fix: suppress sigh interjection annotations

backlog/tasks/task-319 - Suppress-annotations-for-expressive-interjection-subtitles.md

@@ -0,0 +1,58 @@
+---
+id: TASK-319
+title: Suppress annotations for expressive interjection subtitles
+status: Done
+assignee:
+  - Codex
+created_date: '2026-05-03 03:18'
+updated_date: '2026-05-03 03:20'
+labels:
+  - bug
+  - subtitle-annotations
+dependencies: []
+references:
+  - src/core/services/tokenizer/subtitle-annotation-filter.ts
+  - src/core/services/tokenizer/annotation-stage.test.ts
+priority: medium
+---
+
+## Description
+
+<!-- SECTION:DESCRIPTION:BEGIN -->
+Interjection-only subtitle tokens such as ハァ and はっ should remain hoverable as tokens but must not receive known, N+1, frequency, or JLPT annotation styling. Current behavior can still annotate these forms when dictionary/POS metadata does not trip the existing exclusion gate.
+<!-- SECTION:DESCRIPTION:END -->
+
+## Acceptance Criteria
+<!-- AC:BEGIN -->
+- [x] #1 Standalone ハァ/はっ-style interjection tokens have annotation metadata cleared even when dictionary metadata exists.
+- [x] #2 Filtering remains scoped so content-bearing non-interjection tokens still receive annotations.
+- [x] #3 Regression coverage exercises the reported subtitle pattern: ハァ… / （ガーフィール）はっ！
+<!-- AC:END -->
+
+## Implementation Plan
+
+<!-- SECTION:PLAN:BEGIN -->
+1. Add failing regression coverage around annotation filtering for the reported interjection forms, including katakana ハァ and small-tsu はっ with surrounding subtitle punctuation/name text.
+2. Tighten the shared subtitle annotation exclusion gate so expressive kana interjections clear annotation metadata without relying only on MeCab pos1=感動詞.
+3. Run the focused tokenizer/annotation tests, then update acceptance criteria and notes.
+<!-- SECTION:PLAN:END -->
+
+## Implementation Notes
+
+<!-- SECTION:NOTES:BEGIN -->
+Implemented via shared subtitle annotation exclusion term normalization: added はぁ so katakana ハァ normalizes into the existing term gate. Existing small-tsu kana SFX logic already covers はっ. Regression confirms both reported forms clear known/N+1/frequency/JLPT metadata while a normal noun keeps frequency annotation.
+<!-- SECTION:NOTES:END -->
+
+## Final Summary
+
+<!-- SECTION:FINAL_SUMMARY:BEGIN -->
+Summary:
+- Added a regression for the reported subtitle pattern ハァ… / （ガーフィール）はっ！, with annotation metadata present on both interjection tokens.
+- Extended the shared subtitle annotation exclusion term set so ハァ normalizes to はぁ and is stripped of annotation styling. Existing はっ handling remains covered by small-tsu kana SFX filtering.
+- Added a change fragment for the user-visible bug fix.
+
+Verification:
+- bun test src/core/services/tokenizer/annotation-stage.test.ts
+- bun test src/core/services/tokenizer/annotation-stage.test.ts src/core/services/tokenizer.test.ts src/renderer/subtitle-render.test.ts
+- bun run typecheck
+<!-- SECTION:FINAL_SUMMARY:END -->