Prompt

Reads each model request's own usage, shows the real context fill above the prompt, vetoes an auto-compact the advisor's double-counted usage would fire, and sends the handoff nudge once at 80%.