fix(dashboard): normalize Unicode correction keywords

[SaulMoro]

Summary

A team keyword such as réessaye matched an NFC prompt but missed the same word written with a combining accent. This undercounted course corrections. The matcher now normalizes the prompt and keywords to NFC before case-insensitive matching. It preserves the original prompt summary, existing word boundaries and the 60-second window.

The English and Chinese usage guides and changelog describe the fix.

Type of change

• Bug fix

Dashboard comparison

The real dashboards use the same three prompt and keyword scenarios, with events captured by each build's hooks.

Keyword / prompt form	Main	This PR
NFC / NFC	1 correction	1 correction
NFC / NFD	0 corrections	1 correction
NFD / NFC	0 corrections	1 correction

Before:

After:

Test plan

• Added 13 regression tests for both normalization directions, uppercase input, long prompts, accent distinctions, word boundaries, the correction window and legacy events. Six of the initial regression cases failed before the fix.
• npx tsc --noEmit passed.
• npx vitest run passed with 226 files and 3,167 tests.
• npm run build passed.
• Real built CLI passed 180/180 scenarios. The matrix runs 15 scenarios across Claude, Codex, CodeBuddy and OpenCode with team configs declaring git, gitlab and github. Each scenario dispatches Stop and prompt-submit hooks, checks the persisted correction flag and original summary, then checks aggregate counts through the dashboard API. The same matrix against main had 120 passes and 60 failures.
• Real built CLI passed 5/5 project-scope cases after teamai init --self --agent cursor --scope project. These cover both normalization directions through Cursor workspace_roots and Claude cwd, plus isolation from an unrelated directory.
• The real dashboard shows the correction counts in the table above. No browser console or page errors were reported.
• git diff --check passed.

Related issues

Fixes #573.

Follow-up to #567, which addressed the original issue #564.

Notes for reviewers

Normalization runs only during matching. Existing persisted correction flags are not recalculated. Legacy events without a flag still use the built-in keyword fallback, which now also normalizes Unicode.

The CLI matrix uses isolated local homes and local team configurations. It does not exercise live provider authentication or agent GUIs. Validation ran on macOS with Node 22.22.2.