PriyanshuHF/dead-internet-detective-trainer-p3

debug: add /infer/debug endpoint with full traceback

fix: SKIP_TRAINING flag — inference-only mode, no retraining on rebuild

fix: load LoRA adapter from HF Hub (not ephemeral local path)

feat: add /infer and /infer/batch endpoints for trained model testing

speed: ROLLOUTS=2, MAX_EP_STEPS=6, MAX_EP_SECS=15, STEPS=25 → ~25min

phase3: fix max_new_tokens=128, MAX_EP_STEPS=12, 40 train steps

initial commit