Issue 40005 performance evidence driver

Requirements: Linux, Node 26.13.1, Python 3, git, gh and glab installed; tea and fj absent.
No CLI shims, fake clocks, mocked services, or unit tests are used.

1. Create the two worktrees named in run.py from these exact commits:
   before: 1ed4f6a96f93b86839f998ede47c46261c08a4d0
   after:  87f1bfe6aaedea1346a5dd29c262927ed3a4689a
2. Run vp i in each checkout.
3. Put run.py, measure.ts and summarize.py together in an empty scratch directory.
4. Run python3 run.py. It creates a timestamped directory and fresh server homes,
   starts each checkout's real server from source with T3CODE_DEV_SEED=0,
   creates the same synthetic git repository and origin (myhost:repo), and
   auto-bootstraps its project. One completed Replay (dev) turn creates real history.
   It then issues an authenticated session and a WebSocket ticket using the app CLI/API.
5. A real Effect RPC client calls vcs.refreshStatus with localOnly=true.
   This follows GitManager.readLocalStatus -> resolveHostingProvider ->
   SourceControlProviderRegistry and therefore the reported unknown-host probe path.
   localOnly avoids fetching from an unrelated nonexistent SSH endpoint.
6. One warmup is recorded and excluded. Five subsequent samples each follow
   a real 310-second wait (longer than the old five-minute unknown-host cache TTL).
   Both servers run on the same machine; calls are staggered to avoid overlap.
   The driver records monotonic round-trip timing, the full RPC result, and trace spans.
   It stops its captured server PIDs after the measurements complete.
7. Run python3 summarize.py <the printed timestamped run directory>.
   This exports before/after.txt and .json plus comparison.json. Only synthetic
   repository paths are replaced by <synthetic-repository>. Timings are untouched.
   Probe and CLI counts are limited to descendants of refineUnknownRemoteProvider
   in each measured RPC trace. Server timings are included as a cross-check.

The performance comparison uses the five post-expiry samples. The head must have
an identical normalized result and a median improvement larger than the maximum
of the two observed max-minus-min ranges. Warmup is not included in the median.

The source fix extends a settled verdict's cache TTL from five minutes to one hour,
and unifies Forgejo's refinement context. It does not add a plain-git host setting.
This driver measures the VCS refresh path and startup traces; it does not claim to
exercise every PR sweep, UI entry point, the one-hour expiry, or login invalidation.

Additional export commands after run.py completes:
  python3 startup.py <run-directory>
  python3 render.py
  python3 export-logs.py <run-directory>

render.py formats the exported measurement text as PNGs and makes a small HTML chart.
export-logs.py preserves the source logs with local paths and credentials redacted.
CLI versions recorded in this run: gh 2.97.0; glab 1.124.0. Linux x86_64, 8 logical CPUs.
