v1.0.0 — Record linkage that shows its arithmetic

Finds the same person written down in different ways — a typo in the surname, another phone format, a swapped day and month, Latin letters from the system that was replaced two years ago.

• Blocking instead of brute force — several coarse keys per record turn 2.1M possible pairs into 5 934 candidates on the demo base without losing a single true duplicate; the seeder measures recall and prints it
• Metrics written by hand — Jaro-Winkler with a prefix bonus, bounded Levenshtein, Jaccard over address tokens, and a date comparison that knows the two most common entry mistakes
• Transliteration built in — «Соловьёва» meets «Solovieva», «Королёв» meets «Korolev»
• Weights normalised over comparable fields, with every field's contribution stored beside the pair
• Two thresholds with a live preview — move the sliders and see what would be merged automatically, what would go to a human, and how many already-decided pairs the current settings would send the wrong way
• A human decision outranks the machine — «do not merge» keeps a pair apart at any score, and rescoring never deletes decisions
• Clusters with a weakest link, so chains where A resembles B, B resembles C and A has nothing to do with C surface first
• 35 tests on PostgreSQL, light and dark themes, README in three languages