v1.2.3

Bug Fixes

• Fix OOM/segfault on wide datasets: Auto-configure now caps columns at 40, fuzzy fields at 5, and skips exact matchkeys on >10K rows to prevent O(N²) self-joins
• Fix _GEO_PATTERNS regex: city no longer matches inside "Capacity" — uses negative lookbehind instead of word boundaries to preserve city_desc/county_code matching
• Fix golden record schema mismatch: pl.DataFrame(golden_rows) now uses schema_overrides to prevent mixed-type inference across cluster rows
• Fix BlockingConfig.keys required when unnecessary: keys now defaults to [] with model_validator enforcing per-strategy requirements
• Cast to Utf8 in run_dedupe_df: Prevents schema mismatch from mixed-type columns (e.g. birth_year as i64 vs str)
• Guard auto_select against empty keys: Prevents crash when blocking config has no keys

Testing

• Verified on UK GIAS schools register (52,288 rows, 135 columns)
• Zero-config dedupe_df() completes in 175s with 3,854 clusters
• 1,181 tests passing