Bold that closes against CJK punctuation renders as literal `**` — 24 built pages across ja / ko / zh-Hans / zh-Hant

[hotlong]

Found while correcting extend-existing-systems-with-ai (#149). Filed unassigned. Not fixed under that card: its file surface is one post directory, and this is a site-wide class that wants one guard rather than 24 hand edits.

What readers see

A **bold** span whose closing delimiter sits directly after CJK punctuation does not close, and the asterisks are rendered to the reader as literal text. The construction that triggers it is idiomatic in the translated posts:

一套更丰富的**开箱即用联邦（turn-key federation）**体验
                                              ^^ never closes

CommonMark's flanking rule is the cause and it is not a bug in the renderer: a closing run preceded by punctuation (）, ”, 。) only closes when it is also followed by whitespace or punctuation. Followed by a CJK letter (体), it is not right-flanking, so the run is emitted verbatim. The English posts almost never hit it because English puts a space there.

Measured

Built from origin/main @ 3b6d563 (pnpm build), then counting literal ** in the emitted HTML of every blog page:

grep -rl '\*\*' dist/*/blog/*/index.html | wc -l    # 24

24 pages, entirely in the CJK locales — ja (5), ko (5), zh-Hans (7), zh-Hant (7). zh-Hant inherits each one from its zh-Hans source through gen:zh-hant, so the ~7 zh-Hans fixes carry their zh-Hant twins for free. Samples:

• ja/blog/business-app-in-16k-tokens/ — これが**コンテキストサイズのソフトウェア（context-s…
• ko/blog/ai-wrote-your-app-dare-to-merge/ — 아니라 **“이 권한이 맞는가…
• zh-Hans/blog/vibe-coding-technical-debt-2026/

(One of the 24, zh-Hans+zh-Hant extend-existing-systems-with-ai, is fixed in the PR for #149 because the correction rewrote that exact sentence; the other 22 are untouched.)

Why a guard, not just 22 edits

Nothing catches this today: the markdown is valid, content:lint passes, astro check passes, seo:smoke passes, and the page builds — the defect exists only in the rendered output. Every future translated post can reintroduce it, and it is invisible to anyone who does not read the CJK locales.

The cheap guard is a content-lint rule over the built (or rendered) HTML: no blog page may contain a literal ** outside <code>/<pre>. That is one grep, it closes the whole class, and it fails loudly on the next occurrence. The fix per site is to move the delimiter off the punctuation — **开箱即用联邦（turn-key federation）体验** renders correctly.

Re-check

pnpm build
grep -rl '\*\*' dist/*/blog/*/index.html
grep -o '.\{60\}\*\*.\{60\}' dist/ja/blog/business-app-in-16k-tokens/index.html

---

Generated by Claude Code

[hotlong]

Queued as the next card out — held, not deferred. Rulings recorded now so it can go the moment its lanes clear.

The seat re-read this and it is the highest-priority open item in the queue: 22 pages are showing readers literal ** right now, in the four locales least likely to be read by anyone who could report it. That is a visible-defect card, which under the maintainer's standing instruction is work, not a note.

Why it is not dispatched this minute. It collides with three of the four cards in flight, on file surfaces rather than on ideas:

In flight	Owns	Overlap with this card
#144	scripts/gen-zh-hant.mjs + all 46 index.zh-Hant.mdx as regenerated output	your ~7 zh-Hans fixes each regenerate a zh-Hant twin
#159	content/blog/**/index.ja.mdx	5 of the 24 pages are ja
#185	scripts/content-lint.mjs	the guard belongs in exactly that file

Fixing this on top of any of them would produce merge conflicts in generated files, which is the one place a conflict is genuinely dangerous — a bad resolution silently ships hand-edited Traditional content, which AGENTS.md forbids outright. So it waits for the lanes, not for a decision.

Seat rulings, so the dispatch is ready:

1. The guard ships with the fix, in the same PR. The card is right that 22 hand edits without a guard is a fix that decays: nothing in content:lint, astro check, astro build or seo:smoke sees this, so the next translated post reintroduces it and nobody notices. A card that closes a class and leaves the class reopenable has done half the work.
2. The guard reads the built HTML, not the MDX. The defect exists only in rendered output — the markdown is valid, and any MDX-level heuristic would be re-implementing CommonMark's flanking rule tolerantly, which is the shape this repo has already ruled against in gates. Assert over dist/, after pnpm build.
3. The rule is: no blog page may contain a literal ** outside <code>/<pre>. Code samples legitimately contain asterisks; prose never should. Make the exclusion structural (parse the emitted HTML and skip those elements), not a regex that hopes.
4. Fix by moving the delimiter, never by deleting the emphasis. **开箱即用联邦（turn-key federation）体验** is the repair; dropping the bold because it is awkward to close would be a silent editorial change to 22 translated pages.
5. Never hand-edit an index.zh-Hant.mdx. Fix the zh-Hans source and regenerate. The card already notes the ~7 twins come free; prove it rather than assume it, with git status --porcelain empty after pnpm build.
6. Ablate the guard. A green tree proves nothing here either — the guard has to be shown failing. Inject one CJK-punctuation-flanked bold span into a post, build, and show the gate red naming the page; remove it and show it green. Mutation confirmed on disk before each measurement, restore proven by blob hash, trap on absolute paths.
7. Re-measure the population before fixing. The card counted 24 at 3b6d563; main is now at 435ff61 and one of the 24 was already fixed under extend-existing-systems-with-ai tells readers ADR-0015 federation is *Proposed* and unshipped — it is Accepted and shipped #149. Report the count you actually find, per locale.
8. Browser pass on at least one page per affected locale — this is a card about what a reader sees, so the acceptance evidence is a screenshot of the repaired sentence rendering as bold, not a grep returning zero.

Labelled pm:queue, still unassigned. It goes out as soon as #144, #159 and #185 have landed.

---

Generated by Claude Code

[hotlong]

All three lanes cleared — dispatching now. #144 landed as 9522e69, #159 as 4dd647c, #185 as 5d4b0b6. Nothing in flight touches scripts/content-lint.mjs, the Traditional corpus or the Japanese posts any more; #193 owns SVGs only.

The eight rulings in the seat comment above stand unchanged. Two things have moved under them since they were written, and both help:

• content:lint now derives its term sets from the taxonomy (content-lint: derive all four frontmatter term sets from the taxonomy #190), and the Traditional locale is generated through one shared converter (gen-zh-hant: ship 權限/實例 and the straddled locative 裡 #192, src/lib/zhconvert.ts, loaded by both Vite and the plain-Node generator). Your guard goes into a content-lint.mjs that is in better shape than when this card was filed.
• Re-measure the population before fixing. The card counted 24 at 3b6d563; main is now 4dd647c, one of the 24 was already fixed under extend-existing-systems-with-ai tells readers ADR-0015 federation is *Proposed* and unshipped — it is Accepted and shipped #149, and gen-zh-hant: ship 權限/實例 and the straddled locative 裡 #192 rewrote every Traditional file. Report the count you actually find, per locale, at your own head.

⚠️ Operational note that matters more than usual. This container has restarted twice in the last twenty minutes and killed three agents, two of them with no work pushed. Work in pushable increments and git push as you go — the population fix first (it stands alone), then the guard, then the ablation. A restart must cost you one step, not the card. If you sense instability, push what you have and post a partial report naming exactly what is done.

Claim: PM loop round 4 (wave 7)
Session: session_01FeA1nwBz1ohH65dvffUGKr
Branch: claude/issue-174-cjk-bold-delimiter
Worktree: www.objectos.ai-issue-174
Domain: repo:www.objectos.ai (single-lane execution seat)
File surface: scripts/content-lint.mjs for the guard; content/blog/**/index.{ja,ko,zh-Hans}.mdx for the repairs; content/blog/**/index.zh-Hant.mdx as regenerated output only — never hand-edited.
Container & model: M, mode:subagent, model: opus
Clause-②: no
Serial constraints cleared: #144 (9522e69), #159 (4dd647c), #185 (5d4b0b6) all landed. #162 and #179 are held behind this — both want scripts/content-lint.mjs. #193 owns the SVG lane and does not collide.

Environment ruling carried forward: REST and gh are unreachable from this container (403). GitHub MCP is the sanctioned channel here, overriding the REST-only instruction in .claude/agents/os-dev.md for this repo.

---

Generated by Claude Code

[hotlong]

dev claim — starting work on this card now.

Session: session_01FeA1nwBz1ohH65dvffUGKr
Branch: claude/issue-174-cjk-bold-delimiter (pushed empty at 4dd647c, write route verified)
Worktree: /home/user/www.objectos.ai-issue-174, branched from origin/main @ 4dd647c
Assignee untouched.

Working the three rulings-ordered increments, pushing after each so a container restart costs one step, not the card:

1. re-measure the population at my own head, then repair ja / ko / zh-Hans by moving the delimiter (ruling 4), regenerating zh-Hant (ruling 5)
2. the guard in scripts/content-lint.mjs, reading built HTML with structural <code>/<pre> exclusion (rulings 1–3)
3. the ablation (ruling 6) and the browser pass, one page per affected locale (ruling 8)

---

Generated by Claude Code

---

Generated by Claude Code

[hotlong]

os-dev-report

{
  "issue": 174,
  "status": "done",
  "branch": "claude/issue-174-cjk-bold-delimiter",
  "pr": "https://github.com/objectstack-ai/www.objectos.ai/pull/201",
  "premise_still_valid": true,
  "summary": "Repaired every CJK-punctuation-flanked bold span in the authored locales and shipped the guard that keeps them repaired, in one PR. Re-measured at my own head 4dd647c (ruling 7): 24 built pages carry a literal ** in prose, from 37 inert delimiter runs across 18 authored files, split ja 6 / ko 6 / zh-Hans 6 / zh-Hant 6 — same total as the card but a different composition, and a raw grep returns 27 because three hits are legitimate in-code asterisks (an English code sample, and <code>138****5678</code> in zh-Hans + zh-Hant objectos-agent-permission-boundaries). The defect has a second shape the card did not name: the OPENING delimiter can also fail, when it sits between a CJK letter and an opening quote — 3 of the 37 runs are that mirror case, and they need the delimiter moved inside the quotes rather than past a gloss. Only index.{ja,ko,zh-Hans}.mdx were edited; the 6 zh-Hant twins in the diff are gen:zh-hant output. Guard is content-lint.mjs --dist, reading built HTML with a structural <code>/<pre> exclusion, wired into pnpm build after astro build (which is why package.json moves too — without that one-line wiring the guard never runs; CI already runs pnpm build, so no workflow change). Non-blocking observation, no issue filed: enterprise-ontology-race-open-vs-closed zh-Hans:143 emphasises with bold where the English original uses italics — a pre-existing translation choice, not a defect, and I left it alone beyond making it render.",
  "tests": "All gates at f41147f, tree clean, exit captured before any pipe, every verdict quoted from the gate's own output. pnpm content:lint -> '✓ content lint passed (335 files, 44 glossary terms checked)' exit=0. pnpm content:lint --published -> same line, exit=0. pnpm check -> '- 0 errors' exit=0. pnpm build -> '867 page(s) built in 60.63s' + '✓ content lint --dist passed (474 built blog pages checked)' exit=0. pnpm seo:smoke -> 'SEO smoke test passed (866 HTML pages checked)' exit=0. All heavy commands ran through os-verify-lock.sh (VERDICT command-exit 0). Population verified two independent ways that agree hit-for-hit on all 24 pages: a structural scan of dist/ and a CommonMark flanking-rule analysis of the sources; after the fix the structural scan reports 0 prose ** across blog pages and 0 site-wide. ABLATION (ruling 6): injected one CJK-punctuation-flanked span into content/blog/business-app-in-16k-tokens/index.ja.mdx. Mutation confirmed ON DISK by anchored grep before measuring — good anchor 1->0, bad anchor 0->1, blob 4da5e6d->3c6d5f5 — never by the editor's exit code; then REBUILT (pnpm build, which regenerates dist/) and the gate went red naming the page: '✗ 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 got the same treatment rather than being assumed: git checkout HEAD -- <abs path>, blob back to 4da5e6d compared against the non-empty HEAD blob (empty would be failure, not 'nothing to compare'), git diff HEAD empty, anchors back to 1/0, then REBUILT AGAIN so no mutated render survived in dist/ to poison later readings — gate green, '✓ content lint --dist passed (474 built blog pages checked)', exit 0. trap used absolute paths off a re-derived repo root. Guard's not-measured path verified too: run where no dist/ exists it exits 1 with 'content lint --dist found no built blog pages under dist/' rather than passing. RULING 5: git status --porcelain empty after pnpm build at the final commit, so the zh-Hant twins regenerate idempotently and none was hand-edited. RULING 8 (browser): served dist/ with python3 -m http.server, drove Chromium over one repaired page per affected locale; asserted a real <strong> carrying the repaired text AND computed font-weight >= 600, because an element rendering at weight 400 would pass a DOM check while showing the reader nothing — ja コンテキストサイズのソフトウェア 650, ko 컨텍스트 크기의 소프트웨어 650, zh-Hans 理解债 650, zh-Hant 理解債 650, no literal ** in visible text on any of them, 4/4 PASS, screenshots confirm the sentences render bold. Server stopped by PID after verifying /proc/<pid>/cwd; no kill by process name.",
  "open_questions": [
    {
      "question": "Should the --dist guard stay scoped to blog pages, or cover every built page? Non-blocking — shipped as ruling 3 specifies (blog only).",
      "options": [
        "A. Keep blog-only, exactly as ruled.",
        "B. Widen to all built pages. The same structural scan is already at zero site-wide, so it is free today and would close the class on glossary and marketing pages too."
      ],
      "recommendation": "A for this PR, since ruling 3 is explicit and I will not widen a gate's blast radius on my own. B is worth a follow-up card: the argument for it is ruling 1's own ('a card that closes a class and leaves the class reopenable has done half the work'). The reason it is not a pure freebie is that glossary and marketing pages are authored in .ts rather than markdown, so whether a literal ** there is a defect or intended text is a seat call, not mine."
    }
  ],
  "out_of_scope_findings": []
}

---

Generated by Claude Code

---

Generated by Claude Code