Performance evidence driver (scratch, no code modifications)

Prerequisites: Linux, Node 26.13.1, installed T3 Code worktrees for the base and head.
Install each with vp i. Copy driver.mjs into apps/server/proof-driver.mjs in each
checkout so the script resolves the checkout's workspace dependencies.

For each side, workload N in 0,50,500,2500 and replicate in 1,2,3:
  node apps/server/proof-driver.mjs CHECKOUT NEW_OUTPUT_DIRECTORY REDIRECT_PRELOAD N REPLICATE
Run from that checkout. REDIRECT_PRELOAD is the absolute path to redirect.mjs.
Every output directory must be new. Run one server at a time; alternate side order
between replicates. No real account is used. The driver creates and stops its server.

The driver starts the actual source server with fresh state, authenticates through
its CLI, creates a synthetic git project through HTTP, binds unrelated ids through
WebSocket RPC, then reads detail and an older comment page. It measures 300 detail
RPCs after 30 warmups. A Node fetch preload redirects only api.github.com requests
to a local HTTP fixture; the actual GitHub tracker adapter, credential resolution,
GraphQL decoding and HTTP transport still run. All responses are synthetic.

Trace timing: binding SELECT start to GitHubIssueTracker.getIssue start, matching
the triage metric. SQL duration, IssueService.detail duration and client round-trip
are also recorded. The first detail and 30 warmups are excluded from timing.

Outputs: results.json has all 300 RPC and trace measurements, response samples,
normalized detail SHA256, and fixture request counts. Uploaded raw-measurements.json
contains these results for every run, without server logs or data homes.

summarize.py ROOT writes measurement.txt and summary.json from ROOT/{side}-{N}-r{rep}/results.json.
