feat: local judge server (serve_judge.py) — replaces NVIDIA NIM entirely

fix: is_terminal bug in curriculum_basic + wire local judge into rollout_test

feat(training): local judge + parallel rollout + optimized batch config

fix(rewards): LLM judge load balancing, premature query, keyword stuffing, SLA decay

fix: rate limit bypass for admin key + JSON 429 + LLM judge killswitch

fix: natural conversation flow — greet before querying DB

chore(round2): full audit pass — harden rewards, patch frontend, tighten training

feat(round2): DB-grounded multi_domain task with 30-ticket curriculum stage

docs: reorganize .md files into docs/ and rewrite README for Round 2

fix(security): close all 7 audit findings from V2/V3 security audit

feat(env): double ticket pool size for better training generalization

feat: hackathon improvements — benchmark dashboard, GRPO training pipeline, grader fixes

feat: Progressive 4-stage curriculum (v2.1.0)

feat: Round 2 — Hierarchical multi-agent environment (v2.0)

fix: add nightmare grader, regenerate uv.lock without sentence-transformers

fix: remove sentence_transformers, restore TF-IDF scoring, fix nightmare tickets

feat(hackathon): Win plan implementation - grader exploit fixes and prompt tuning

fix: Phase 17 final production lock (auth, PII redact, structlog)

feat: complete hackathon hardening and nightmare mode integration

fix: production hardening — rate limiting, session cap, body limit, logging, deep copy, e2e tests

fix: address 6 audit report issues — input validation, CORS, session TTL, health check, log safety, history truncation

feat: complete CustomerSupportEnv RL environment for OpenEnv hackathon