v0.189.0

• Project reachability for Haskell. Each function now gets one of four verdicts: entrypoint-reachable, unreachable-in-project, external-api or unknown. A negative verdict needs a complete scan, a hole-free call graph, a real entry point, a known exposed-module list and a function that is not an instance method; anything less is unknown, never "unreachable". Template Haskell, foreign export, the function cap, unresolved names and withheld code are holes. The verdict is a separate record from the trust-boundary verdict.
• Measured effect: none on real code. On the 14 labelled real projects no finding changed severity and no labelled site landed in unreachable-in-project; 0 of about 41,000 functions were judged unreachable. Template Haskell blocks 11 of the 14 projects, the test-directory scan policy blocks 7, unresolved calls block 8. The mechanism is correct and inert here, so this release makes no precision claim.
• Sealed sets, measured once. Independent Haskell set: 1 true positive in 35 decided sites, 2.9% (0.5% to 14.5%), identical to 0.188.0, so by the rule fixed beforehand (under 2 points means no generalisation) this did not generalise. unseen-v4 taint recall is 66.7% (10 of 15), unchanged from 0.188.0: the case lost in 0.188.0 was not restored.
• A regression can now be seen. npm run bench:haskell-precision:check compares the development and fresh results for equality and the sealed result against a recorded baseline, and is part of the pre-push gate and the release check. A labelled true positive in unreachable-in-project fails it.
• Two labelers. The sealed-set tooling accepts a second labeler, computes percent agreement and Cohen's kappa and records an adjudication pass. Only exercised on a synthetic fixture so far; the sealed set still has one labeler.
• Smaller changes. One shared Wilson interval; originalSeverity is now recorded when the neutral dead-code pass demotes a finding; relevance ranking can consume a reachability verdict (rank only, severity untouched).
• Left out on purpose: treating external-api as "callers unknown" lowered dev precision (35.7% to 23.8%) with no true positives gained, so it is recorded but not acted on.