fix(i18n): reject translated chunks that leak reasoning or lose table rows

[diegosouzapw]

Summary

run-translation.mjs accepted whatever the model returned. The drift check only compares hashes, so broken mirrors still passed it. In the two refreshes on 2026-09-23/24 (#14676, #14724), fallback models:

• leaked their reasoning into about 90 mirrors, as <think>…</think> blocks and English meta-prose such as "I'll keep the table header…" and "My plan became clear";
• dropped table rows or code fences in about 60 mirrors. For example, es PROVIDER_REFERENCE.md lost 23 rows, and bs REMOVED_PROVIDERS.md collapsed from 67 lines to 32.

Both PRs had to be repaired by hand before merge.

Change

validateTranslatedChunk(source, output) (exported, pure):

• strips <think>…</think> blocks and a ```markdown fence that wraps the whole answer;
• rejects the chunk when the code-fence count or the table-row count differs from the source chunk;
• rejects a stray <think> tag;
• rejects English meta phrases (I'll, I need to, Okay,, My plan, …) that appear more often than in the source.

translateBody() retries a rejected chunk up to 3 times. If the chunk still fails, the doc fails and the mirror is not written. The runner already reports failed docs and retries them.

Validation

• New tests/unit/i18n-translation-output-guard.test.ts (9 cases, built from the real refresh-5/6 failures). It failed before the change (no export) and passes after.
• Existing i18n suites pass: chunking 6/6, sections 17/17, state-adopt 12/12, language-bar 10/10.
• I could not run a live check: the translation backend was down at the time (cinf 503, auto/gemini → dva 502).

⚠️ base-red inherited: #14547