Real-server ACP subagent evidence

Run each revision with T3CODE_DEV_SEED=0 vp run dev in its own installed worktree. Configure the Grok provider enabled with binaryPath pointing to grok-wrapper.mjs. This wrapper only adds xai.api_key to initialize authMethods; requests and streaming replies pass through. Place driver.py and the wrapper together. Create a synthetic git repository named lighthouse-project next to driver.py. Then run python3 driver.py <checkout> base (or head). An installed, authenticated Grok CLI is required. No real T3 userdata is used.

The driver starts a real Grok turn through the running T3 CLI, samples persisted child event counters every 500ms, and saves the resulting events. analyze.py selects the child assistant stream, compares concatenated incoming ACP text to its final persisted answer, and checks final-snapshot ordering against child completion. Public data omits unrelated provider/tool events and local paths.

base/head-counters.mp4 are visual playbacks of captured server counters, not browser screen recordings. comparison.png is a chart of real event timestamps. Native model replies vary in length, as expected. No test suite or mock supplied this evidence.

Base: 9441a7285c5f3466e743d6fa1489fbefed9043f1
Head: 52f0daa5e31f5adb02a5e31e4f2d3df73cae5a37
