[argszero]

Summary

ERR_MAP in ui/js/i18n.js maps the backend's Chinese error strings to English translation keys, and mapErr() returned on the first entry whose literal was a substring of the message. The table mixes generic and specific literals, so two specific entries were permanently unreachable:

literal	key	shadowed by
验证码不存在或已过期	err.codeExpired	不存在 → err.notFound
验证码错误次数过多	err.codeLocked	验证码错误 → err.badCode

In English mode this made the verification-code flows show the wrong text, even though both correct translations already existed in the pack:

• 验证码不存在或已过期，请重新获取 → "Resource not found" (should be "Code invalid or expired, please request a new one")
• 验证码错误次数过多，请重新获取 → "Incorrect verification code" (should be "Too many wrong attempts, please request a new code")

The backend messages live at src/routes/mod.rs:284, :291 (reset-password) and :394, :401 (email verification); mapErr is applied to every user-visible error message in app.js, so the wrong text is what users actually see.

An all-pairs sweep of all 22 entries found exactly two shadowing pairs, both with different keys — only these two mappings were affected. No other message changes.

Related Issue

Changes

• ui/js/i18n.js — mapErr() now selects the longest matching literal instead of the first one, which makes resolution independent of table order. The table is additionally reordered longest-first so the most specific entries read first and newly inserted entries cannot reintroduce the same problem. The mapping set is unchanged (22 entries in, 22 out; none lost, added, or duplicated).
• ui/index.html — cache-bust ?v=20260911-8 → -9 (5 references).

No new keys and no new vocabulary: both err.codeExpired and err.codeLocked already existed.

Tests

• cargo test — 148 passed; 0 failed
• cargo fmt --check — clean
• cargo clippy --all-targets -- -D warnings — clean
• node --check on all four ui/js/*.js — clean
• New unit tests added (n/a — see note below)

Verification of the actual behaviour (the repo has no JS test harness — its convention is Rust #[cfg(test)], and CI runs fmt/clippy/test only, so I did not add a new JS test runner as that would exceed the scope of this fix):

• Replayed the real backend messages through the real mapErr in EN mode — both now resolve to the specific entry, and previously-correct mappings (点数余额不足, 该邮箱已注册, the generic 不存在 fallback) are unchanged.
• Order-independence proof: 200 random shuffles of the table × 19 real messages → 0 mismatches. Running the same shuffling against the previous first-match algorithm yields 288 mismatches, confirming the test is sensitive to order and that resolution no longer depends on it.
• Controls: zh-mode passthrough ✅, unmatched-message passthrough ✅, null input passthrough ✅.
• Set-equality check: 22 entries before → 22 after, no loss/duplication.

Checklist

• Branch naming follows the convention (fix/<description>)
• Commit message uses Conventional Commits (fix(ui): …)
• Single responsibility, minimal diff (2 files, +31/−21)

Follow-up (not in this PR)

A separate, larger gap is measurable but deliberately left out to keep this change reviewable: of the 40 distinct Chinese error strings the backend can return, 18 are covered by ERR_MAP and 22 are not, and the uncovered ones render raw Chinese in English mode (e.g. 该邮箱未注册 on the login path, src/routes/mod.rs:388). Happy to follow up with that coverage expansion if wanted.