v0.10.0

[heavykenny]

A precision release. A sweep through the ai-slop rules to cut false positives on real human-written code, validated against popular open-source libraries surfaced at the HN launch. One library dropped from 426 findings to 92, every removed finding a false positive, with clean code still scoring 100.

Fixed (false positives)

• thin-wrapper only flags a true passthrough (a function forwarding its own parameters unchanged). A call that transforms its arguments is real work, not a wrapper.
• narrative-comment drops the bare-section-label branch, recognises reasoning markers, and the length-based branches now require an actual AI-narration signal rather than length alone. Long, well-written explanatory comments are kept.
• trivial-comment only fires when the comment restates the next line. Comments that add a condition or reason are kept.
• silent-recovery only fires when the caught error is dropped, not when it is logged.
• meta-comment only flags leading plan markers, not UI "step N" wizard prose.
• hardcoded-id / hardcoded-url exempt env-var-name literals, migration files, and stable vendor API hosts.
• todo-stub spares TODOs that reference a tracking issue.
• unused-import no longer flags from __future__ import.
• hallucinated-import parses PEP 735 [dependency-groups], recognises psycopg2 from psycopg2-binary, and no longer flags import-shaped text inside docstrings, imports under if TYPE_CHECKING:, the code / codeop / rlcompleter stdlib modules, or non-production paths (validated on flask: 19 to 0).
• docs/ build tooling is treated as non-production, like scripts/ and examples/.

Changed

• Scoring: style rules (comments, file size) weigh half so the score is driven by genuine slop rather than house style. The scoring curve is otherwise unchanged.

Full notes in CHANGELOG.md.

---

This discussion was created from the release v0.10.0.