Real server evidence for issue 36416. Set EVIDENCE_DIR, HEAD_CHECKOUT and BASE_CHECKOUT to your own directories. Install dependencies in each checkout with vp i. run.py before and run.py after seed fresh server databases via thread.create RPC, then measure seven MCP calls at 5, 500, 2000 and 10000 threads. No provider turns. capture.py records the seven additional calls on the 10000-thread database; its required arguments are side, generated run-directory name, and synthetic project UUID printed by run.py. Videos render actual stdout as it arrives, at 10 fps; they are terminal transcript recordings, not a web UI. The initial scaling measurements and later recorded requests are separate runs. summary.json and the span extracts refer to the initial benchmark. No tokens or raw server logs are included. The span extracts contain only named listing/shell/index spans.
