Windows real-server branch pull-request lookup measurement Prerequisites: Windows Server 2025, Git, Node >=24.10, vite-plus (vp). Create two git worktrees, with dependencies installed by vp i: base = git merge-base origin/main 96b6b3071bf6c81dc81a2d64e392859cf5415462 = 1ec7f234ab5fe460f1ae53cfca73c4c830ac57f9 head = 96b6b3071bf6c81dc81a2d64e392859cf5415462 Keep measure.mjs and probe.mjs together in a scratch directory. Copy seed.ts into apps/server/seed.ts in each checkout (do not commit it). Run: node measure.mjs 15 The driver creates a new timestamped output directory. It creates one tiny git repository: a bare origin, main pushed with upstream, and feature/local-only with a local commit and no upstream. It creates each side's fresh isolated T3 home using its own real server, and creates one project and idle thread through HTTP and WebSocket APIs. It does not read or copy an existing T3 database and starts no provider turn. Agent providers are disabled in both isolated homes. Each round starts src/bin.ts serve directly from the relevant apps/server, with T3CODE_DEV_SEED=0, on loopback. The initial automatic thread-settlement sweep calls the real branchPullRequest service. The driver waits for that lookup's captured completion, then stops that exact child process. It excludes one warm-up pair, measures 15 pairs, and alternates which commit runs first. Each measured lookup therefore has a newly constructed manager and empty in-memory lookup caches. Both sides use the same physical synthetic git repo. probe.mjs uses Node's synchronous module load hooks to add read-only observation at the service boundary and immediately after successful git child spawns in the git driver. It also observes ProcessRunner spawns if any occur in the same lookup. It does not replace a git result, stub a service, change lookup options, add a delay, or change lookup control flow. Timing starts before the existing branchPullRequest effect and ends after its successful result, before writing the JSONL record. Only spawns for the measured repo during that lookup count. The capture omits physical paths and includes the commands and lookup result. measurement.json contains every retained sample and summary statistics. The spread is the full maximum-minus-minimum range, without outlier removal. The success criterion is that the median improvement exceeds both sides' spreads and every result remains null (no pull request for this unpublished branch). This measures the actual lookup at server startup, not total server startup.