fix(vllm): only pass documents via chat_template_kwargs, not top-level

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

debug: improve vllm-direct endpoint to test context overflow

debug: add /api/debug/vllm-direct endpoint

fix(vllm): strip document-role messages before sending to vLLM

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

fix: remove vLLM probe + move warmup post-planner

fix: remove vLLM probe + move warmup post-planner

fix: conditional Ollama timeout — 5s when vLLM primary, 240s otherwise

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

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

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

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

fix(llm): raise vLLM first-token timeout to 360s for RunPod cold start

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(warmup): fire vLLM warmup before planner so RunPod loads during planner+stones

fix(warmup): fire vLLM warmup before planner so RunPod loads during planner+stones

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

fix(sse): send keepalive comments to prevent proxy idle timeout

fix(fsm): inject valid doc_ids into system prompt to prevent rag_npcc4 hallucination

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

fix(reconcile): correct rag_npcc4→npcc4_slr in system prompt example citation

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

fix(fsm): fallback uses non-streaming reconcile to avoid double hang

fix(fsm): fallback to non-strict reconcile when Mellea returns empty paragraph

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

fix(cornerstone): replace ThreadPoolExecutor with sequential loop — fixes Burr post-action cleanup hang

fix(burr): remove SQLitePersister cache — was poisoning state on first broken run

sync: upload app/reconcile.py from local main

feat(burr): LocalTrackingClient, SQLitePersister cache, StepEventHook, conditional transitions

deploy(l4): self-contained Riprap mirror