Recall: keyword fallback should support short Unicode/CJK tokens

[GoZumie]

Summary

Keyword fallback currently drops tokens shorter than 3 characters:

.filter(|s| s.chars().count() >= 3)

This can suppress valid short Unicode terms (especially common CJK 1–2 character words), making fallback a no-op for multilingual queries where short tokens are meaningful.

Why this matters

The fallback path is explicitly intended to improve multilingual/short-phrase recall in text-only or weak-semantic scenarios. The fixed length threshold is language-biased and can miss relevant matches.

Proposed direction

• Replace fixed >=3 rule with language-aware or script-aware token policy.
• At minimum, allow short non-ASCII tokens (or CJK script classes) while still filtering noisy short Latin stopwords.
• Add tests for short CJK-token fallback behavior.

Acceptance criteria

• Query terms using short CJK tokens can produce fallback terms and recover related matches.
• Existing noise control for short Latin stopwords remains acceptable.
• Regression test(s) added for multilingual short-token fallback.

Context

Found during review of PR #49.