jev-lint 0.2.1

Since 0.2.0

• Evals per rule. Every rule lives in rules/<id>/ with rule.yml and evals/ (cases, labels.json, an accepted baseline.json). jev-lint eval rules/<id> --repeat 3 scores a rule at its shipped cutoff against its baseline; --accept promotes a run; --replay re-scores every baseline with no request and refuses one whose question changed; --compare a.json b.json diffs two records. npm run ci ends in eval --replay.
• Labels out of the files. The corpus carried // DEFECT markers inside the code the model was shown; the fits measured the labels. All labels are JSON now, and every figure was re-derived without them.
• Rule improvement round. 113 → 200 labelled defects; 20 of 23 rules at precision and recall 1.00 at their shipped cutoffs, the three misses named in the rule files. fn-name-promises cutoff 0.82 → 0.55; comment-describes-block recall 0.5 → 0.92.
• Tool fixes found by the round. A test() / it() callback is a container named by its title, so a statement inside one promotes on every arm (the preamble false-positive band on real tests: median 0.57 → 0.11). A captured // comment is its whole run of lines, not the last one. The module outline is capped at 16,000 characters and nests a symbol under the one that contains it, so a 25k-line module gets a verdict. --format json is no longer truncated at 64 KiB.
• Review on a diff. --staged, --fail-on <severity>, and jev-lint init --pre-commit for a hook that reviews only what is being committed.
• Install for a coding agent. /plugin marketplace add mizchi/jev-lint, or npx skills add mizchi/jev-lint --skill jev-lint.

Full run on this repository: 1,996 subjects, 78 requests, $0.058, 8.4 s.