Repo-wide em dash (U+2014) violates CLAUDE.md's house style outside @ref glosses

[philcunliffe]

What

CLAUDE.md bans the em dash (U+2014) repo-wide: "No em dashes ... anywhere: code,
comments, JSDoc, strings, or docs." That is a standing, repo-wide violation on
master today, independent of any open PR.

Verified counts (measured directly against master at this triage)

• 2459 lines across 193 tracked files contain at least one U+2014.
• Confirms the specific examples raised during triage of PR LLP corpus hygiene: illustrative @refs are marked and no gloss uses an em dash (#463 items 2 and 3) #468:
  • llp/0045-client-attach.design.md: 66 lines
  • llp/0041-central-config-client-actions.design.md: 54 lines
  • hypaware-plugin-kernel-types.d.ts: 36 lines
• Heaviest offenders overall: hypaware-core/plugins-workspace/{claude,codex}/skills/hypaware-report-to-html/SKILL.md (72 lines each), llp/0063-login-auto-provision-forward-sink.decision.md (71), llp/0045-client-attach.design.md (66), llp/tombstones/0020-hypaware-design.explainer.md (66), llp/0041-central-config-client-actions.design.md (54), llp/0022-iceberg-export-partitioning.spec.md (51), llp/tombstones/0018-implementation-plan.plan.md (51).

Scope note

Issue #463 item 3 (fixed by PR #468) covers only em dashes used as the
separator inside @ref gloss lines, and PR #468 lands that fix. This issue is
the broader, repo-wide CLAUDE.md violation: em dashes in ordinary prose,
comments, and docs outside @ref annotations, which #463 never claimed to
cover and which no other open or closed issue tracks (search performed at
filing time: em dash, em-dash, U+2014, 463 all checked).

A related gap worth folding into the fix

While triaging #468 I confirmed the new test/core/llp-ref-hygiene.test.js
em-dash check only inspects the single line containing the @ref token, so
an em dash in a continuation line of a multi-line gloss is invisible to it.
Three live examples exist today, undetected by the new test:

• hypaware-core/plugins-workspace/format-iceberg/src/table-format.js:456
• llp/0067-session-opt-out.design.md:114
• src/core/runtime/activation.js:64

Worth widening that test's scan to the full multi-line gloss (or just running
it as part of whatever mechanism resolves this issue), so the corpus-hygiene
invariant PR #468 establishes doesn't have a standing blind spot.

Backlink

Found and filed during neutral triage of #468 (branch fix/issue-463, head
ef04fcd709c145866bcf7d6e439d4ea12b056bb9). Not part of that PR's declared
scope, so filed separately rather than blocking it.

[philcunliffe]

🤖 neutral: stuck on sequencing, not on the finding (the sweep must wait for the held PRs to merge)

The finding is real and correctly filed. Neutral is declining to attempt it right now, and recording why, rather than opening a PR that would damage work already waiting on you.

The problem with sweeping now

The violation spans 2628 lines across 198 files on master (verified independently: git grep -c $'—' origin/master). A mechanical sweep is therefore a ~198-file PR. There are currently seven open PRs, and five of them are green, reviewed, and carrying neutral:approved awaiting your merge:

PR	Files changed	State
#468	64	held, neutral:approved
#466	10	held, neutral:approved
#467	7	held, neutral:approved
#458	10	held, neutral:approved
#461	6	held, neutral:approved
#456	10	held, neutral:stuck
#462	6	held, neutral:stuck

A 198-file sweep touching prose and comments across the tree would conflict with essentially all of them. Every conflict would move those PRs' head SHAs, which strips neutral:approved (the label tracks the current reviewed-clean head by design, LLP 0030) and re-opens the whole rung ladder: merge-base, CI, and a fresh review round each. Five PRs that are ready for you to merge today would go back to being unreviewed work in progress, and the sweep would have bought nothing but style compliance.

That trade is clearly wrong, so neutral is not making it unilaterally.

How this unsticks

Merge (or close) the held PRs, then drop the neutral:stuck label or reply here. With the queue drained, the sweep is a single mechanical PR against a quiet tree with almost no conflict surface, which is the cheapest moment it will ever have.

Note on scope, for whoever runs it

• Confirmed no prior issue tracked this, which is why PR LLP corpus hygiene: illustrative @refs are marked and no gloss uses an em dash (#463 items 2 and 3) #468's triage filed it.
• PR LLP corpus hygiene: illustrative @refs are marked and no gloss uses an em dash (#463 items 2 and 3) #468 already fixed the 164 em dashes inside @ref glosses, so measure again after it merges; the residual on a post-LLP corpus hygiene: illustrative @refs are marked and no gloss uses an em dash (#463 items 2 and 3) #468 tree is roughly 2464 lines.
• CLAUDE.md's rule is repo-wide and not limited to @ref glosses: code, comments, JSDoc, strings, and docs. In prose use the punctuation the sentence wants (a comma, colon, parentheses, or a sentence split); in runtime strings prefer -.
• PR LLP corpus hygiene: illustrative @refs are marked and no gloss uses an em dash (#463 items 2 and 3) #468's triage also found the em-dash gate only scans the line containing @ref, so multi-line gloss continuations are undetected. Three live instances: hypaware-core/plugins-workspace/format-iceberg/src/table-format.js:456, llp/0067-session-opt-out.design.md:114, src/core/runtime/activation.js:64. Worth folding into the same sweep, and worth extending the gate so it cannot recur.
• This is style compliance with no runtime behaviour change, so it should be a low-risk, high-file-count, mechanical PR. Do not let it reword any rationale.

[philcunliffe]

The five held PRs all merged today, so the queue concern is cleared. Run the sweep now: one mechanical PR, full CLAUDE.md scope (code, comments, JSDoc, strings, docs), no rewording of rationale. I accept the small rebase risk on the remaining open PRs (#456/#462/#491 etc); do not wait for them.

Fold in the gate widening too: extend the em-dash check in test/core/llp-ref-hygiene.test.js to scan full multi-line glosses, and fix the three known continuation-line instances.