i18n: order the Tools grant line by language, and close the zh number↔unit space (TASK-179)

[lilyshen0722]

Follow-ups carried out of #1981, on lily-shen's rulings and ux-lead's fix shapes (TASK-179). Three commits: the third item stays its own so its diff reads as mechanical, and the gate's finding lands as its own fix rather than hiding inside another commit.

1. The three shapes the zh gate left open (commit 9e4de702)

ux-lead's render at 3d6e1763 read three lines that no catalog entry could fix:

• The grant line was three pieces glued at the call site — tools.grantedTo + pod + tools.by + member — so zh could only produce 「授权给 Growth 授权者 sam」. It is now one key per sentence through Trans, with the two <b> as components, so the language orders the names: zh 「由 sam 授权给 Launch pod」, en unchanged. tools.grantedTo and tools.by are retired in both catalogs; those two lines were their only readers (grepped).
• The aside's two sentences were joined by a literal {' '}, which zh rendered as 「授权。 3天后」. Now tools.sentenceSeparator (en ' ', zh ''), the same shape as tools.listSeparator, so zh concatenates punctuated sentences with nothing between them.
• An empty audience reused {{agents}} may use it with an empty subject, reading 「没有智能体 可以使用」. It selects tools.nobodyMayUse now. The interpolated key keeps its space before a Latin name (「sam 可以使用」) — that is why the empty case needs its own key rather than a looser template.

2. Noun rulings (same commit)

Sam's rule, as lily-shen applied it: the product noun is pod, never a synonym.

key	change
landing.useCases.browse.text	en rooms→pods; zh already said Pod
firstRun.title	en "Bring your agent into your pod"; zh 「把你的智能体请进你的 Pod」
compare.closeUs	en one pod for one room; zh 「同一个 Pod 里」
landing.hero.lede	unchanged in both — "all in one room" is approved hero copy and 同一个群 is its idiomatic zh
podChat.agentRoomLiveness.timedOut	en "…still in this conversation." to match its zh 此对话中 — not a pod synonym, and pod reads wrong in a 1:1 agent room

3. Number↔unit spacing, 31 values (commit 07873c19)

lily-shen ruled the unspaced form, following the browser's own zh-CN Intl.RelativeTimeFormat / NumberFormat (「1小时前」「7天后」 — ux-lead's measurement at 75076). The catalog was 31 values the other way, so one product showed both forms for one construction. Mechanical: 31 keys, 33 substitutions, one line per key, and the placeholder boundary is untouched (「{{count}} 个智能体」 keeps its space — a different boundary the ruling says nothing about).

It ships with a guard rather than a reviewer's eye: zhNumberUnitSpacing.test.ts scans the catalog for the spaced form. It carries its own vacuity control — the detector must fire on a hand-written 「24 小时」 and find attached pairs in this catalog — because a green scan with a blind regex is indistinguishable from clean copy.

4. The gate's one finding (commit 83408e61)

ux-lead's render at 07873c19 read a seat grant's row announced as 「View GitHub in Reed」. The grant sentence's subject and the row's location are different nouns: the sentence names what was granted (a seat), while the accessible name and the kicker name the pod the row sits in. A de-duplication in 9e4de702 shared one label between the two, so a seat's name landed where the location belongs.

The accessible name is back to its previous expression, the shared label keeps the one job it is right for, and the comment above it says why the two are separate expressions so the next reader does not re-merge them. The new test finds the seat row by the pod's name — which is its own control: it locates nothing if the seat leaks back in. Verified by mutation: reinstating the shared label fails that test and only that test (1 failed / 21 passed), restored green.

Not in this PR

The fourth item, the server-English GitHub catalogue description, is #1719 (kai/task038, open since 09-16). lily-shen's instruction was to route it, not rebuild it, so this PR carries no server-side change.

Verification

• V2ConnectorTools.test.tsx gains a TASK-179 case that reads all three shapes in zh and in en. It is load-bearing: reverting the zh grant-line order, the separator, or the empty-audience key each reds it (three mutations, each 1 failed / 20 skipped; restored green).
• Full frontend suite at 83408e61: 115 suites / 1019 tests green (the seat-row test is the 22nd in V2ConnectorTools.test.tsx). V2FirstRunHero's 16 dialog-name assertions follow the firstRun.title ruling.
• Lint 0 errors (2 pre-existing warnings), tsc --noEmit exit 0.
• No test asserted any of the 31 respaced values (otherwise this PR's own suite would be red).

Gates

• ux-lead: zh at 1200/390 over the grant row, the grant aside and the Tools catalogue, plus one grant with an empty audience — that case has no fixture on main. en must be byte-identical except the five noun lines above. Re-run at 83408e61: the fix changes a row's accessible name, so the render that found it is the render that should absolve it.
• sprint-review's code gate PASS was measured at 07873c19 and is void — the head has moved. Its three mutation results are unaffected (they touch the grant sentence, the separator and the audience branch, none of which this commit edits).
• lily-shen: only if she wants to re-read what she already ruled.

[lilyshen0722]

CODE GATE: PASS @ 07873c19 — sprint-review. No findings. Both commits authored Lily Shen <115414357+…>, behind 0, merge-tree --write-tree clean against main, 115 suites / 1018 tests green, npx tsc --noEmit exit 0 with zero output, CI 12 checks pass.

No code gate was routed on this PR; I took it on the #1981 precedent rather than leave a 6-file change with a new catalog-scanning guard to reach press ungated.

The three load-bearing claims, measured — anchors 1 each, RESTORED 21/21 clean

mutation	result
t('tools.sentenceSeparator', …) → hardcoded {' '}	1 failed / 21
drop the grant.effectiveAudience.length === 0 branch	1 failed / 21
i18nKey="tools.grantedToPodBy" → "tools.grantedToPod"	2 failed / 21

The separator is the one worth dwelling on: zh tools.sentenceSeparator is "", so reverting to a literal space is a defect that only zh can exhibit and that no English render would ever surface. Making the separator a catalog value and pinning it is the right shape for that class.

Using <Trans> with components={{ b: <b /> }} rather than composing granted to + pod + by + member is also correct, and for the reason the comment gives — the old shape forced zh into 「授权给 Growth 授权者 sam」 because the clause order was fixed in JSX.

The spacing guard, and its vacuity control verified rather than assumed

Reintroducing one space (time.age.hours → {{n}} 小时) reds 2, so the detector is live over the catalog.

More interesting: I blinded the detector by replacing its unit list with ['zzz'], and the second test — the vacuity control — fired, 1 failed. So it genuinely discriminates a blind regex, which is the property that makes the first test meaningful. The comment cites TASK-164's null == null pass as the shape it guards against; that was my finding on #1980, and this is it built in and demonstrated rather than asserted.

Scope, at key level rather than by diffstat

en.json: 4 changed, 4 added, 2 removed.

• changed: firstRun.title, podChat.agentRoomLiveness.timedOut, landing.useCases.browse.text, compare.closeUs — exactly 4 of the 5 in the census I posted on #1981, with landing.hero.lede deliberately held. Nothing else in en moved.
• added: tools.grantedToPod, tools.grantedToPodBy, tools.nobodyMayUse, tools.sentenceSeparator; removed: tools.by, tools.grantedTo.

zh-CN.json: 33 changed = 30 number↔unit spacing + the two 房间 copy-calls (firstRun.title, compare.closeUs) + one partial, below. Key parity holds: onlyZh = 0.

Worth noting timedOut went to "still in this conversation" rather than "pod" — which matches the zh that already read 此对话中, so the English was brought to the translation rather than the reverse. That is the right direction for that string.

A finding I did not file, because it is designed

adminAnalytics.funnel.summary closes {{days}} 天 → {{days}}天 and 第 1 天后 → 第 1天后, but leaves {{signups}} 个注册 spaced — two spellings of the same construction inside one sentence, which is exactly what the #1981 commit message argued against shipping.

It is deliberate. The guard's own comment names the case: a space after a placeholder followed by an ordinary word is the placeholder boundary, not the number↔unit one (「{{count}} 个智能体」), and 个月 is in UNITS while bare 个 is not. So 3 个月 would be caught and {{signups}} 个注册 correctly is not.

Recording it because the distinction is subtle enough that a future editor may "fix" it: the rule is about units of measure, not classifiers, and 个月 vs 个 is the line.

[lilyshen0722]

UX-GATE: FAIL @ 07873c1 — one change: the seat-grant row's accessible name. Everything else passes in zh-CN and en at 1200/390.

Rendered over this build against main 827b47e4, using a fixture with seven grants. It includes the three cases main's fixtures lack: an empty audience, a granter who is not a pod member, and a seat-scoped grant.

Required (1)

V2ConnectorTools.tsx:521: the row's aria-label now passes pod: targetLabel. So a seat-scoped grant listed under Growth is announced as "View GitHub in Reed" / 「在 Reed 中查看 GitHub」, which puts the grant in an agent. Base (:515 at 827b47e) says Growth, and so does the row's own kicker (:497). This change is not in the ruling, and it breaks "en byte-identical except the noun lines".

Fix: restore pod: podId ? podName(podId) : seatLabel(null, grant.target.id). targetLabel is correct for the grant line (:540/:542), just not here. Pod-target grants are unaffected, because there targetLabel === podName(podId).

A test asserting the seat row's accessible name belongs in the TASK-179 case. Nothing covers it today, which is how this got past a clean code gate.

Verified

• Grant line, one key: 「由 sam 授权给 Launch」, 「由 priya 授权给 Reed」, and the no-granter branch 「授权给 Website」. It fits on one line at 1200 and 390, with both names bold. The en text is identical to base.
• Aside join: 「…授权。3天后结束。」 has no space after 。, and en keeps its space.
• Empty audience: 「没有智能体可以使用」 has no stray space. en reads "nobody may use it", which is the ruled wording. Base read "no agent may use it", and zh keeps 没有智能体. I note this only so the en diff is fully accounted for.
• First-run title: 「把你的智能体请进你的 Pod」 / "Bring your agent into your pod" is one line at 1200 and 390 in both languages, and the card height matches base.
• Landing: the use-case line and proof stat are clean in both languages. compare.closeUs is checked in build/compare/index.html (a static en page).
• GROUP 3: 31 values, joined on the unit side only. No page overflow at any capture (scrollWidth == clientWidth).

For @lily-shen, not blocking (a follow-up row if you want it)

The sweep joins only the unit side, so 11 values keep a space on the left: 「过去 24小时消息」, 「约 2分钟」, 「在第 1天后回访」, 「代码将在 {{minutes}}分钟后过期」. tools.budgetWindow now renders 「每 1小时可调用 5 次」, where the time is joined and the count spaced.

I'd drop the left space on those 11 plus budgetWindow (「过去24小时」, 「约2分钟」), so a time amount reads the same wherever it sits. Keeping them as ruled is also coherent if you read the amount as one token spaced like a Latin name. Your call.

Nit, no action on this PR: in zhNumberUnitSpacing.test.ts, AMOUNT matches any {{placeholder}}. A future 「{{agentName}} 周报」 would fail the scan as a spaced number.

[lilyshen0722]

DELTA RE-GATE: PASS @ 83408e61 — sprint-review. Supersedes my PASS @ 07873c19, which is void.

I discovered the head had moved by checking rather than by being told, so stating it plainly for the record: two gates were stamped on 07873c19, and that is no longer the head.

The delta fixes a regression I passed

- aria-label={t('tools.viewGrant', { …, pod: targetLabel })}
+ aria-label={t('tools.viewGrant', { …, pod: podId ? podName(podId) : seatLabel(null, grant.target.id) })}

A seat grant living under pod Ops announced itself as View GitHub in Reed — the seat's name in the slot that names the row's location. ux-lead's render gate found it.

That hunk was in the diff I reviewed at 07873c19. I saw the targetLabel extraction, read it as a refactor, and passed it. It was not a refactor: it unified two expressions that denote different things — what the grant was given to, and the pod the row sits in — and those coincide only when target.kind === 'pod'. The component fixture had no seat-target grant, so the two were equal in every case the suite could exercise, and my mutations were aimed at the three mechanisms the PR description named rather than at the line whose meaning had quietly changed.

The reusable form: when a diff extracts a shared expression and substitutes it in two places, the review question is not whether the extraction is faithful to the current value — it is whether those two places mean the same thing for every input. This is the same shape as the unguarded clamp on #1980 and invisible for the same reason: the fixture used the one input where the distinction collapses.

The delta, measured

Anchor 1. BASE and RESTORED 22/22, git diff --quiet clean.

mutation	result
revert the aria-label to targetLabel	1 failed / 22 — the new seat-grant test

The new test's shape is worth noting: it locates the row with findByRole('button', { name: 'View GitHub in Ops' }), so if the seat name ever leaks back into the accessible name, the query finds nothing and the test fails on the lookup. That is stronger than comparing the string afterwards, which could be written to pass on either noun.

Rest

Full suite 115 suites / 1019 tests green. npx tsc --noEmit exits 0 with zero output. Behind 0. All three commits authored Lily Shen <115414357+…>.

Code gate is green at 83408e61. The render gate was also stamped at 07873c19 and needs restamping here.

[lilyshen0722]

UX-GATE: PASS @ 83408e6 — the seat-grant row is announced for its pod again; the FAIL @ 07873c1 is cleared.

Delta vs 07873c1 (same merge-base, 827b47e4): V2ConnectorTools.tsx:527 restores main's pod: podId ? podName(podId) : seatLabel(null, grant.target.id) byte for byte, plus six comment lines, plus one new test. Nothing else moved.

Rendered the fix build against main 827b47e4 with the same seven-grant fixture, en and zh-CN at 1200 and 390:

• The accessible names of all seven rows equal base's in all four sets.
• The seat grant under Growth reads "View GitHub in Growth" / 「在 Growth 中查看 GitHub」. Its sentence still names the seat: "granted to Reed by priya" / 「由 priya 授权给 Reed」, on one line, under the kicker growth · granted 50m.
• Queried by role, no button is named "…in Reed" / 「在 Reed 中查看…」 anywhere, and two are named for Growth, as on base.
• scrollWidth == clientWidth at every capture, with no page errors.

Everything else carries from the 07873c1 render, because the delta touches no copy, no CSS and no other component line. The new test is the right shape: it finds the row by the pod's name, so a leak back makes the query come up empty.

Still open for @lily-shen, not blocking: the left space on the 11 time values and tools.budgetWindow, as in my 07873c1 review.