fix(dashboard): match correction keywords as whole words, add team keywords

[SaulMoro]

Summary

isCorrectionPrompt tested every prompt with lower.includes(keyword). The built-in undo and redo are common Spanish and Portuguese word endings, so "dame el segundo fichero" sent within 60 s of a Stop was recorded as a course correction. One correction scores CONTRIBUTE_CORRECTION_WEIGHT (20), which equals CONTRIBUTE_SMART_THRESHOLD, so a single false match fired the /teamai-share-learnings nudge and skewed the Human Intervention panel of teamai dashboard.

Two changes:

1. Whole-word matching for spaced scripts. A keyword that contains Han, Hiragana, Katakana or Hangul keeps substring matching, because those scripts do not separate words. Every other keyword must match as a whole word. The boundary is Unicode-aware ((?<![\p{L}\p{N}_])…(?![\p{L}\p{N}_])), so "malísimo" does not match mal and the identifier test_undo does not match undo; JavaScript \b would have treated í as a boundary.
2. Team keywords. sharing.intervention.correctionKeywords in teamai.yaml is merged with the built-in list under the same matching rules. The prompt_submit hook resolves the team config and stores a correction: boolean on the event. This has to happen at capture time: ~/.teamai/dashboard/events.jsonl is machine-level and mixes sessions from every team, so rebuildSessions cannot know which team's keywords apply. A side effect is that the full prompt is checked, not only the 200-character summary. Events written before this change carry no flag and fall back to the built-in list over promptSummary, now with whole-word matching.

The built-in list still covers only Chinese, English and Japanese. A correction typed in another language counts once the team adds its words. Interrupts, tool rejections and tool errors are unchanged.

Dashboard, same seeded events, main vs this branch

Latest prompt	main	this PR
wrong, redo it	⚠ 1	⚠ 1
esto está mal, rehazlo (event flagged by the hook via a team keyword)	none	⚠ 1
dame el segundo fichero (legacy event, no flag)	⚠ 1	none

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)

Test Plan

• npx tsc --noEmit passed.

• npx vitest run passed: 226 files, 3154 tests.

• New unit tests (15): dashboard-collector.test.ts covers the seven false positives from the issue, whole-word and multi-word English keywords, underscore as a word character, Chinese and Japanese substring matching, team keywords with accented letters, full-prompt checking past the 200-char summary, and the legacy fallback in rebuildSessions. types.test.ts covers the schema and getInterventionSharing. hook-handlers.test.ts covers passing team keywords to parseHookEvent, the unreadable-config fallback, and that non-prompt hooks do not read team config.

• npm run build passed.

• Real built CLI, user scope: 72 of 72 checks passed across Claude, Codex, CodeBuddy, OpenCode and git, gitlab, github team configs. Each combination used an isolated HOME with a teamai.yaml declaring correctionKeywords: [rehazlo, "no era eso", mal], dispatched stop then prompt-submit through teamai hook-dispatch, and read the persisted event:

  Prompt	correction
  dame el segundo fichero	false
  esto está mal, rehazlo	true
  wrong, redo it	true
  no está malísimo, sigue	false
  这不对，重来	true
  run test_undo again	false

• Real built CLI, project scope: 5 of 5 checks passed. A git repo initialised with teamai init --self --agent cursor under an isolated HOME, keywords added to .teamai/teamai.yaml. Hooks dispatched with the process cwd set to ~/.cursor and the project passed only through workspace_roots (Cursor payload shape) applied the team keywords; a Claude payload with cwd set to the project did too; a prompt from an unrelated cwd did not see the project's keywords.

• Real teamai dashboard on main and on this branch with the same 21 seeded events (table above).

• git diff --check passed.

Related Issues

Fixes #564

Notes for Reviewers

• teamCorrectionKeywords resolves the team via autoDetectInit(), the same path contributeHintAllowed uses for the Stop hook. hook-dispatch-cli has already chdir'd to the hook payload's cwd by then, so this also works for Cursor, which runs hooks from ~/.cursor; the project-scope check above exercises that path.
• I did not add built-in Spanish, Portuguese, French, Italian or German keywords. Short words such as "mal" are ambiguous even with word boundaries ("no está mal" means "it's fine"), and the team list covers that need without growing a per-language table in types.ts.
• Docs: English and Chinese usage guides describe the matching rule, the built-in language coverage, and the new teamai.yaml key in the Human Intervention section and in the teamai.yaml reference. CHANGELOG.md has one Bug Fix entry for the whole-word change and one Feature entry for team keywords and the new event field.
• The real-CLI matrix exercises hook dispatch and event persistence with isolated homes, a local team directory, and an offline single-repo init. It does not exercise live provider authentication or the agent GUIs.