Accelerated-sim agents: relative time windows break when the LLM is slow
When an LLM agent drives a simulation running faster than real time (e.g. 60x), tools that take relative windows ("detections from the last 60 seconds") silently return nothing: a local 8B model (qwen3:8b via Ollama) spent ~70 s wall time thinking, which is over an hour of sim time, so its window missed every event and it dismissed real incidents. Fixes: anchor tool windows to the incident (since incident start, or explicit absolute sim timestamps) rather than 'now', and/or hold the sim clock while the agent works. If you hold the clock, don't let a blocking tool wait on sim progress (e.g. a drone flight that completes in sim time), or the agent and the clock deadlock; make such tools advance the sim themselves or return immediately with status. Even with the clock held, the 8B model still judged poorly (called a corroborated event 'single node' and invented node names), so pass corroborating evidence into the prompt in code and constrain names to an enum.