OpenClaw local LLM timeouts resolved

Local agent setups are currently a configuration nightmare where "cloud-first" defaults like 60-second timeouts break on local context prefill. Hardcoded 15s/60s timeouts are insufficient for local Llama 3 or Gemma 4 prefill cycles, causing silent fallbacks to cloud APIs. The "slug generator" hook creates a head-of-line blocking issue in the Ollama queue by sending requests that are abandoned but still processed. Injecting ~40,000 characters of system prompt is a "prefill tax" that local models can't pay quickly, making trimming bootstrap files essential. Parallel request processing (OLLAMA_NUM_PARALLEL) and VRAM pinning are mandatory but undocumented requirements for a stable experience.