[zzet]

Adds a startup scenario to the bench/load harness. It records what agents actually receive while a daemon starts: from before its socket exists (including a first boot that runs a schema upgrade), through loading, to ready, on an initial boot and on warm restarts. It produces the runtime evidence #1023 asks for. Its time-to-ready and time-to-first-success figures can be compared between builds with delta.

Includes PR #1162. This branch was cut from #1162's refreshed head (252b9d2), so it contains all of #1162's commits. Only the last commit (ac0447a, feat(bench): probe startup admission from before the socket exists) is new here. #1162 must land first; after that, this PR's diff is that one commit. The commit touches bench/load only, and bench/load on main is the same as on #1162's head, so it also applies to main on its own.

Draft status: the probe is written, but none of its checks have run yet (see "Testing").

What it does

• --scenario startup is opt-in and never part of all. It configures every manifest repository and adds --worktrees linked checkouts of the first one.
• --seed-store copies an existing store into the private run first, read-only from its source, so the initial boot runs its migration before the socket listens. --restarts N adds N warm restarts.
• Session probes. Every second, it probes from three directories: the primary, a linked worktree, and a directory whose .gortex.yaml names a workspace claimed by a private servers.toml roster. From each, it opens fresh sessions and records:
  • initialize (which instructions variant was served), search_symbols and resources/read gortex://index-health on the legacy surface;
  • workspace(operation:"index") on the compact surface.
• Hook and CLI probes. From the same directories it records the PreToolUse hook's decision on a native Read every second. It also runs gortex call search_symbols through a pipe and through a pty, each until its first success in that boot.
• What each answer keeps: error_code, retryable, phase, percent, latency, exit code, and whether a refusal advises gortex track or gortex repos reconcile. Negated guidance is ignored.
• Watcher. It polls the readiness probe every 100 ms. For 2 s after the primary registers, it also reads the checkout coordinator count from daemon status.
• Output:
  • a per-second startup-probe.jsonl;
  • a startup_probe gate that counts answers per surface, directory and window (no socket, loading, ready), plus per-boot timings;
  • headline metrics for delta, such as startup_repo_not_tracked_before_ready, startup_nonconforming_loading_refusals, startup_restart_time_to_ready_p95_ms, startup_initial_time_to_ready_ms and startup_coordinators_within_2s_min. Restart figures are reported unavailable unless every restart produced them.

The scenario lives in new startup_probe*.go files. The shared harness files gain only its registration, a few lines each. The fixture daemon now answers the CLI and hook probes, so the whole scenario runs end to end in tests.

Testing

The tests are written but have not been run: the local check run (gofmt, go vet on three platforms, focused race tests, an end-to-end test run three times, pinned lint) was interrupted before it started. CI on this PR is their first run. The written tests cover:

• unit tests for the decoder, advice detection, CLI and hook classification, response framing, first-success timing and the per-window counters;
• an end-to-end test that runs the whole scenario against the harness's fixture daemon across one restart.

Not done

• Runs on the reference host: warm restarts with many linked worktrees, a cold start with worktrees, and an upgrade from an older store. They wait for perf(daemon): add fair lanes for MCP dispatch #1159 and fix(startup): preserve loading admission until repository identity is ready #1162 to land, and the results go on Prove startup repo_loading across warm restart, cold start and upgrade, and close the gaps #1023.
• Merge conflicts. Against feat(bench): add strict budget collection and nightly evidence checks #1160 and feat: measure edit-loop phases and add edit-and-probe diagnostics #1169, only bench/load/run.go conflicts, and every conflict keeps both sides' additions.

Refs #1023

[zzet]

Superseded by #1162, which now contains the exact probe commit ac0447a in its published ancestry. The combined integration passed focused race and prescribed pinned checks before publication; fresh exact-head CI runs on #1162. The original probe branch and its evidence remain preserved. Startup runtime and performance acceptance remain open in #1023.