fix(vllm): reduce num_predict 512→350 to stay under max_model_len=2352

fix(mellea): restore vLLM probe with auth header + non-5xx check

fix: remove vLLM probe + move warmup post-planner

fix: vLLM probe with auth header + accept 401 as ready

fix: vLLM readiness probe + fast Ollama fallback timeout (app/mellea_validator.py)

fix(mellea): raise first_token_timeout to 400s to match new 360s LiteLLM timeout

fix(mellea): set _timed_out on LiteLLM exception before first token to prevent concurrent retries

fix(mellea): fallback to best_paragraph when later attempts time out

fix(mellea): two-phase timeout — 250s first token, 45s inter-token

fix(warmup): fire 1-token LLM ping at request start to warm RunPod GPU

fix(mellea): 120s cold-start timeout + drop run_reconcile fallback

fix(mellea): abort retry loop after streaming hang to prevent concurrent vLLM requests + fix closure late-binding

fix(mellea): increase num_predict 350→512 to avoid citations_dense truncation

fix(mellea): reduce num_predict to 350 for vLLM context headroom; reject empty paragraphs

fix(mellea): per-token timeout with queue-based streaming to prevent hang

deploy(l4): self-contained Riprap mirror