Sanitize empty finding descriptions

Harden reward parsing for truncated JSON

Round2 hardening, campaign fixes, training quickstart, and CI gate

Round 2 Implementation: Multi-period campaign, regulatory shocks, adversarial grading, and GRPO training infrastructure

Emergency fix: drop scoring limit to 2 decimal points and 0.01 / 0.99 to pass deep verification

Fix phase 2 deep validation grader scoring out-of-bounds error

Refactor graders.py: add explicit partial credit comments, strict duplicate handling, and weighted false negative tracking

Fix inference stdout formatting and update graders tests for strictly clamped reward constraint bounds

fix: increase score epsilon to 0.01 for validator safety

feat: v2.0 — fraud detection task, severity grading, investigation mode, security hardening, 78 tests