feat(contribute): recognize Japanese course-correction prompts

[Rererr]

Summary

CORRECTION_KEYWORDS only covers Chinese and English, so a Japanese user correcting the agent right after a stop (e.g. 「違う、やり直して」) is never counted as a correction. For Japanese-speaking teams this quietly lowers the friction score (the share-learnings hint rarely fires) and under-reports corrections in digest / dashboard intervention metrics. This PR adds the common Japanese correction phrases and documents the language coverage.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature causing existing behavior to change)
• Documentation only
• Refactor / internal cleanup

Test Plan

• npx tsc --noEmit passes
• npx vitest run passes (full suite; the only failures on my macOS machine are lock-atomic.test.ts and push-*.test.ts concurrency/timing tests, which fail identically on a pristine main checkout and are unrelated to this change)
• Added/updated tests for the change

New tests in dashboard-collector.test.ts:

• a Japanese correction prompt within the 60s window counts as correction: 1
• a Japanese follow-up task (「次はテスト環境へデプロイして」) does not count

Mutation check: removing the new keyword line makes the first new test fail, so the test actually guards the change.

Related Issues

None. Found while integrating TeamAI with a Japanese team's personal retrospective loop.

Notes for Reviewers

• Matching stays a plain substring check on the lowercased prompt summary, so no logic changes — only the keyword list and two doc lines (docs/usage-guide.md, docs/usage-guide.zh-CN.md) that describe the correction signal.
• Keyword choice: I kept to phrases that are almost always a course-correction in a follow-up prompt (違う / ちがう / そうじゃな / そうではな / やり直 / やりなおし / 間違って / 間違え / 勝手に / 戻して). I deliberately left out bare 「間違」 because 「間違いない」 ("no doubt") is a common affirmative, and 「不要」-style single words that read differently in Japanese.
• Happy to trim or extend the list if you'd prefer a different balance between recall and false positives.