Changelog

Petrock follows semantic versioning. Rule-table changes that alter output for existing input are treated as breaking and only ship in a major release.

2.4.1 — 6 August 2026

• Fixed Dutch house-number suffixes being swallowed when separated by a hyphen (12-A now parses as number 12, suffix A).
• parse_many no longer leaks a worker process when the consumer stops early.
• Rule table rebuilt against the July postcode lists for CZ, SK and PL.

2.4.0 — 14 July 2026

• Added partial rule sets for Hungary, Slovenia, Croatia and Romania.
• New Address.as_dict().
• CLI gained --stdin and now writes UTF-8 with BOM on Windows only.
• About 18 % faster on batches above 100 000 rows — the tokeniser no longer recompiles its pattern per call.

2.3.2 — 29 May 2026

• Czech orientation numbers with a letter (7a) kept their letter but lost the descriptive part. Fixed.
• Confidence was reported as exactly 1.0 for single-token inputs. It now caps at 0.5 when only one field was assigned.

2.3.0 — 2 April 2026

• Dropped Python 3.8.
• Auto-detection rewritten: postcode shape first, then street-suffix vocabulary. Detection accuracy on the mixed test corpus went from 71 % to 89 %.
• strict= added to parse.