fix(holographic): add LIKE fallback for CJK full-text search

[maurovideosmr-wq]

FTS5s default unicode61 tokenizer does not segment CJK text, so Chinese characters without whitespace or punctuation boundaries are indexed as single tokens. For example, "香港腾讯云" is indexed as one token, and searching for "香港" returns no results because FTS5 does exact token matching.

This fix:

1. Catches FTS5 MATCH exceptions (e.g. hyphens in queries like "sing-box" are parsed as column references by FTS5).
2. Always runs a LIKE "%query%" fallback alongside FTS5 MATCH.
3. Merges results from both sources, deduplicated by fact_id — FTS5 results keep their ranking, LIKE additions get a neutral rank.

The LIKE scan is O(n) but the fact store is typically small (<1000 entries), so the performance impact is negligible.

Testing:

• Searching "香港" now returns all 3 matching facts (was: 0 or partial)
• Searching "sing-box" finds facts with hyphens (was: FTS5 crash)
• Searching "43.132" matches numeric IP fragments
• Existing FTS5-exact searches (English/ASCII) remain unaffected

[teknium1]

Thanks for isolating the CJK recall gap. The current schema still uses a default FTS5 table (plugins/memory/holographic/store.py:48-49), and current query preparation only splits on whitespace before emitting quoted FTS phrases (plugins/memory/holographic/retrieval.py:605-619), so the fallback idea remains relevant.

Problems

• The proposed LIKE "%{query}%" uses raw % and _ from user input as SQL wildcards. Escape literal LIKE metacharacters and add an ESCAPE clause so fallback results remain a literal substring search.
• The fallback only scans f.content, although FTS indexes both content and tags (plugins/memory/holographic/store.py:48-49). Include tags so the fallback preserves the searchable surface.
• The patch predates current main's sanitizer/OR recall path at plugins/memory/holographic/retrieval.py:499-503 and 585-619, introduced in cb6d6d46ab6b20b173c8215a1f066b53847e9ee5.

Suggested changes

• Rework the fallback around the current sanitizer, retain its existing behavior, and add integration coverage for CJK content/tags plus literal % and _ queries.

This is an automated hermes-sweeper review.

[maurovideosmr-wq]

Updated per @teknium1's review:

1. LIKE metacharacters escaped — added _escape_like_pattern() that escapes %, _, and \ in user input, with ESCAPE '\' clause in the SQL. Literal % and _ in queries are now treated as characters, not wildcards.

2. Tags included — LIKE fallback scans both content and tags (f.content LIKE ? OR f.tags LIKE ?), matching the FTS5 index surface.

3. Built on current sanitizer — the FTS5 path still uses _sanitize_fts_query() (cb6d6d4) unchanged. LIKE only adds recall for sub-string matches the tokenizer cannot segment (CJK, numeric fragments). No duplicate logic.

4. Integration tests added — 7 new tests covering:

   • CJK sub-string in content ("香港" → matches "香港腾讯云...")
   • CJK sub-string in tags ("腾讯" → matches via tags column)
   • FTS5 + LIKE merge dedup (no duplicate fact_id)
   • Literal % escaped (searching "95%" does not wildcard-match everything)
   • Literal _ escaped ("DATABASE_URL" matches literally)
   • Numeric IP fragments ("43.132")
   • Existing English search unaffected

All 17 tests pass (10 existing + 7 new, 0.48s).