim was already false before the merge.

Worth naming because it is the class the entry itself is about: a claim about corpus state, written from what the author had in their working tree rather than from a query against the repo. Either wait for #2185 to land before citing its constants, or cite what exists on main today.

Everything else in that pass came back clean, and the clean results are worth recording since they bound the risk:

• Merge broke nothing. Main touched only 2 of the 7 files, on lines the branch never edited (the d-morrison -> the repository owner purge). The other 5 are byte-identical to the branch parent. Zero duplicated headings or bold-lead lines.
• All 8 relative link targets resolve, and every quoted section title still exists with that wording.
• The shlex count is exact, not merely safe: 9 hooks import it at HEAD and 9 at 10909ff1, the form the claim is anchored to.
• Occurrence numbering, claim-pr.md block ordering, and the AGENTS.md marker literal (byte-identical by repr comparison) all hold.
• check-links.py: exit 0, 2357 links across 555 files. check-context-closure.py: exit 0, both notes pre-existing.

Posted by Claude Code (AI agent) --- not written by a human.