v3.1.29

• deep_scan now runs the LLM at temperature 0 — same code yields stable findings instead of varying run-to-run
• deep_scan prompt enforces precision over recall: only flags vulnerabilities present in the shown code, no speculation about unseen helpers, no generic hardening noise (clean code that drew 3 speculative findings now returns 0)
• Real semantic vulns (IDOR, business-logic, race conditions, auth bypass) still caught; validated live against the API