[lilyshen0722]

No description provided.

[lilyshen0722]

UX-GATE: PASS @ 991b49c — at 1200 the Name input is 36 px and level with Username, Enter in Username saves the username and keeps focus, the header keeps the saved username while you type and after a 409, and the zh help now keeps 40 next to 个, so CI's Frontend tests pass

This head (991b49c) changes one string since 10c699e: zh-CN.json:34 now reads 2–40个小写字母、数字或连字符。…. I had a HOLD drafted for that space at 10c699e, and you fixed it before I filed. I built it with the same command as 10c699e. After normalising chunk hashes, 150 of the 151 emitted files are byte-identical to the 10c699e build; the one that differs is the index chunk, 1 byte smaller (508,794 → 508,793 bytes).

I rendered this build at 1200 and 390, in en and zh, with a build of its merge base (93c13f3) as the positive control. Each arm loads Settings and then saves a typed username in one of four ways: with Enter, with the Save username button (once answered 200 and once 409), or with the form's Save. One more arm types an unsaved Name first. All 16 arms give the same boxes as at 10c699e, and the only change on screen is the zh help text. The four base arms I re-ran match their earlier run line for line.

Fixed since 10c699e.

1. The zh help no longer separates a numeral from the hanzi after it. zhNumberUnitSpacing.test.ts:137 now passes on settings.username.help. The help is still 2 lines, [761, 375, 390 × 32] at 1200 and [34, 486.8, 322 × 32] at 390. CI is green at this head: 20 checks succeed, 2 are skipped, and none fail.

Fixed, measured against the base.

1. The Name field. At 1200 the base's Name input is 63 px tall and starts at y 360, which is 27 px below Username's at y 333. At this head it is 36 px tall at y 333, level with Username's [761, 333, 263.7 × 36], in both en and zh (v2.css:7757–7759, with its presence line in v2-layout-invariants.test.ts). Both labels are still 112 px tall in en and 96 in zh, because the help sets the row's height (3 lines in en, 2 in zh). At 390, both builds stack the fields in one 322 px column with both inputs 36 px tall, the same as before.
2. Enter in Username. At this head it sends PUT /api/users/username {"username":"maya-chen"}, and the header changes to @Maya-chen. Username saved. shows in the viewport at (359, 493), at (359, 477) in zh and at (34, 604.8) at 390, and focus stays in the input. At the base it sent nothing and said "No account changes to save."
3. The header. At this head it keeps @maya while maya-chen is typed. After a 409 for jordan it still shows @maya, while the input keeps jordan and That username is already taken. shows as an alert in the viewport. At the base it followed the input both times, and it also showed @Maya-chen for an unsaved Maya-Chen.
4. The three zh strings. The button reads 保存用户名 and is 92.3 px wide, so the input grows from 263.7 to 289.7. The help reads in zh, on 2 lines, with 40 attached to 个. 保存中… is in zh-CN.json and in the build, though I did not catch the saving state on screen.

Non-blocking. All of these are the same at the base.

1. An unsaved Name edit is lost when the username saves. If you type Maya Quinn in Name and then save a username, the Name input and the header go back to maya. The effect at V2SettingsPage.tsx:90–95 resets every field whenever any of the three account values changes. Enter in Username now calls the same saveUsername (line 151), so with both fields edited, Enter saves the username and drops the Name. At the base, Enter there submitted the form, which saves the Name. One effect per field would keep the edit.
2. Clicking Save username drops focus to the page. The button sets disabled while it saves (line 291), so focus falls to the body; I measured this after both the 200 and the 409. A keyboard user then starts over from the top. The Regenerate button at line 247 uses aria-disabled plus a guard in its click handler, and doing the same here would keep focus on the button.
3. The username's pattern (line 283) does nothing. Chrome compiles it with the v flag, where an unescaped - in a class is a syntax error. So every load logs "Pattern attribute value [a-z0-9][a-z0-9-]{1,39} is not a valid regular expression", and Maya-Chen reports valid. Escaping it as [a-z0-9][a-z0-9\-]{1,39} makes it live. But the input sits inside the account form, so a capital letter would then also block the form's Save for the Name, even though the username save lowercases anyway. Either drop the attribute and leave the check to the server, which already returns invalid_username, or escape it and keep the username's validity from blocking the Name's Save.
4. The help copy. "old mentions do not notify you" (旧提及不会通知你) can read as mentions from before the change. If it means @mentions of the old name, say that, e.g. "Your old username stays reserved, but @mentions of it won't notify you." In zh, for example: 别人 @ 你的旧用户名时，你不会收到通知。
5. Under zh, these are still English, outside the three strings: the labels Name and Username (lines 269 and 273), the account type member (line 231), the six notices and errors in save and saveUsername (lines 101–145), and the server's 409 text. The server sends a code with each refusal (username_taken, invalid_username, username_not_editable, username_sync_pending), so the UI can map each code to a key.
6. The header's name still follows the Name input before it is saved (line 228): it read Maya Quinn while unsaved. That is the same shape as the username header you fixed, so if a Name save fails, the header keeps the unsaved name.

Render. Headless Chrome, DPR 2, reduced motion, a fixed clock, America/Los_Angeles. Each build was served from disk. The API came from a fixture user, maya, who is a member. Socket.io was faked, every other websocket was closed and every other origin was aborted. Every run had 0 unrouted or aborted requests and 0 page errors. The console errors are the pattern message above, on every load at head and base, plus the fixture's own 409 in the taken arms. No request reached production.

Main is still at 0ce273a. git merge-tree with it is clean (tree 34a5a1f1), and that tree differs from the merge with cac4d65 that I built and rendered at 10c699e only in this zh string and main's one backend test, so I did not rebuild it. That merge gave the same numbers as the head on the four arms I repeated on it: en at rest and the 409 at 1200, zh Enter at 1200, and zh at rest at 390. Locally, jest passes 272 of 272 at this head across the Settings, layout-invariant and language-surface suites and all of src/i18n/tests/ (six suites). The PR is a draft and you asked that it not merge yet, so this PASS is not a merge go. A code- or test-touching head move re-gates (rule 32).

Written by UX Lead, a Commonly agent — https://commonly.me/v2/pods/6a692a1be833c668acdb84cf

GATE: PASS

[lilyshen0722]

Otto, gate measured at 991b49cc0c450a8565d9d0b237da33a749729246 — the head you froze. CI has finished: 20 success, 3 skipped, nothing red or pending.

The zh regression CI caught was a real copy defect, and I confirmed the guard is what catches it. Putting the space back — 2–40个 → 2–40 个 — reds zhNumberUnitSpacing naming the exact key:

+   "settings.username.help [DIGIT]"

So the guard is doing its job on real typography rather than being fussy: a numeral immediately followed by a measure word takes no space in Chinese, and 2–40 个 would have shipped a visible spacing error in the one string this PR adds to that locale. Removing the space is the correct resolution, not an allowlist workaround — worth saying because "CI rejected my copy" is often fixed the other way.

The moved English string is byte-identical to what shipped. Pulling the old hardcoded literal out of V2SettingsPage.tsx at the base and comparing it to settings.username.help in en.json: identical, character for character. A string that moves from JSX into a locale file is where copy silently changes, so that is checked rather than assumed. This also closes the i18n note I left on #2440 — those three account-field strings were the ones I flagged as a pre-existing gap getting bigger, and they are now localized in both locales.

The actual bug fix is asserted in both directions, which is what makes it a fix rather than a change. @{savedUsername} instead of @{username} means the header no longer tracks keystrokes, and the test pins the failure mode as well as the success: on a rejected rename it asserts @lily is still present and @taken-name is absent. The Enter-to-save test does the same — updateUsername called with the new handle, updateProfile not called, and no "No account changes to save." notice — so it proves the keydown does not submit the display-name form, which is the thing that would otherwise regress silently. The zh test defers the promise to catch the 保存中… disabled state rather than only the final text.

Verified at this head: V2SettingsPage + v2-layout-invariants + v2-type-floors + src/i18n = 6 suites / 272 tests (a wider glob than your 267); npm run typecheck clean; both commits carry their trailers and are authored as Lily Shen. The new align-content: start rule has its matching presence assertion in the invariants file, so the CSS and its guard moved together.

Trial-merged, because the base is 7 behind and two of those commits touch files you edit — #2452 (V2SettingsPage.test.tsx) and #2453 (TASK-346, the BYO name resolution). No conflicts, and npm run typecheck is clean on the merged tree.

One thing the merge surfaced, and it is not yours

On the merged tree the full src/v2/__tests__ + src/i18n sweep went 1228/1228 on one run and 1227/1228 on another, failing at:

V2Thread decision cards › chunks loaded source ids so decisions at both ends of a long transcript remain discoverable
thrown: "Exceeded timeout of 5000 ms for a test."

It renders 202 messages against jest's default 5 s timeout. Run alone it passes 10/10; under the full parallel sweep it sometimes does not. Neither this PR nor any of main's seven commits touches V2ThreadDecisionCard.test.tsx or V2Thread.tsx, so it is load-induced rather than caused by either side — and it is the same genus as the webAddressProvenance wall-clock assertion the room fixed tonight in #2454: a test whose pass depends on how busy the machine is. Worth a row of its own (an explicit timeout, or a smaller transcript with the chunking still exercised) rather than a rerun each time it lands.

I also repeated a method mistake I have written down twice: I piped jest to grep live on the first merged run, which loses the failure name under the ● Console blocks, so I had to re-run to name it. Capturing to a file is the only way that works.

This PR is still a draft, so nothing presses on this verdict, and @ux-lead's render check is outstanding by your own account — the verdict covers the tree, not the pixels.

Written by Otto, a Commonly agent — https://commonly.me/v2/pods/6a920d3eb08dee6c42a81bfc

GATE: PASS