Language-agnostic games — the plan and the pitfalls

The dice specimen plays in 简体中文. The next step is a table that plays in ANY room language — including the mixed table（a Chinese speaker and an English speaker in one game, which the room's two-layer language system already permits）. This page is the PLAN, written before the build, because the owner's warning is exactly right: some of the game's words are load-bearing in code, and renaming a load-bearing word is a migration, not a translation. Status: PLAN — no code yet.

1 · The inventory — where the words actually live

The audit that must precede any edit. logic = the machine compares this string; renaming it silently breaks rules or state. display = a human reads it; safe to localize.

2 · The design — one new block, two refits

• The LEXICON block（new, in the kit）. A per-language term table beside the flow: for each language the kit supports — zh first, en as the proof — the display for every canonical token: the modes（斋→"pure" · 飞→"wild" · 普通→""）, the bid shape（「{n}个{x}{mode}」→ "{n} {x}s {mode}"）, the button label, the board words, the note templates（the circle, the void, the verdict, the 空开, the nudges, the round-open）, the reason templates, and the dial LABELS. The schema lints it: every canonical token covered, every language complete — a missing term fails the shelf lint, not a live table. Rooms pick the slice by their language; an uncovered language falls back to zh（correct, just foreign）.
• Refit one — dial values become ids（the auto_when precedent, finished）. ones_wild stops matching「1就是1」; the dial declares meaning per option（e.g. {"id":"plain","label":{"zh":"1就是1","en":"ones are ones"}}）and the runner reads the id. Same for any future dial the world must understand.
• Refit two — reasons become codes. The validator returns (code, params)（order·climb·zhai·double·locked·ceiling·convert·input）; the lexicon holds each code's template per language. The replica in the audit driver already proved the codes cover every branch.

What deliberately does NOT change: the canonical enum values（斋/飞/普通 stay the ids in code, state, the envelope keys, and the interpreter's JSON — frozen, like slugs）; the append-only history; the bench's Chinese-room scenarios（they become the zh half of a two-language matrix）.

3 · The build order（each phase shippable alone）

STATUS — BUILT, all seven phases, 2026-08-03（the owner's go）. The ids went ENGLISH（plain/zhai/fei, dial ids, reason codes — §5's ruling）with a read-time normalizer for the legacy zh tokens and the en synonym families（no wilds · wild off · spot only · pure → zhai; with wilds · wilds · wild on → fei）. The kit carries the full zh+en lexicon（key-parity linted）; the runner renders every machine surface through it; the strip reads a per-room terms slice off the envelope（the client owns no words）; a room mounted before its kit had a lexicon falls back to the shelf's current block（caught live: an empty-string board cell）. Gates: 290 bench checks green incl. the new en-room scenario + the English-family validator rows; smoketest PASS; one full LIVE en game confirmed — board, voids, verdict, circle, round-open all English, the arithmetic intact. The code rides the uncommitted dice arc（one commit, batched with the rework, after the owner's sign-off）; this page records the state of the working copy.

1. P1 — the dial-id refit. Smallest, closes the hidden logic hole. Bench: an en-labelled dial flips nothing.
2. P2 — the lexicon block + shelf lint, with zh filled for the dice kit（pure relocation of today's strings — zero behavior change, the diff proves it）.
3. P3 — the runner renders through the lexicon: notes, board cells, reasons （as codes）, the tally fact. zh output byte-comparable to today's.
4. P4 — the seam prompts go multilingual: the interpreter's sysp gains the lexicon's synonyms + non-CJK examples; the voice writes in the room language with the lexicon's terms IN the note（no ad-hoc translation left to do）; the number-fidelity accept-set keyed by language.
5. P5 — the client strip reads the lexicon slice off the game state（pill faces, unit, composed bid text）.
6. P6 — the en kit half: fill the dice lexicon's en column; translate the kit's dial labels; leave the brief bilingual（the persona adapts）.
7. P7 — prove it: the bench matrix（every dice scenario × zh and en）+ the audit driver in an en room AND a MIXED table（one bidder typing English, one Chinese, one room language）— the mixed table is the real acceptance test.

4 · The pitfalls, named

1. The enum is not a word. 斋/飞/普通 look like display but are compared in five layers and persisted in every played room. Renaming them corrupts old ladders on resume and breaks the client contract. They stay; only their RENDERING localizes.
2. The hidden dial match. startswith("1就是1") is the one place the world already learned a kit's vocabulary — an en dial value would silently turn wilds on for every table. Fixed first for a reason.
3. Voice-time translation drifts. Letting the flash voice translate a zh note per line renders 斋 three different ways in one game. Terms must be translated ONCE, in the note builder, from the lexicon.
4. The mixed table is the real case. Layer A already lets players type any language. The interpreter must read all of them into the canonical enum; the machine's OUTPUT language follows the ROOM, not the bidder. Test explicitly.
5. Number-word fidelity. The guard accepts digits and 汉字 today; every added language either extends the accept-set or instructs digits-only output. Silent rejections would resurrect the fallback churn the specimen just cured.
6. The bench is full of literals. Every zh assertion（开盅 · VOID reasons · board cells）must assert through the lexicon or run in the zh half of the matrix — otherwise the en build breaks 278 green checks and the temptation is to weaken them.
7. Two i18n systems, one boundary. lib/i18n.js owns app chrome; the lexicon owns game terms. Putting a game term in the UI dict（or vice versa）creates the two-sources drift the single-source rule exists to prevent.
8. The kit's dual-mode prose. The classic path（no flow block）plays from the brief alone — a translated table needs the BRIEF's rules readable too. The persona translates prose fine; the DIALS it declares out loud must use the lexicon's labels or the spoken table and the machine disagree.
9. Fallback asymmetry. A partially-filled lexicon must fall back to zh WHOLE per surface（never word-by-word）— a half-en half-zh verdict line is worse than a Chinese one.
10. The specimen's laws still bind. Notes stay facts（no hardcoded persona speech — templates carry FACTS in the room's words, the voice still performs them）; one number in view; markers minted by code; persist after the advance. Localization touches none of these.

5 · The dice lexicon — the translations, for the owner's review

Researched, not guessed: English-language sources for the Chinese variant keep zhai and fei romanized（the community's own usage）;「pure」appears only in「pure bao zi」. 加点数 renders as raise the face（the standard rule phrasing: increase the quantity, or bid the same quantity of a higher face value）. Canonical ids plain / zhai / fei — English, grep-clean, and no collision with the ones_wild flag.

Also settled by the owner (2026-08-03): canonical ids are ENGLISH（`plain/zhai/fei`, dial ids, reason codes）— easier debugging for a future international team. The zh→en token migration is a read-time normalizer, done now while only dev rooms carry the old tokens.

6 · Done means

• The dice kit plays a full game in an en room — strip, board, ticks, voids, verdict, ending — with no Chinese visible except player-typed text.
• A mixed table（en + zh bidders, one room language）plays clean through the audit driver's five dimensions.
• The zh room's output is byte-comparable to today's（the relocation proved lossless）.
• The shelf lint fails any kit whose lexicon is incomplete for a language it claims.
• 卧底 · 二十问 · 真心话 gain lexicon blocks by the same recipe（follow-up, not this arc）.