The harness always loads this directory (no env needed) and fuzzes these programs plus mutated/crossed-over descendants — concentrating inputs on the detection logic that has historically been weakest.

Two seed families, split by expected rule verdict:

regressions/ — every file is a confirmed false positive: correct, idiomatic code that a rule once wrongly flagged.
true-positives/ — every file is a confirmed true positive: a genuine bug a rule must flag, kept as a mutation seed so its shape keeps applying pressure.

Seeds with // verdict: pass or // verdict: fail are replayed deterministically by both the smoke suite and targeted fuzzing. Seeds without a verdict remain mutation-only inputs. firedProgramCount is still a coverage stat rather than a correctness oracle for generated programs.

The evolving loop (see the fuzz skill): whenever a new false positive is confirmed — from a user report, an RDE eval, a react-bench run, review, or a fuzz invariant finding — add a minimal reproducer here as regressions/<rule-id>--<slug>.tsx with a header comment naming the rule and the weakness class. The next fuzz run picks it up automatically.

Header format:

// rule: <rule-id>
// weakness: <alias-guard | copy-tracking | name-heuristic | paren-shape |
//            framework-gating | test-gating | control-flow |
//            wrapper-transparency | library-idiom | cross-file | other>
// source: <PR/issue/session reference>
// react-major: <major, only when the false positive depends on React version>

Files may use .ts, .tsx, .js, or .jsx and must parse cleanly (pnpm test enforces it).

Exact audit corpora

react-bench-0.9.7-audit/ and dummy-threejs-v14-audit/ preserve complete source files from exhaustive production-corpus reviews. Their manifests pin the source-report hashes, audited callsites, source-line hashes, expected verdict counts, and fixture mappings. Regression files use // audit-verdict: pass because the exact source can contain unrelated valid findings; they remain mutation seeds instead of claiming the entire file must be diagnostic-free. True-positive files use // verdict: fail and are replayed deterministically.

Regenerate the Dummy corpus from its archived scan artifacts with:

node scripts/import-dummy-threejs-audit-corpus.mjs \
  <before-diagnostics.tsv> <after-diagnostics.tsv> <selected-roots.txt> \
  corpus/dummy-threejs-v14-audit corpus/dummy-threejs-v14-audit.json