Summary

Three fixes found while auditing the codebase for obvious issues. lint / build / test (163 passing) all green; split into one commit per logical change.

1. emptyNoteWordThreshold setting had no effect (functional bug)

The empty-notes scanner only checked for zero-length content and never read ctx.emptyNoteWordThreshold, so the settings slider did nothing — only completely empty notes were ever flagged.

• New pure helper countWords() with CJK-aware counting: each Han / Hiragana / Katakana / Hangul character counts as one word (CJK text has no word separators), Latin/other text splits on whitespace, mixed content sums correctly. Without this the threshold is meaningless on Chinese vaults.
• Judgment changed to countWords(body) <= ctx.emptyNoteWordThreshold.
• Non-empty short notes now get a stub-specific message ("only has N words") and a wordCount evidence field.
• Added regression tests including one that toggles the threshold and asserts the setting actually changes behavior — this is the direct guard against the original bug.

2. Dead resolvedLinks fast path in orphan-attachments

The optimization cast resolvedLinks values to string and checked typeof === "string", but Obsidian stores link counts there as number, so the branch never executed. It also couldn't work as intended: resolvedLinks only tracks markdown-to-markdown references and cannot resolve attachment link targets to disk paths. Removed the dead branch; every link already falls through to resolveVaultLinkTargets, so behavior is unchanged.

3. Duplicated formatBytes helpers

render-issues.ts and markdown-export.ts each carried a near-identical local formatBytes, while utils/format.ts already exports formatSize for the same purpose. Removed both copies and routed all byte formatting through the shared helper.

Validation

• npm run lint ✅
• npm run lint:obsidian-warnings ✅
• npm run build ✅
• npm test ✅ 163/163