324554f041
A/B analysis across 13 PRs (convo/cto/threads/2026-06-qa-ab-review/02-analyst-findings.md): - DeepSeek: high false-positive rate (absint()="code exec", PLUGIN_DIR="dir traversal") - Kimi: ~60% model failure rate, 1 unique-real finding in 13 PRs - Neurometric: summarizer only, 0 unique-real findings, unreliable heuristic verdict label - Haiku: 0 failures, lowest noise, equivalent real-signal catch vs 4-voice panel Removes DeepSeek, Kimi, and Neurometric voices. Haiku is now sole reviewer. Also removes NEUROMETRIC_API_KEY env var (no longer needed). No plugin version bump -- CI-only change (precedent: prautoblogger #150). Agent-Session: engineer-panel-haiku-20260612