v0.5.2

Detector precision work (ADV-206) plus the false-negative regressions it introduced (ADV-233), released together. The FP reduction was never published on its own: v0.5.1 was tagged from 2b7d5ff, which predates it.

Fixed

• Restore detection in 26 detectors that the ADV-206 false-positive guards had silenced on genuinely vulnerable code (ADV-233)
• Guards now key on what code does, not what it is named or spelled. fn_lower.contains("validate") treated a helper named validate_pda_seeds as a signer check; body_src.contains("is_signer") treated a mere mention as a check; body_has_bound_vocab treated v.len().checked_sub(1) as a bounds check. All replaced with AST-structural analysis (ADV-233)
• Evaluate #[cfg(..)] predicates structurally instead of substring-matching their rendered tokens. #[cfg(not(test))] stringifies to cfg (not (test)), which contains "test" — production-only code was being skipped as test code. Affected INK-001, INK-004, NEAR-004, NEAR-005, CW-013 (ADV-233)
• Reduce false positives across all 61 detectors: 253 empirically-proven FPs eliminated with AST-based guards (ADV-206)
• Recognize validation delegated to a helper by resolving the callee body via the in-file call graph, not name-based skips (ADV-206)
• Match parsed tokens instead of raw source so patterns in comments and string literals no longer trigger findings (ADV-206)
• Skip #[cfg(test)] modules and test/mock helpers so non-shipped code is not flagged (ADV-206)
• Recognize const/literal and floating-point arithmetic operands to avoid spurious integer-overflow findings (ADV-206)
• test_corpus_repos cloned corpora into a path containing integration_tests, which the scanner skips as a test-path filter — suppressing ~80% of findings and failing CI for ~4 months (ADV-228)

Changed

• Add a must-still-flag regression test per fixed detector, asserting the detector still reports the vulnerability it exists for. The suite was previously ~388 should-not-flag vs ~111 must-still-flag, so a guard that silenced a detector entirely passed CI — which is how ADV-206 regressed 26 detectors with 570 green tests. Suite now 661 tests, 0 failures (ADV-233, ADV-235)
• Pin integration corpora to exact commits and tighten expected ranges; the old bounds were wide enough to stay green through a catastrophic recall loss (ADV-228)