Real-server PR batching proof

Create two installed worktrees at the commits in measurements.json.
Set PROOF_SCRATCH to a fresh scratch directory holding driver.py, mcp-token.mjs and pr-urls.json.
Set PROOF_MODEL_SELECTION to a JSON object with instanceId and model for an enabled local provider. The thread is idle; no agent turn is started.
Run python3 driver.py base <base-checkout> then python3 driver.py head <head-checkout>.
For repeat.py set PROOF_BASE_CHECKOUT and PROOF_HEAD_CHECKOUT, and replace its two seed-directory names with those printed by driver.py.
All application state is created by MCP commands. The cache is moved aside only while the server is stopped.
Normal runs are unrestricted; loaded runs pin the server and one busy-loop worker to the same allowed CPU.
Trace excerpts retain only sweep, syncGroup, summary API and GraphQL request spans in the first cold sweep. Arbitrary attributes and events are removed for privacy. Counts, timestamps, IDs, response status and span exits are retained.
The PNGs visualize these trace measurements, not an application screenshot. The text transcripts list every measured request.
No tests, mock API, replay provider or code instrumentation were used as evidence.
AI-assisted; verified with the tests and tools listed above.
