Real-server proof for issue #33300 Environment: Windows Server 2025 x64; Node 26.13.1; T3 server 0.1.10. Base: 5ec4f500ef36adf517c9944d1e2190aad719fb84 Head: 4c77e7988541fde9d2f1e5ef89087ae6b87f89e6 Two separate installed worktrees. Fresh independent data directories. T3CODE_DEV_SEED=0. No client connected. No Google login. No model turn or replay provider. Settings: backgroundActivityProfile=performance; one enabled acpRegistry instance, source=registry, agentId=antigravity-acp, distribution=auto. An empty tools/antigravity-acp/0.0.0/windows-x86_64 directory marks the agent prepared. T3 itself downloaded registry runtime 1.5.0 on both sides. Launched node apps/server/src/bin.ts --base-dir --port 3999 --no-browser. Resolve node to process.execPath before Start-Process to bypass the Windows shim. Three startup cycles per commit; 36 samples per cycle, each after a five-second wait. Recursive directory scans add overhead, so elapsed times differ between sides. Only TEMP directories created after each side's baseline are counted. Existing leftovers are excluded. Each sample records Win32_Process entries for agy_acp_server.exe and file counts, sizes, and agy_acp_licenses.txt marker counts for new TEMP and T3-owned _MEI directories. All 108 samples per side have a live server. Each probe ends in Authentication required. Servers are stopped only after their observation window, using the captured process ID. Base cycle final totals: 2 folders / 243199683 bytes; 4 / 530583447; 6 / 817967211. Head: unpacking observed under T3-owned directories in all three cycles; zero new TEMP folders throughout, zero unpack folders after each probe. Final owned root has zero files and zero run directories. Head did not delete the base's pre-existing TEMP leftovers. Videos and stills are rendered views of the raw filesystem sampler output, not desktop screen recordings. Videos replay one sample every 0.5 seconds, explicitly accelerated. Raw samples and a sanitized summary of actual trace spans accompany the visuals.