Five more `s2twp` over-substitutions are wrong in any register — 介面卡 is a network interface card, not an adapter

[hotlong]

Filed by the PM seat from the corpus-wide survey in PR #192 (card #144), which the ruling on that card asked for and reserved the filing of. Every number below was re-verified by the seat against opencc-js and the corpus at 5d4b0b6, not copied from the survey.

Blocked on #144 / PR #192 landing — that PR builds the override mechanism this card would use, and the two would collide on the same table.

What

s2twp's TWPhrases group contains a handful of entries that are not contested vocabulary at all. They are mechanical over-substitutions: the preset appends or swaps a word fragment that is simply wrong in Taiwanese usage, in any register, for any audience. Same shape as the 許可權 defect #144 just fixed — only rarer, which is why nobody noticed.

zh-Hans	twp ships	ordinary Taiwanese	occurrences	what went wrong
全局	全域性	全域	4	a spurious 性 turns a noun into an adjective
扩展	擴充套件	擴充 / 擴充功能	4	a spurious 套件 ("package")
适配器	介面卡	配接器	3	介面卡 is a network interface card — different hardware entirely
控件	控制元件	控制項	2	the standard UI-control term is 控制項
镜像	映象	映像	1	wrong second character

Verified directly:

全局   twp-> 全域性     扩展   twp-> 擴充套件    适配器 twp-> 介面卡
控件   twp-> 控制元件    镜像   twp-> 映象

and every one is live in the built Traditional corpus at exactly those counts (全域性 4, 擴充套件 4, 介面卡 3, 控制元件 2, 映象 1).

Plus one the seat has already ruled on: 脚本

Seat ruling: normalize on 腳本, by shadowing the TWPhrases entry — the same identity-entry mechanism PR #192 built for 權限 and 實例. 指令碼 is the Microsoft-glossary rendering; 腳本 is the ordinary Taiwanese word, and in this corpus the referent is always shell/webhook glue in an automation argument, never a screenplay.

The more important half of this is a defect regardless of which term wins: the corpus already ships both.

指令碼  26 occurrences
腳本     5 occurrences

That split is not editorial. It is the same straddling-phrase accident PR #192 diagnosed for the locative 里: the two-character STPhrases entry 本里 matches across a word boundary in 脚本里, which blocks the phrase layer as well as the character layer, so five sites escape the 指令碼 substitution. Whichever term is chosen, the inconsistency has to close in the same pass.

Shape of the fix

PR #192 leaves behind exactly the mechanism this needs, and (per the patch round on it) a shared override module that both scripts/gen-zh-hant.mjs and src/lib/zhconvert.ts load. Each entry here is:

• one identity or replacement entry in that shared table, and
• one fixture case, so a preset upgrade cannot silently reinstate it.

Nothing else. No content file is hand-edited; the Traditional locale stays fully generated.

Acceptance

• All five artifact spellings go to 0 in the built output (dist, not just the MDX — gen-zh-hant: ship 權限/實例 and the straddled locative 裡 #192's patch round is the reminder that these two are different populations), and their intended forms appear at the same counts.
• 指令碼 → 0 and 腳本 → 31, matching 脚本 in the zh-Hans sources.
• Every changed byte explained by the ruled terms alone, checked mechanically the way gen-zh-hant: ship 權限/實例 and the straddled locative 裡 #192 checked its own: apply only the ruled substitutions to the main version of each file and require byte equality.
• git status --porcelain empty after pnpm build.
• Fixture ablation: remove one override, show it red.

Not in scope

The genuinely contested Microsoft renderings — 元数据 → 後設資料 (174 occurrences), 类型 → 型別 (67), 发布 → 釋出 (48), 审核 → 稽核 (40) and the rest — are a different question and are filed separately. This card is only the ones that are wrong on their face.

[hotlong]

dev claim — session session_01FeA1nwBz1ohH65dvffUGKr, branch claude/issue-194-twp-artifacts (pushed empty at ca3983b, worktree ../www.objectos.ai-issue-194).

Assignee left as-is (PM already claimed). Starting on the five mechanical TWPhrases over-substitutions plus the 脚本 → 腳本 normalization, per the rulings in the body. Will verify each term is actually emitted by TWPhrases (not STPhrases/STCharacters) before relying on the identity-entry shadow.

---

Generated by Claude Code

[hotlong]

os-dev-report

{
  "issue": 194,
  "status": "done",
  "branch": "claude/issue-194-twp-artifacts",
  "pr": "https://github.com/objectstack-ai/www.objectos.ai/pull/204",
  "premise_still_valid": true,
  "summary": "All six ruled terms corrected in TWP_PHRASE_OVERRIDES (src/lib/zhconvert.ts) with 18 new fixture cases in CONVERSION_CASES; content/ is regenerated output only (18 files), nothing hand-edited. Ruling 3 checked before the mechanism was used, not assumed: every one of the six wrong strings is emitted by TWPhrases alone, keyed on the post-s2t form, so the shadow reaches all of them. The one hit outside TWPhrases is STPhrases' IDENTITY entry 介面卡 -> 介面卡, which cannot introduce the string and is in fact why the repair belongs in the phrase layer rather than a post-pass — a post-pass could not tell a converted 适配器 from a source 介面卡. Four of the six are REPLACEMENT entries, not identity ones: an identity entry on 全局 would have disabled 全域性 but left the mainland spelling standing, which is its own defect; only 腳本 is an identity entry. Also found and deliberately left: TWPhrases carries the same 象/像 slip a second time in 顯像管 -> 映象管, inert here (映象管 is 0 in dist because no source writes 显像管), recorded in the code comment beside the 鏡像 override. main moved to 8a36292 (#202, a new content-lint rule) mid-round, so origin/main was merged in and the whole gate set re-run at f650569 rather than carried over.",
  "tests": "All five gates through bash /home/user/objectstack/scripts/pm/os-verify-lock.sh -c '<cmd>', exit captured before any pipe, re-run in full at f650569 (git rev-parse --short HEAD) after merging origin/main. Verdict lines as the lock printed them: gen:zh-hant 'VERDICT command-exit 0 · held the lock 2s'; check 'VERDICT command-exit 0 · held the lock 11s' with astro's own 'Result (135 files): 0 errors, 0 warnings, 0 hints'; content:lint --published 'VERDICT command-exit 0' with '✓ content lint passed (335 files, 44 glossary terms checked)'; build 'VERDICT command-exit 0 · held the lock 66s' ending '✓ content lint --dist passed (474 built blog pages checked)' — #201's --dist pass and #202's new rule both green; seo:smoke 'VERDICT command-exit 0' with 'SEO smoke test passed (866 HTML pages checked)'. git status --porcelain EMPTY after pnpm build. ACCEPTANCE COUNTS, measured on dist (not only MDX): 全域性 6->0, 擴充套件 5->0, 介面卡 18->0, 控制元件 3->0, 映象 6->0, 指令碼 35->0; intended forms 配接器 0->18, 控制項 2->5, 映像 0->6, 腳本 5->40, and 全域/擴充 0->6/0->5 standalone (a bare grep on main returns 6 and 5 there, but every one is a substring of the artifact itself). In the generated MDX corpus: 4/4/3/2/1/26 -> 0 and 腳本 5 -> 31, matching 脚本 in the zh-Hans sources. BYTE EQUALITY: each of the 18 changed files taken from origin/main, ONLY the six ruled substitutions applied, required byte-identical to the regenerated file — '18 changed content file(s), 40 ruled substitution(s), 0 unexplained', and 40 = 4+4+3+2+1+26. #187 guard: grep -rn '/zh-Hans/' content/blog --include=index.zh-Hant.mdx = 0. ABLATION: dropped ['適配器', '配接器']; mutation proven ON DISK before any reading — anchor line count 1 -> 0 and blob hash 767ba60 -> e15fe3a, both asserted, with trap '<restore>' EXIT INT TERM using an absolute path; generator then went RED with the exact expected diagnostic (in: '适配器' expected: '配接器' actual: '介面卡', gen exit 1); restore proven by blob hash back to 767ba60 = the HEAD blob AND empty git diff HEAD, then green again, tree clean. No rebuild leg was needed or skipped: scripts/gen-zh-hant.mjs imports src/lib/zhconvert.ts directly via Node type stripping, so no dist artifact sits between the edit and the measurement — verified by reading the import path rather than assumed. BROWSER: chromium at /opt/pw-browsers/chromium-1194 (the path in the dispatch, /opt/pw-browsers/chromium, has no chrome binary), dist served by python3 -m http.server 8194, PID killed after checking /proc/<pid>/cwd. Three pages x 1440x900 and 390x844, all six artifact spellings absent and the intended forms present in rendered innerText on every load: /zh-Hant/security/ (映像 derived by s2t from src/lib/security.ts — a UI string that exists in no MDX), /zh-Hant/blog/ (配接器 in the FDE card, from that post's description), /zh-Hant/blog/forward-deployed-engineer-tools/ (配接器 in the article body). Zero horizontal overflow, zero console errors, nav rendering 'ObjectOS 產品 開發者 資源 定價 繁體中文 開始使用'.",
  "open_questions": [
    {
      "question": "Non-blocking, implemented as ruled — but worth the seat knowing. 全局 -> 全域 is right on the mechanical reading (strip the spurious 性) and I applied it to all four blog occurrences plus the glossary and marketing-page ones. Two of the six are the technical sense where 全域 is plainly correct ('错误却是全局的', '一个全局定时服务'). The other four are the idiom for 'the whole picture' — '项目经理就能看清全局', '老板能看全局', '从片面变成了全局', 'Agent 能在一个上下文里掌握全局' — and there 全局 is itself an ordinary Traditional word (綜觀全局), so 全域 is understandable but reads as 'the whole domain'. The defect (全域性) is gone either way; the question is only which of the two correct forms those four take.",
      "options": [
        "A — leave as shipped: 全局 -> 全域 for all six, one rule, no sense-splitting. What is on the branch now.",
        "B — 全局 -> 全局 (identity) instead: fixes the 性 defect and keeps the idiom, but leaves 全域 unavailable for the two technical uses, where the mainland spelling would then stand.",
        "C — split by sense by hand in the four idiom sites, which the dictionary cannot do and which would break 'nothing in content/ is hand-edited'."
      ],
      "recommendation": "A, as ruled and as shipped. B trades a visible defect for a subtler one and C forfeits the generated-locale invariant, which is worth more than the nuance. Recording it only because #195 is the register for exactly this kind of judgement and this one was decided inside a card scoped to mechanical artifacts."
    }
  ],
  "out_of_scope_findings": [
    "filed as #203: 连接器 -> 聯結器 is a 16th contested term the #195 survey missed — 17 source occurrences (34 in dist, one of them a post title), which would rank fifth in #195's table, above ten of the rows already there. Contested, not mechanical, so not touched.",
    "not filed, already covered by #184: the zh-Hant FDE page embeds the Simplified-text cover SVG (/_astro/cover.*.svg, the same asset as zh-Hans, while /en/ gets its own cover-en) — 27 zh-Hant blog pages, 104 blog SVGs with 60 -en variants and 0 zh-Hant variants. Confirmed against the built output while doing the browser pass; searched before filing and #184 already holds it.",
    "not filed, recorded in the PR and in the code comment: TWPhrases entry 顯像管 -> 映象管 is a second instance of the exact 象/像 slip this card fixes for 鏡像, still live in the dictionary but inert in this corpus (映象管 = 0 in dist, no source writes 显像管). Left alone deliberately — a CRT is not this card's business — but it is the thing to reach for if 映象 ever reappears.",
    "not filed, marginal: 十进制 -> 十進位制 (1 occurrence). 進位制 is a legitimate Taiwanese term, so this fails the #194 'wrong in any register' test and is not worth a card; noted only so the next sweep does not spend time rediscovering it."
  ]
}

---

Generated by Claude Code