Obfuscation-Resistant and Calibrated Extraction of Indicators of Compromise: evaluation code and cached predictions

Pipeline and evaluation harness for the paper Obfuscation-Resistant and Calibrated Extraction of Indicators of Compromise. Includes the rule-based deobfuscation layer, confidence fusion and post-hoc isotonic calibration, the split-conformal risk-control procedure, the adversarial study, and the cached LLM predictions needed to reproduce every headline number offline via scripts/reproduce_all.py. The PRISM gold benchmark itself is not redistributed here; see README.md for acquisition instructions.