check-non-standard-chars skips .md files (and U+00D7), so gha's own README has 39 unflagged em-dashes

[d-morrison]

Noticed while working #321. Not blocking that PR (whose own added lines are clean); filing so it isn't lost.

What's wrong

check-non-standard-chars/check-non-standard-chars.py scans only two extensions:

extensions = ['.qmd', '.R']   # line 93

so Markdown is never scanned. The consequence is visible in this repo:

$ grep -c $'—' README.md
39

39 em-dashes in README.md, and the chars / check-chars selftest job reports
success on every run. CHANGELOG.md, REVDEPS.md, changelog.d/*.md, and the
examples/ stubs' comments are equally unscanned.

Markdown files are source too: version-controlled, diffed, and rendered, so they
carry the same smart-quote / copy-paste corruption and whole-paragraph-reflow
diff problems .qmd and .R do. The ai-config corpus's
shared/coding/ascii-punctuation-in-source.md states the rule as covering every
tracked source file, .md explicitly included, and names extending this very
check as the enforcement follow-up.

Second, smaller gap: the glyph set

NON_STANDARD_CHARS (line 25) covers six code points: curly quotes (U+2018,
U+2019, U+201C, U+201D) and en/em dashes (U+2013, U+2014). It has no entry for
the multiplication sign (U+00D7), which the same shared fragment names
alongside the others as a glyph that slips in from rendered-text copy-paste and
should be written as $\times$ / × in prose or × in a string
literal.

Proposed scope

1. Add .md to the scanned extensions.
2. Add U+00D7 to NON_STANDARD_CHARS.
3. Make the extension list an input rather than a hard-coded literal, with
   the current set as the default, so a consumer can narrow or widen it without
   forking the script (per shared/coding/configurable-parameters.md).
4. Clean up this repo's own existing violations in the same PR, or the newly
   widened check turns chars / check-chars red on main immediately. This is
   the reason to treat (4) as part of the change rather than a follow-up: 39
   hits in README.md alone, plus whatever the other .md files hold.
5. Update the action's own description: (it says ".qmd and .R files"), the
   README/website tables, and website/reference/check-non-standard-chars.qmd,
   per the doc-sync checklist in CLAUDE.md.

Consumer impact

Widening the default scope will flag pre-existing prose in every consumer repo
that calls check-non-standard-chars.yml, not just this one. Worth deciding
whether the new extension list ships as the default (with a heads-up in the
changelog under ### Changed) or as an opt-in input that defaults to today's
.qmd/.R behavior. My inclination is the former, since the point of the rule
is that Markdown was never meant to be exempt, but that's a maintainer call.

[dem-extra1]

Extra evidence for the scope question here, found while working #342: the gap is not only .md. .github/workflows/*.yml is unscanned too, and it has drifted further than the README has.

.github/workflows/claude.yml alone carries 40 em-dashes in its comment prose (measured on main, 2026-07-29). I introduced two more without noticing while adding comments to that file on #343 — copying the surrounding style, which is exactly the failure mode the ASCII-punctuation rule describes — and caught them only with an ad-hoc scan of my own diff, not with any check in this repo.

So the fix here probably wants three things rather than two:

1. scan .md (this issue's original point),
2. add U+00D7 to NON_STANDARD_CHARS (also already noted here),
3. scan .yml/.yaml, where the workflow comment prose lives.

Worth pairing with diff-scoping when it lands, the way check-new-line-breaks does: a whole-tree scan would immediately flag those 40 pre-existing hits and nothing would merge until someone did a repo-wide sweep. Diff-scoping makes it enforce the rule on new prose while leaving the grandfathered lines for a deliberate cleanup pass.

---

Generated by Claude Code