Branch task/issue-1844 · Commits c0f874cad (gate + backfill), 11f0a9718 (drop the CI workflow file — the bot token cannot push it) · Follow-up #1850 · Refs #1840, #1845

The client-side translator falls back to the key itself:

// gui/templates/i18n/i18n.js:175
var str = _strings[key] || key;

apply() then overwrites textContent unconditionally, so a bundle that is silent on a key paints the key on screen — no exception, no console error, no failed build, no Event-Viewer row. Nothing in the toolchain noticed, because the only i18n check in the repository was

python3 -m json.tool <file>      # ai/AGENTS.md rule 3, ai/IMPLEMENT-TASK.md §3,
                                  # ai/FIX-ISSUE.md §5, ai/ISSUES.md §4,
                                  # ai/MERGE-STALE-BRANCHES.md

which proves well-formedness, never coverage. {"role.noRights": null} is valid JSON.

40 unique keys, in five groups:

Keys a bundle has but en.json does not: 0 in all nine — drift is strictly one-directional, which is what makes en.json the right reference for a gate.

The class is not carelessness. 4bcf10e07 shows it is what happens when a revert is re-landed bundle by bundle: each bundle is valid, each diff is reviewed, and nobody compares the key sets.

Written textually before the closing brace, never re-serialised: the bundles are append-ordered, not alphabetically sorted, so json.dumps(sort_keys=True) would have moved ~100 lines per file and buried 184 real lines in reformatting noise. Result per file: +26/-1, +20/-1, +6/-1, +40/-1 — additions only, the single -1 being the closing brace gaining a comma.

• 8 strings reused verbatim from each bundle's own sibling. rmo.grid.* are the Requirements-Monitor-Overview twins of the already-translated ro.grid.* (identical English in en.json), and rmo.resetFilters is the twin of tspec.resetFilters. Reusing the sibling is the only way the two screens keep saying the same thing — a fresh hand translation would have drifted from ro.grid.* inside one UI.
• ~80 hand-written translations for rctc.*, role.*, ts.chapterBug*, rmo.multiSort, rmo.resetToDefaultState and tspec.* (zh), each written in the bundle's existing register and reusing its established terms.
• Placeholders preserved: {count} in rmo.grid.groupItem/groupItems, verified across all nine bundles.
• fr.json — nothing to do; #1840 had already fixed it.

bash ai/verify_i18n_coverage.sh              # exit 1 if a bundle misses an en.json key
bash ai/verify_i18n_coverage.sh --report     # WARN only, triage
bash ai/verify_i18n_coverage.sh --max-print N <dir>

• Flattens nested bundles before comparing — TLi18n looks up flat keys, so a bundle that nests a.b.c as an object must compare equal to a flat en.json.
• Extra keys are informational, not failures — a bundle may legitimately carry a key en.json has dropped; that must not wedge the gate.
• Blocking, not warning. A --report-only gate would have printed all five historical commits and exited 0: the same silent outcome, one step removed.
• Refuses to report success when it cannot compare — missing or empty en.json, invalid JSON in any bundle: all exit 1 with a readable message.
• Read-only: it never rewrites a bundle.

Wired into ai/AGENTS.md rule 3, ai/IMPLEMENT-TASK.md §3 and ai/FIX-ISSUE.md §5 — all three previously stopped at json.tool, which is exactly how the drift shipped.

.github/workflows/i18n-coverage.yml was written and validated (yaml.safe_load → jobs: ['coverage']), but the agents' push token is a GitHub App without workflows permission and the push was rejected. The YAML is preserved verbatim in #1850 for a token that can commit it. Until then the rulebooks are the enforcement point, and they cover every agent run.

$ python3 -m json.tool gui/templates/i18n/*.json     -> 10/10 valid
$ bash ai/verify_i18n_coverage.sh ; echo $?
PASS de.json ... PASS zh.json — 6781 keys, 0 missing
9 bundle(s) passed, 0 failed
0

9-case gate self-test: exit 1 on a single injected missing key (the exact shape of 9d380b9dc), on invalid JSON, on an empty reference, on a missing reference, on a bad --max-print; exit 0 with --report on the same tree; nested layout compares equal; the real tree stays green.

Browser: the pre-fix ja.json (HEAD~2) and the fixed one rendered through the real lookup side by side — 7 of 7 keys were the raw key before, 7 of 7 resolve after:

And in the live UI (admin/admin, locale forced with ?locale=ja):

グループの展開/折りたたむ · 全ての列を表示 · フィルターをリセット · デフォルト状態に戻す · 複数ソート — five strings that used to be raw keys on that toolbar. Role Management, Create Test Cases from Requirements and Test Specification were checked the same way in zh and ja: zero raw keys on screen, and a programmatic sweep of 368 key × locale lookups across the 40 keys × 9 bundles returned zero fallbacks.

Test suite: Task — Issue #1844 in tmp/TLU_Test_Cases.md, 15 cases, 15/15 PASS. ai/verify_test_suites.sh (G1805): 7 PASS / 0 FAIL / 0 SKIP — 12 suites at the merge-base → 13, none lost.