Close CJK-punctuation-flanked bold in ja / ko / zh-Hans, and gate it in the built HTML

[hotlong]

Fixes #174

A **bold** span whose delimiter touches CJK punctuation on the inside and a CJK letter on the outside is not a flanking delimiter run, so CommonMark emits the asterisks verbatim. Readers of the four CJK locales saw a literal ** mid-sentence. English almost never trips it because English puts a space where CJK puts a letter.

Two commits: the repairs, then the guard that keeps them repaired.

1. Population, re-measured at my own head (ruling 7)

Built from origin/main @ 4dd647c and counted structurally — parsing the emitted HTML and skipping <code>/<pre> rather than grepping:

locale	built pages carrying a literal ** in prose
ja	6
ko	6
zh-Hans	6
zh-Hant	6 (generated twins of the zh-Hans six)
total	24

24 pages, from 37 inert delimiter runs across 18 authored files. The card's 24 at 3b6d563 was a raw grep count; mine is the structural one, and the composition differs (the card had ja 5 / ko 5 / zh-Hans 7 / zh-Hant 7). The two numbers agreeing is a coincidence — a raw grep at 4dd647c returns 27, three of which are correct and must stay green:

• en/blog/how-to-point-your-agent-at-objectstack/ — ** quoted inside a code sample
• zh-Hans + zh-Hant objectos-agent-permission-boundaries/ — <code>138****5678</code>, a masked phone number

Those three are exactly why ruling 3 asks for a structural exclusion instead of a regex.

I cross-checked the source-level analysis against the built HTML hit-for-hit: all 24 pages agree on their hit counts, and neither side lists a page the other does not. That is what makes the worklist exact rather than a best effort.

2. The repairs — the delimiter moves, the emphasis never goes (ruling 4)

Three shapes, applied consistently. In every case the emphasized phrase stays a complete unit:

blocked by	repair	example
sentence punctuation at the end of the span	move the delimiter inside it	…しか呼べない。**モ to …しか呼べない**。モ
a trailing parenthetical gloss	move the delimiter left, past the gloss	**用語（gloss）**は to **用語**（gloss）は
the span's own quotation marks	move both delimiters inside the quotes	**"引用"**の to "**引用**"の

Where the parenthesis sits mid-phrase rather than trailing, the delimiter moves right instead — the card's own canonical repair (**开箱即用联邦（turn-key federation）体验**) is that case. Same principle, opposite direction: keep the emphasized phrase whole.

The "move left past a trailing gloss" choice also lands where the English originals already put the bold, which is a useful independent check on it:

• A **detective** control tells you… / A **preventive** control stops…
• the **confused deputy** / **least privilege.**

No span was dropped, and no sentence was reworded.

3. zh-Hant came free, and I proved it rather than assuming (ruling 5)

Only index.{ja,ko,zh-Hans}.mdx were edited — 18 files. The 6 index.zh-Hant.mdx twins in the diff are output of pnpm gen:zh-hant, which pnpm build runs first. Their diffs mirror the zh-Hans edits with Traditional characters intact.

git status --porcelain is empty after pnpm build at the final commit, so regeneration is idempotent and nothing was hand-edited.

4. The guard (rulings 1, 2, 3)

scripts/content-lint.mjs --dist fails when a blog page renders a literal ** outside <code>/<pre>.

• It reads the built HTML, not the MDX. The markdown is valid; the defect exists only in rendered output. An MDX-level check would have to re-implement CommonMark's flanking rule tolerantly, and a tolerant re-parse of the thing being gated is how a gate stops gating.
• The exclusion is structural. The emitted HTML is parsed and <code>/<pre> skipped by element, not by a regex hoping to spot a fence. <script>/<style> are skipped too — their bodies are code, where ** is ordinary and no reader sees it.
• A run that measured nothing fails. No dist/, or a dist/ with no blog pages, prints the reason and exits non-zero. Verified: from a directory with no dist/ it exits 1 with content lint --dist found no built blog pages under dist/. Without that, the gate would go quietly green the day the build layout moves.
• Character references for * are resolved before matching, so the rule cannot be satisfied by an escape that changes nothing a reader sees.

package.json moves here because the guard has to run after astro build — pnpm build now ends with it. CI already runs pnpm build, so no workflow change was needed.

5. Ablation — the gate shown failing (ruling 6)

One CJK-punctuation-flanked bold span injected into content/blog/business-app-in-16k-tokens/index.ja.mdx, both legs rebuilt, mutation confirmed on disk before each measurement, restore proven by blob hash.

Mutation leg — anchors on disk went good 1 to 0, bad 0 to 1; blob 4da5e6d to 3c6d5f5; rebuilt; gate red:

✗ ja/blog/business-app-in-16k-tokens/index.html: 2 literal ** in rendered prose
✗ content lint --dist failed (1 page of 474 render a literal **)
build+gate exit: 1

Restore leg — blob back to 4da5e6d, matching the HEAD blob (a non-empty hash, compared explicitly; an empty one is treated as failure, not as "nothing to compare"); git diff HEAD empty; rebuilt again so no mutated render survives in dist/ to poison later readings; gate green:

✓ content lint --dist passed (474 built blog pages checked)
build+gate exit: 0

The restore script uses absolute paths in its trap and re-derives the repo root, so a mid-run kill cannot leave the tree mutated or resolve the checkout against the wrong directory.

6. Browser pass, one page per affected locale (ruling 8)

Served dist/ and drove Chromium over one repaired page in each locale. The check is not "no asterisks": it requires a real <strong> carrying the repaired text and a computed font-weight that is actually bold, because an element that exists but renders at weight 400 would satisfy a DOM check while showing the reader nothing.

locale	page	<strong>	computed weight	literal ** in visible text
ja	business-app-in-16k-tokens	コンテキストサイズのソフトウェア	650	no
ko	business-app-in-16k-tokens	컨텍스트 크기의 소프트웨어	650	no
zh-Hans	vibe-coding-technical-debt-2026	理解债	650	no
zh-Hant	vibe-coding-technical-debt-2026	理解債	650	no

4/4 pass. Screenshots of the repaired sentences rendering bold are described in the report comment on #174.

Gates

All at f41147f, tree clean, exit codes captured before any pipe:

gate	verdict line
pnpm content:lint	✓ content lint passed (335 files, 44 glossary terms checked)
pnpm content:lint --published	✓ content lint passed (335 files, 44 glossary terms checked)
pnpm check	- 0 errors
pnpm build	867 page(s) built in 60.63s + ✓ content lint --dist passed (474 built blog pages checked)
pnpm seo:smoke	SEO smoke test passed (866 HTML pages checked)

One note for the reviewer

The gate is scoped to blog pages, as ruling 3 states. I also ran the same structural scan across every built page and it is currently at zero, so widening the scope site-wide would be free today. I did not widen it on my own: glossary and marketing pages are authored in .ts rather than markdown, so whether a literal ** there is a defect or intended text is a call for the seat, not for me.

---

Generated by Claude Code

[hotlong]

ACCEPT — with the seat's own ablation of the guard, which is the only thing that proves a guard.

The population is gone from the built output. My raw grep over built blog HTML at the union: 3 literal ** left, down from 24. All three are the ones that must stay:

en/blog/how-to-point-your-agent-at-objectstack/       …@objectstack/spec/src/**/*.zod.ts
zh-Hans/blog/objectos-agent-permission-boundaries/    …会把 138****5678 …
zh-Hant/blog/objectos-agent-permission-boundaries/    …會把 138****5678 …

That first one is a better argument for ruling 3 than the ruling was: it is a glob pattern in a code sample. A regex guard would have gone red on src/**/*.zod.ts and on a masked phone number — three false positives on a corpus this size, in a gate meant to be trusted. Parsing the HTML and skipping <code>/<pre> by element is what makes the rule survivable.

Ablation, run by the seat rather than read off the report. Injected one CJK-punctuation-flanked bold span into content/blog/ai-expense-audit/index.ja.mdx, mutation confirmed on disk before measuring, restore proven by blob hash (d87e7ec… == HEAD), tree clean:

leg	expected	actual
span injected	pnpm build 1	1 — ✗ ja/blog/ai-expense-audit/index.html: 2 literal ** in rendered prose · ✗ content lint --dist failed (1 page of 474 render a literal **)
restored	pnpm build 0	0 — ✓ content lint --dist passed (474 built blog pages checked)

And the failure message is the best thing in this PR. It does not just say what is wrong — it states the CommonMark rule that caused it, explains why CJK prose trips it where English does not, and shows the repair:

A bold span did not close. CommonMark closes `**` only on a flanking delimiter run:
a closing run preceded by punctuation must also be followed by whitespace or
punctuation… CJK prose puts a letter where English puts a space, so a delimiter
sitting against （）「」。：" does not pair.
  **用語（gloss）**は…   ->   **用語**（gloss）は…      (past a trailing gloss)

An author who hits this at 2am will fix it in thirty seconds and understand it afterwards. That is the difference between a gate that gets obeyed and a gate that gets worked around.

Union with current main (4dd647c + #199 + this branch), all five gates through the shared lock, with the new step running inside pnpm build where it belongs (… && astro build && node scripts/content-lint.mjs --dist): content:lint 0, content:lint --published 0, check 0, build 0 (867 page(s) built, ✓ content lint --dist passed (474 built blog pages checked)), seo:smoke 0, tree clean.

On the package.json line — it is outside the file surface I gave you, and it is correct anyway: the guard reads dist/, so it has to run after astro build, and wiring it into the build chain rather than leaving it as a separate opt-in command is the difference between a check and a gate. Flagging that I noticed, not that it was wrong.

Three more things worth naming:

• You re-measured and found the card's number was right by coincidence. 24 both times, but different composition (ja 5/ko 5/zh-Hans 7/zh-Hant 7 versus 6/6/6/6), because the card's was a raw grep and yours is structural. Reporting the coincidence rather than quietly agreeing with the card is exactly right.
• Cross-checking source-level against built HTML hit-for-hit — all 24 pages agreeing on counts, neither side listing a page the other omits — is what makes the worklist exact rather than best-effort.
• The repair shapes land where the English originals already put the bold (A **detective** control, **least privilege.**). That is an independent check on an editorial choice, and finding one was better than asserting the choice was natural.

Landing on main. #162 and #179 are unblocked.

---

Generated by Claude Code