memory-quality-gate 0.2.1

memory-quality-gate 0.2.1

Faster novelty checks, more accurate scoring, and support for Python 3.11–3.14.

Performance

• About 5.5x more candidates per second when one gate scores many candidates against an existing corpus. An inverted token index and a compact word-trigram filter are built on the second novelty lookup, so one-shot CLI runs never pay for them. Scores are identical either way.
• Scoring stops counting regex matches once a dimension reaches its maximum score, so very long candidates are about 30% faster.
• About 9% less startup CPU for the CLI, which no longer loads pathlib, or json for text output.
• New scripts/bench.py measures import time, CLI wall time and throughput on a generated corpus that is the same on every run.

Fixed

• Percentages such as 50% now count toward specificity.
• File paths now count toward specificity when they end a sentence, sit inside backticks or quotes, or are listed back to back.
• reason: and why: labels now count toward reasoning.
• A candidate whose score lands exactly on its threshold now passes. JSON scores print as clean decimals.
• The specificity shortcut never raises a custom threshold.
• Duplicate detection now works for non-ASCII text, such as Japanese or accented French.
• CLI: --text "" is scored instead of falling back to stdin. A closed stdin now gives a usage error.
• scripts/setup.sh accepts Python 3.11 and newer, matching requires-python.

Changed

• CI runs on Python 3.11, 3.12, 3.13 and 3.14.
• The README example output and the scoring-model document now match actual behaviour.

Full Changelog: v0.2.0...v0.2.1