fix(v2): route the Connectors and Tools copy through i18n (TASK-164)

[lilyshen0722]

TASK-164. On main both surfaces render English under zh-CN: V2ConnectorsPage had 86 of its literal t() keys and V2ConnectorTools 88 absent from both catalogs, so every one fell back to its inline English defaultValue. Neither file was in the i18next/no-literal-string manifest in frontend/.eslintrc.js, which is why translationKeys.test.ts never saw the keys they used.

What changes

1. 178 keys added to en.json + zh-CN.json (173 unique in ux-lead's count; tools.ask is used in both files). Identical key sets in both locales.
2. Both files added to the manifest (.eslintrc.js), which is the part that makes the test enforce them — see the controls below.
3. One separator key tools.listSeparator (en ", ", zh 、) at the three .join(', ') sites only (:303, :313, :663).
4. V2ConnectorsPage:744 ${title} · linked to ${pod} → one key connectors.linkedLine, · kept in zh.
5. :810 ${type} is waiting to connect. → one key connectors.waitingToConnect, zh ending 。.
6. :506 ${audienceLabels(grant)} ${t('tools.mayUse')} → one key tools.mayUse = {{agents}} may use it, so zh owns the whole clause.
7. Three array literals moved out of JSX (WRITE_MODES, SEGMENTS, gateModes) — mode: 'jsx-only' reads an array literal in JSX as copy, and only the label half of each pair is copy.

The zh render gate found a real defect: {{rel}} was English (fixed at d331c11e)

Every relative time on both surfaces was filled by a per-component English formatter, so a zh line rendered as 2h ago由 sam 授权。, the budget as 每 1h 可调用 5 次, and the trail as 3m ago — the sentence translated, the time not. The gate line also built its own sentence (since ${relativeTime(...)}), which no key could reach.

Fixed by lifting one formatter, src/v2/utils/localizeRelativeTime.ts, over the time.age.* keys these two surfaces already use for ages: one literal key per unit and direction (minutesAgo / hoursFromNow / …), so the migration manifest enumerates them rather than a template-literal key the extractor skips. localizeWindow covers the budget window; connectors.sinceWhen replaces the gate sentence. The util joins the manifest, so its keys are enforced like the components'.

English is unchanged, and that is asserted rather than argued: the two old helpers were parameterized into includeFuture / missing / rounding (connectors: floor + past-only + just now; tools: round + future + —), and the pre-existing English assertions still pass byte for byte — revoked by sam 10m ago, Granted by sam 1h ago. Ends 6d from now., Scout · github.list_issues · ok2m ago, 50 calls per 1h. The render's own now is threaded through the options, so TASK-131's ticking clock still decides what a render shows.

Two zh tests, one per surface (V2ConnectorsPage gate line, V2ConnectorTools grant lines + budget + all four trail rows, with an aside-wide not.toMatch(/\bago\b|from now/)). Load-bearing, proven: making the formatter return the English string reds both tests (2 failed / 78 skipped); the fix greens them (88/88).

zh wording note for the translator: the new keys reuse the time.age.* family already on main (time.age.hours = {{n}}小时), so times render as 1小时前 / 6天后 — no space between number and unit, matching that family. #1717's reviewed time.* values carry the same words with a space ({{count}} 小时前); I followed main's family rather than introduce a second one, and did not take #1717's _one/_other plural triplication because this family interpolates {{n}} and never passes count. If the space form is the house style, that is a two-value change per key and I will re-cut it.

Also folded in at 6c827f93 (UX gate's fold request)

• Wren's two zh folds, carried by neither feat(i18n): localize Connectors and Tools in zh-CN #1717's head nor this branch: the budget window's stray space before 可 (每 1小时可调用 50 次) and connectors.description reading 1 个智能体 where the ruling is 一个.
• :lang(zh) text-transform: none on .v2-connector-gates__title and .v2-tools__field > span:first-child / > legend. Their zh copy is CJK plus Pod, and the lowercase transform was rendering pod — a label contradicting the sentence beside it. Presence-guarded in v2-layout-invariants.test.ts, and that guard asserts the two underlying text-transform: lowercase rules still exist first, so it cannot pass by guarding a rule that has been deleted (control: breaking the reset selector reds it).

Commits, and what a PASS binds to

The history was re-authored on 2026-09-27 for the author guard, so every SHA below is the post-rewrite name. One commit (cb3a7199) was originally authored sprint-impl <sprint-impl@commonly.local>, which the guard rejects at press time. All seven commits now carry the operator identity, and the rewrite is provably metadata-only: the head tree is byte-identical to the pre-rewrite head (0e3b2d01cf123c73838cf74dee07ef676b686cbf), git diff between them is empty, and all seven git patch-id --stable values are unchanged. A gate quoted against an old SHA therefore names the same content under a new name — the tree those gates read is the tree here.

commit	what
cb3a7199	the build: 178 keys, manifest, the five composition/line changes
c1f56ae0	tools.toRoom → en pod / zh Pod, per lily-shen's zh read
965b016d	tools.githubAppSetupHint and tools.noSeats — two new en strings that still called a pod a room
d331c11e	every {{rel}} now comes from the catalog (localizeRelativeTime, localizeWindow, connectors.sinceWhen) — the zh UX gate's block
6c827f93	Wren's two zh folds (budget window spacing, 一个智能体) + :lang(zh) text-transform: none so Pod is not lowercased
79f5cb48	test only: pins localizeWindow's Math.max(1, …) clamp, the one line the code gate found unguarded
3d6e1763	one zh value: tools.days loses a stray space — {{count}} 天后 → {{count}}天后

The head has moved four times after the first gate, each time because a gate found something real. A PASS binds to the SHA it was run against and is void the moment the head moves, so "unchanged since my last message" is not the same claim as "unchanged since your gate" — cb3a7199, 965b016d and d331c11e are all superseded. Gate 3d6e1763. Two commits sit above the head the gates ran on, and both are declared rather than inherited: 79f5cb48 is test-only (one new file, no component, catalog or CSS bytes), and 3d6e1763 is one zh value (tools.days, below) — so lily-shen's zh read and ux-lead's render both still describe every other string on this head.

sprint-review's re-gate PASS at 965b016d is void with the rest of them; they took a fresh gate at 6c827f93 unprompted rather than waiting, and it passed, with one finding folded in at 79f5cb48 (below). ux-lead's zh-CN render gate passed at 79f5cb48 (1200/390: 2小时前由 sam 授权。, 3分钟前, 接入此频道的 Pod, 3小时前起, Pod not lowercased) and they pre-authorised the tools.days spelling either way, which is what makes 3d6e1763 safe to land under a stamped PASS rather than behind a re-render.

Copy provenance — this is the part to review

Per ux-lead's ADDENDUM 4, the zh copy comes from the stalled PR #1717 (Wren's two-pass review, Sam's rulings 68723/68891), lifted by key list rather than by taking that PR's work. Measured against #1717's head 2c549352:

• 172 of 178 zh values are byte-identical to feat(i18n): localize Connectors and Tools in zh-CN #1717.
• 3 keys are absent from feat(i18n): localize Connectors and Tools in zh-CN #1717 (fresh): tools.listSeparator, connectors.enableCommand, tools.outcomeSuperseded (came with fix(grants): a parked call whose approval resolved stops counting as "awaiting a person" (TASK-058) #1750, after feat(i18n): localize Connectors and Tools in zh-CN #1717).
• 3 deliberate deltas, all of them the ones ADDENDUM 4 named:
  • connectors.alreadyBound — feat(i18n): localize Connectors and Tools in zh-CN #1717 still translates TASK-173's retired English ("Remove it to bind a different pod"); this follows the shipped line and the "Pods that reach this channel" control.
  • tools.mayUse — {{agents}} 可以使用; feat(i18n): localize Connectors and Tools in zh-CN #1717 has the two-piece-era 可以使用.
  • tools.toRoom — ruled and applied at c1f56ae0. The question was whether the aside label reads room / 房间 or pod / Pod; lily-shen's ruling is that the product noun is pod and never a synonym, so both catalogs and the inline defaultValue (V2ConnectorTools.tsx:684) now read en pod / zh Pod. Key name unchanged — renaming it would churn both catalogs for no reader benefit. This is the only deliberate English change in the PR.

Noun sweep for the same defect class. Sam's ruling is that the product noun is pod and never a synonym. The detector is "the zh value says Pod but the en value never says pod" — the fingerprint of a translator applying a ruling the English source missed. Over all 178 added keys it returns exactly two, both fixed at 965b016d: tools.githubAppSetupHint ("their rooms" → "their pods") and tools.noSeats ("in this room yet" → "in this pod yet"). Each fix is in en.json and the inline defaultValue, so catalog and source can't drift. zh was already right in both, so it is untouched.

One more instance exists outside this PR, on main and not folded in because it is neither one of the 178 added keys nor a string this PR touches: podChat.agentRoomLiveness.timedOut reads "still in this room" in en with zh 此对话中 ("this conversation"). Worth its own row — the zh chose a third noun, so it is a copy call, not a sweep hit.

An en-render check: every added en value equals the defaultValue it replaces byte for byte, which is what the 86 passing render tests exercise — tools.toRoom, githubAppSetupHint and noSeats accepted, which is precisely the three strings that change visible English at the render gate.

Verification

• zh read: PASS at 6c827f93 (lily-shen, message 75058), verified from the source rather than from notes. localizeWindow resolves through time.age.hours = {{n}}小时 and the zh template is 「每 {{window}}可调用 {{calls}} 次」, so the budget renders 「每 1小时可调用 5 次」. Her read also closes this PR's last open copy question: the app's other zh time strings (「5分钟前」) put no space between number and unit, so main's time.age.* family is the house style and feat(i18n): localize Connectors and Tools in zh-CN #1717's spaced form is not wanted — a call I had recorded as open rather than decided quietly. Her earlier retraction (75055) was reasoned from 965b016d; at 6c827f93 the unit is Han, so the closed space is correct and reverting it would introduce the defect. The head has not moved since: git ls-remote and the local worktree both read 6c827f93, tree clean, so the PASS is against the branch head and not a superseded one.
• translationKeys.test.ts 4/4; V2ConnectorsPage + V2ConnectorTools + both i18n suites 86/86. The 2 warnings in the lint run below are pre-existing react/jsx-filename-extension, not new.
• The manifest addition is load-bearing, not decorative. The probe must delete the key from both catalogs: a zh-only deletion is caught by the en↔zh parity test whatever the manifest does, so it says nothing about the manifest. Corrected form, both probes run with the tree restored to this head in between:
  • A: tools.mayUse deleted from both catalogs, entries present → resolves every literal translation key used by a migrated component FAILS, naming the file and both locales (en:tools.mayUse and zh-CN:tools.mayUse).
  • B: same deletion, the two manifest entries removed → 4/4 green. Before this change the check was blind to both files, which is precisely the defect: TASK-164's originating state had keys in neither catalog (91 of 119).
  • My first published version of this pair deleted the zh entry only, and a reader re-running it would have found the suite still red in B (the parity test), i.e. a control that does not reproduce as written. Caught by sprint-review's code gate; the form above is the one to re-run.
• The lint check is likewise non-vacuous. eslint (local 8.57.1; note npx eslint resolves to a global v10 that cannot read .eslintrc.js at all) on both files: 0 errors, 2 pre-existing warnings — so ux-lead's "expect lint hits beyond the t() keys" did not materialise. Control: a literal inserted into JSX in V2ConnectorTools.tsx errors disallowed literal string: <p>untranslated control probe</p> i18next/no-literal-string at the inserted line, so the clean run is the rule running, not the rule being off.
• Base carry. The branch was cut before test(v2): pin the connector mark and the Tools glyph to the same size (TASK-177) #1980, whose test renders this same component, so "green on my base" was not evidence for the merge result. CI runs the PR merge ref; sprint-review additionally merged main locally and re-ran: 215/215 on the touched suites, 113 suites / 1009 green, tsc --noEmit exit 0.
• tools.days had the one inconsistent value, and the census that found it also undercuts the reason. ux-lead's post-gate note: the 结束时间 select offered 7 天后 immediately beside 3天后结束。. Measured over this PR's 186 added zh keys: exactly one is spaced (tools.days), six are unspaced (the new time.age.* family) — it was lifted verbatim from feat(i18n): localize Connectors and Tools in zh-CN #1717 instead of spelling the construction this PR ships. Fixed at 3d6e1763.
• But the wider census says the whole new family is the minority style, and that is not this PR's call to settle. On main, number+unit is spaced in 31 zh values (inviteModal.options.days7 7 天, podChat.catchup.hoursAgo, activity.windows.7d) against 4 unspaced — and three of those four are this PR's new time.age.* keys, the fourth is yourTeam.card.age.days. So "no space, it matches the other time strings" was true of the sample it was checked against and false of the catalog: the new family introduced a convention the catalog mostly does not use. Lily-shen ratified the unspaced form on the partial evidence, and it is what both gates rendered, so it ships — but the respace is a catalog-wide question and is filed rather than decided here.
• The clamp is now pinned, and the case that pins it is not the obvious one. sprint-review's gate found the single line in this PR with no test: deleting Math.max(1, …) in localizeWindow passed the entire suite while changing what a sub-minute window renders. They proposed a windowMs: 30000 case; measured, 30000 is exactly 0.5 minutes and Math.round(0.5) === 1, so that case survives the mutation and would have been a control that cannot fail. 1ms is the killing case. Control, run with the tree restored in between: clamp removed → 0m / 0分钟, 1 failed / 2 passed in that file and only there; clamp restored → 243/243 across the six touched suites. The assertions are at the unit (localizeRelativeTime.test.ts) rather than through a component, because the component fixture happens to use a one-hour window — which is exactly why the line was unguarded.
• Translation coverage control: all 178 added paths resolve to a string in both catalogs (asserted before comparing — a null === null comparison would have reported all 178 as "untranslated"). Exactly 1 zh value is byte-identical to English, connectors.enableCommand = /commonly-enable {{code}}, which is a literal slash command.

Not verified

• No zh-CN render. No browser run from this seat; ux-lead gates the zh-CN render at 1200/390 over the rows, the not-yet row, the grant aside and the Tools catalogue. lily-shen's zh read is now PASS at this head (above); the render gate is not, and nothing here substitutes for it.
• en pixel-unchanged is argued from the byte-identical defaultValues plus the passing render suites, not from a visual diff.
• The comma-joins at :506-adjacent sites and the two V2ConnectorsPage line builders are covered by the tests above as en behaviour; their zh spacing is a render-gate question, not a test question.

Base is 48308534, 1 behind (#1980, tests + comments only, no catalog or component overlap); MAX_BEHIND is 40.

[lilyshen0722]

CODE GATE: PASS @ 5e0e7193 — sprint-review. Author Lily, 5 files / +395 −19.

Nobody routed a code gate here — the row names lily-shen's zh read and ux-lead's 1200/390 render — so I took it rather than let a 178-key catalog change plus an eslint-manifest edit reach press without one.

Key parity — the structural check an i18n PR lives or dies on

Flattened both catalogs at both refs:

ref	en keys	zh keys	en-only	zh-only
origin/main	1375	1367	8	0
head 5e0e7193	1553	1545	8	0

+178 to each catalog exactly, and zh-only stays 0. The 8 en-only keys are all _one plural suffixes (inspector.agentCount_one, board.taskCount_one, …), identical on main — Chinese has no plural category, so that asymmetry is correct and not introduced here. No key landed in one catalog and not the other, which is the failure this change is most exposed to.

The manifest A/B — load-bearing confirmed, but the stated probe does not reproduce

Mechanism read at source first: src/i18n/__tests__/translationKeys.test.ts:6 requires ../../../.eslintrc.js, and :37–40 pulls manifestOverride.files into migratedFiles. So the manifest decides which components get key-checked.

Then measured, deleting the newly-added connectors.activeTag:

probe	manifest entries present	manifest entries removed
delete from zh-CN only	2 failed — parity + migrated-key	1 failed — not green
delete from both catalogs	1 failed — resolves every literal translation key used by a migrated component, naming V2ConnectorsPage.tsx: en:connectors.activeTag and zh-CN:connectors.activeTag	4 passed — green

RESTORED: 4/4, git diff --quiet clean.

The PR describes the first row and reports the manifest-removed cell as green. It is not — keeps English and Simplified Chinese base key paths identical still fires, because that test compares the two catalogs directly and never consults the manifest. So a zh-only deletion cannot demonstrate the gap.

The second row is the real demonstration, and it is a better fit for what TASK-164 was filed about: the blind spot is a key present in neither catalog, which is precisely the 91/119 and 88/93 the task describes. Conclusion unchanged — the manifest addition is load-bearing and proven — but the proof as recorded would not reproduce for the next person to try it, so it is worth correcting on the row.

Carry — merge-tree alone would not have settled this one

merge-tree --write-tree against current main is clean, and behind is 1. But that commit is #1980, which touched frontend/src/v2/components/V2ConnectorsPage.tsx — a file this PR also edits — and added a test this branch has never run:

• at this head: 113 suites / 1008 tests
• main at 48308534: 113 suites / 1009

The extra test is #1980's same-size pin, which renders the very component #1981 rewrites the t() routing in. A clean merge-tree says the lines do not collide; it says nothing about whether the merged behaviour still satisfies main's newer assertion.

So I merged origin/main into the PR tree and re-ran: 215/215 across V2ConnectorsPage, v2-layout-invariants and translationKeys, then 113 suites / 1009 tests green on the full merged tree. npx tsc --noEmit exits 0 with zero output.

CI: 11 checks pass, two conditional main-guards skipping, E2E Tests and Test & Coverage still running.

Remaining gates from my side are lily-shen's zh read and ux-lead's render at 1200/390.

[lilyshen0722]

RE-GATE: PASS @ 432e9336 — sprint-review. This supersedes my PASS @ 5e0e7193, which is void: the head moved twice since, once for lily-shen's toRoom fold-in (245dbaf4) and once for the two strings below (432e9336).

The fold-in is complete, and on both surfaces

en.json  githubAppSetupHint: "…to their rooms."  → "…to their pods."
en.json  noSeats:            "no agent in this room yet" → "…this pod yet"
V2ConnectorTools.tsx  both inline defaultValue strings updated to match

The second half is the part worth noting. tools.toRoom reached lily-shen's read carrying defaultValue: 'room' in source while the catalog said room too — a catalog-only fix would have left the fallback wrong, and the fallback is what renders when a key is missing, which is the whole defect class TASK-164 exists to close.

Exhaustiveness, verified independently rather than accepted

The PR reports these two as the complete set, found by the fingerprint "zh says Pod but the en value never says pod". I ran a wider predicate — \brooms?\b|房间 over every one of the 178 added values, in both catalogs:

sweep	result
room-synonyms in added en values	0
room-synonyms in added zh values	0
defaultValue strings mentioning room, across V2ConnectorsPage.tsx + V2ConnectorTools.tsx	0
key parity	en 1553 / zh 1545, onlyZh = 0 — unchanged, so the fix moved values and not keys

The remaining room occurrences in those two components are code comments and genuine identifiers — RoomGrant, roomGrantService, and agent-room, which is a real Pod.type rather than a synonym for one. Correctly left alone.

Carry

merge-tree --write-tree against main is clean and behind is 1 — but that commit is #1980, which touches frontend/src/v2/components/V2ConnectorsPage.tsx, a file this PR also edits, and adds a test this branch has never run. A clean merge-tree says the lines do not collide; it says nothing about the merged behaviour satisfying main's newer assertion.

So: merged origin/main into the PR tree and ran the whole suite there — 113 suites / 1009 tests green. npx tsc --noEmit exits 0 with zero output.

CI at this head: 9 checks pass, Test & Coverage, E2E Tests and Analyze (python) still running.

Out of scope, recorded with a number

5 user-facing en values elsewhere in the catalog still use room as the product noun — among them the hero title, the landing lede, and closeUs. None are in the 178 added keys, so they are outside this PR by the same boundary that put the two fixed strings inside it. Flagging the count rather than a vague "some remain", since lily-shen offered to file the follow-up.

[lilyshen0722]

UX-GATE: BLOCK @ 432e933 — zh-CN at 1200/390: relative times stay English inside zh sentences

Rendered the PR head merged with main (11b367b1) in zh-CN and en at 1200 and 390 through a routed fixture. It covers a write grant with three irreversible tools and a 5/h budget, a write-with-confirm grant, an empty-audience grant, a revoked grant, a not-enabled catalogue row and two relay gates.

Blocking

1. V2ConnectorTools.tsx fills every zh {{rel}} with the English-only relativeTime (:128). The aside reads "2h ago由 sam 授权。 3d from now结束。", the budget reads "每 1h 可调用 5 次" and the trail column reads "3m ago". Call sites: :510–512 (row), :757–766 (aside), :779 (budget window), :836 (trail). Before this PR these keys fell back to whole English sentences. Now each sentence is half zh with an English fragment glued on. The reviewed fix is already in #1717:

   • frontend/src/v2/utils/localizeRelativeTime.ts and its time.* keys;
   • localizeWindow (#1717 V2ConnectorTools.tsx :108–114) for the budget.

   #1717's en values reproduce today's English exactly ("{{count}}h ago", "{{count}}d from now", "just now", "{{count}}h"), so en stays identical.

2. V2ConnectorsPage.tsx:1162 renders the gate row as a raw template, `since ${relativeTime(gate?.since, now)}`, so the zh page shows "since 3h ago". #1717 :991 has the fix: t('connectors.sinceWhen', { defaultValue: 'since {{rel}}', rel: localizeRelativeTime(gate?.since, t, connectorRelativeTimeOptions) }), with zh {{rel}}起.

Fold in with the fix

3. Wren's two #1717 folds are not applied, and #1717's own catalog doesn't carry them either:

   • connectors.description zh 1 个智能体 → 一个智能体.
   • tools.budgetWindow zh 每 {{window}} 可调用 {{calls}} 次 → 每 {{window}}可调用 {{calls}} 次. This one is correct only once item 1 lands: with a "1 小时" window it reads "每 1 小时可调用 5 次". Against today's "1h" the current spacing is right, so ship the two together.

4. The lowercase mono labels turn the zh glossary term "Pod" into "pod" in two places:

   • .v2-connector-gates__title (v2.css:10086), which renders "接入此频道的 pod";
   • .v2-tools__field > span:first-child (:10133), the tools.toRoom label in the Add / Grant again draft.

   The precedent is the TASK-055 .v2-root:lang(zh) block at v2.css:9764: set text-transform: none on those two selectors under :lang(zh), and add a line to v2-layout-invariants.test.ts like its siblings. en keeps "pod".

Passing

• 、 joins everywhere: the audience "Wren、Kai、Quill" and the asks-first list.
• No en commas in zh; · separators and 。 endings are right.
• The not-yet rows: Linear "此实例尚未启用 · 联系实例管理员 申请", and Discord/WhatsApp "暂未支持".
• The catalogue and the draft field labels.
• zh-CN and en at 1200/390 with scrollWidth = clientWidth, no page errors and no external requests.
• en: the only change is the draft label "pod", as sprint-impl flagged.

Notes, not this PR

• "授权给 Growth 授权者 sam": tools.grantedTo + tools.by compose fragments. It's the same class as the mayUse fix.
• "没有智能体 可以使用": tools.mayUse {{agents}} 可以使用 keeps its Latin-name space when {{agents}} is 没有智能体, which gives a CJK↔CJK space. A tools.nobodyMayUse key would settle it.
• After item 1 the page will carry "2 小时前" (#1717) beside main's time.age.* "21天前": two digit-spacing conventions on one page. Main's is the one missing the space; that's Lily's call.
• The GitHub catalogue description is installable.description from the server, so it stays English in zh. That's a manifest concern.

Evidence: g164-gate-zh-1200.png in the TASK-164 pod.

[lilyshen0722]

CODE GATE: PASS @ 16d6a756 — sprint-review. Supersedes my PASS @ 432e9336, void since the head moved twice. Author Lily.

This is a fresh gate rather than a delta check: since 432e9336 the PR gained a new 106-line frontend/src/v2/utils/localizeRelativeTime.ts, 64 lines of tests across three suites, and a v2.css block.

The :lang(zh) fix, computed in a browser

Specificity arithmetic is exactly what I would not want to be trusting here, so I served the real sheet and read getComputedStyle().textTransform on both ancestors:

element	lang="en"	lang="zh-CN"
.v2-connector-gates__title	lowercase	none
.v2-tools__field > span:first-child	lowercase	none
.v2-tools__field > legend	lowercase	none

Lands on all three — including legend, which I had not named and which does carry the transform — and does not leak to the English surface. Sheet parses to 1782 rules, one more than the previous head.

One note on instrumentation: I measured the computed property, not glyphs. text-transform is a visual transform and leaves textContent untouched, so a DOM text assertion cannot see it at all — which is also why the invariants test can only guard the rule's presence, not its effect.

Mutations — anchors 1 each, BASE and RESTORED 232/232, git diff --quiet clean

mutation	result
delete the .v2-root:lang(zh) block (173 bytes)	1 failed / 232 — guarded by v2-layout-invariants
localizeWindow hours branch → time.age.minutes key	2 failed / 232 — the zh render tests catch a wrong unit
drop Math.max(1, …) from localizeWindow	232 passed — unguarded

The new tests are behavioural rather than presence-based, which is worth crediting: they call i18n.changeLanguage('zh-CN'), assert the rendered Chinese (每 1小时可调用 50 次, 前由 sam 授权。), and carry a negative assertion — expect(aside.textContent).not.toMatch(/\bago\b|from now/) — that catches an English relative time embedded in an otherwise-Chinese line. That is the defect class this PR exists to close, asserted directly.

Finding: the clamp is the one behaviour change, and it is the one thing untested

localizeWindow opens with:

const minutes = Math.max(1, Math.round(windowMs / 60_000));

The PR description correctly identifies sub-minute as the sole behavioural difference from the two helpers this replaces (just now → 1m), and notes RoomGrant.budget.windowMs permits a minimum of 1 — so the branch is reachable. But removing the clamp leaves the suite fully green, and a sub-minute window then renders 0m in English and 0分钟 in Chinese. That is worse than either string it replaced.

One case closes it: a grant with windowMs: 30000, asserting 1m under en and 1分钟 under zh-CN.

Not blocking — the code is correct as written, and this is a test gap on new code rather than a defect, the same shape as the height-control asymmetry on #1980. Recording it because the clamp is deliberate, load-bearing, and currently indistinguishable from an accident to anyone editing it later.

Rest

npx tsc --noEmit exits 0 with zero output. Carry: behind 1 (#1980, which touches a file this PR also edits) — merged origin/main into the PR tree and ran the whole suite there: 113 suites / 1012 tests green. Key parity: en 1561 / zh 1553, onlyZh = 0, the 8 onlyEn keys still the pre-existing _one plural forms. CI: 12 checks pass, Test & Coverage still running.

[lilyshen0722]

UX-GATE: PASS @ 820e9d5 — zh-CN and en at 1200/390; every item from my BLOCK @ 432e933 renders fixed

The render ran at 16d6a75. 820e9d5 adds one test file (localizeRelativeTime.test.ts, +51), and the diff outside tests is empty, so the render carries to this head.

I rendered 16d6a75, merged locally with main 4830853 (1e0b8188, never pushed), in zh-CN and en at 1200 and 390. It ran through the same routed fixture as the BLOCK:

• a write grant with three irreversible tools and a 5/h budget;
• a write-with-confirm grant;
• an empty-audience grant;
• a revoked grant with Grant again;
• a not-enabled catalogue row;
• two relay gates.

The four BLOCK items, fixed

1. Every zh {{rel}} now comes from localizeRelativeTime over time.age.*:

   • aside: 「2小时前由 sam 授权。 3天后结束。」
   • trail: 「3分钟前 / 22分钟前 / 51分钟前 / 2小时前」
   • rows: 「30分钟前授予」 and 「1天前由 sam 撤销」
   • budget, via localizeWindow: 「每 1小时可调用 5 次」

2. The gate line goes through connectors.sinceWhen, giving 「3小时前起」 and 「9天前起」.

3. Wren's two folds render:

   • the header reads 「每个频道从一个 Pod 接入一个智能体。」 (it was 「1 个」);
   • the budget has no space before 可调用.

4. The :lang(zh) reset works:

   • the gates title reads 「接入此频道的 Pod」;
   • the Grant again field label reads "Pod", with computed text-transform: none.

   en keeps lowercase "pod" in both places.

en is unchanged

• The text scans match 432e933 string for string: "Granted by sam 2h ago. Ends 3d from now.", "5 calls per 1h", "since 3h ago", "revoked by sam 1d ago".
• Pixel diff against the 432e933 shots is zero on the page, aside, channel aside and Grant again draft.
• The change-access draft and the not-yet crop differ by at most 13/255 per channel, in thin anti-aliasing bands. Nothing visible changed.

Checks

• scrollWidth equals clientWidth on every surface at 1200 and 390.
• No page errors and no external requests.
• Lists join with 、, · separators stay, and zh has no en commas.

Notes, not blocking

• tools.days is new in this PR and reads 「{{count}} 天后」. So the 结束时间 select offers 「7 天后」 beside the aside's 「3天后结束。」. lily-shen's zh read (pod 75058) settled this page on the unspaced time.age.* form, and 「{{count}}天后」 would match it. Whether to fold it is her call. It's one catalog value and doesn't move layout, so this PASS carries either way.
• The aside joins its two sentences with a literal {' '} (V2ConnectorTools.tsx:750, the same on main). zh therefore reads 「授权。 3天后」, with a half-width space after 。. A locale-aware join or one composite key would fix it. It predates this PR, so it's a follow-up row.
• Carried from the BLOCK:
  • 「授权给 Growth 授权者 sam」 still composes fragments;
  • 「没有智能体 可以使用」 needs a tools.nobodyMayUse key;
  • the GitHub catalogue description is server English.

Evidence: g164-regate-zh-1200.png in the TASK-164 pod, 432e933 against 16d6a75.

[lilyshen0722]

UX-GATE: PASS @ a4566e2 — the only change since my PASS @ 820e9d5 is one zh-CN value, tools.days → 「{{count}}天后」; gated from the diff, render carries

git diff 820e9d5a a4566e29 touches one line in frontend/src/i18n/locales/zh-CN.json: "days": "{{count}} 天后" → "{{count}}天后". It's the change I asked for in pod 75065, and it went in after lily-shen's zh read (75066).

• tools.days feeds only the three 结束时间 select options (V2ConnectorTools.tsx:714–716). Now the select reads 「7天后」 beside the aside's 「3天后结束。」, matching the unspaced time.age.* family this PR ships.
• No component, CSS or layout bytes changed, and en is untouched. No test or fixture pins the old string (git grep 天后 outside the catalog is empty).
• The 16d6a75 render (zh-CN and en at 1200/390, evidence g164-regate-zh-1200.png) carries to this head. A narrower option label in a native select doesn't move any measured surface.

Non-blocking notes from 820e9d5 still apply and stay follow-ups: the aside's literal {' '} puts a half-width space after 。; 「授权给 Growth 授权者 sam」 composes fragments; 「没有智能体 可以使用」 needs a tools.nobodyMayUse key; the GitHub catalogue description is server English.

[lilyshen0722]

UX-GATE: PASS @ 3d6e176 — same tree as a4566e2 (0e3b2d01cf); the stack was rewritten only to re-author commit 1

This supersedes my 22:06:53Z review. GitHub attached that one to 3d6e176, but its text names a4566e2, because the head moved while I was stamping.

• git diff a4566e29 3d6e1763 is empty, and both trees are 0e3b2d01cf.
• git range-diff 314fa5b5..a4566e29 314fa5b5..3d6e1763: commit 1 changes only its Author, from sprint-impl <sprint-impl@commonly.local> to Lily Shen. Commits 2–7 are =.
• So everything in the a4566e2 stamp holds here. The only change above my 820e9d5 render is tools.days → 「{{count}}天后」, gated from the diff. The zh-CN and en render at 1200/390 carries.

[lilyshen0722]

DELTA RE-GATE: PASS @ a4566e29 — sprint-review. Supersedes my PASS @ 16d6a756.

The press is still blocked, and not for a reason in the thread

5e0e7193d — the PR's first commit — is still authored sprint-impl <sprint-impl@commonly.local> at this head:

a4566e29f  Lily Shen <115414357+lilyshen0722@users.noreply.github.com>
820e9d5a8  Lily Shen <…>
16d6a7561  Lily Shen <…>
20ff9377a  Lily Shen <…>
432e93369  Lily Shen <…>
245dbaf42  Lily Shen <…>
5e0e7193d  sprint-impl <sprint-impl@commonly.local>

A squash will stamp that as a Co-authored-by on main, which is the guard that stopped the press at 820e9d5a. Two commits have landed since that was raised and neither touched it.

One consequence worth planning around: re-authoring 5e0e7193 rewrites every descendant sha, so the head that eventually presses cannot be a4566e29. Both gates will need restamping against the rewritten head — cheaper to do the re-author and the merge in one pass than to gate the same tree twice.

The delta

Exactly as described: zh-CN.json one line (tools.days {{count}} 天后 → {{count}}天后) and the new 51-line localizeRelativeTime.test.ts. No component, CSS or en-catalog byte moves, so every string lily-shen and ux-lead read at 16d6a756 is still the string here.

The clamp pin, and my own error in specifying it

I recommended windowMs: 30000 as the case to assert. That was wrong, and the correction in the PR is right:

windowMs	clamped	unclamped
1	1	0
30000	1	1
59999	1	1

Math.round(0.5) === 1, so 30000 sits exactly on the rounding boundary and returns 1m with or without the clamp. The killing range is windowMs < 30000. My suggested control could never have failed — it would have been a test that looks like a pin and pins nothing, which is the defect class I spend most of these gates hunting.

Measured on the delivered test: BASE 3/3; clamp removed → 1 failed, the sub-minute case, with both locales wrong simultaneously (0m and 0分钟); RESTORED clean. Retaining 30_000 beside minutes(1) is the right choice — it documents reachability against RoomGrant.budget.windowMs even though only the smaller value discriminates.

Siting the pin at the unit rather than through a component is also correct, and for the stated reason: the component fixture uses a one-hour window, which is precisely why this line went unguarded through two gates.

Rest

npx tsc --noEmit exits 0 with zero output. Carry: merged origin/main into the PR tree and ran the full suite there — 114 suites / 1015 tests green, which is +1 suite and +3 tests against my last gate, matching the delta exactly. Key parity: en 1561 / zh 1553, onlyZh = 0. CI: 8 checks pass, Test & Coverage and Analyze (javascript-typescript) still running.

Code gate is green on the tree. Authorship is the only thing outstanding.