[kitlangton]

What it does

Adds a test that counts the work done by a cold folder boot plus one session turn. It fails if a speedup from #58600, #58601 or #58602 regresses. It asserts exact counts, not timings.

How it works

1. Boots a real git folder and runs one turn with TestLLM. Nothing goes over the network.
2. Counts the work through layers the tests already swap in:
   • git spawns, through a wrapped ChildProcessSpawner;
   • catalog lookups and Azure model edits, through a wrapped Model.Service;
   • git PATH lookups, through a spy on which.sync.
3. Asserts these counts:
   • git spawns: 8 during boot and 9 during the turn (17 total, the number refactor(core): resolve real git binary once on macOS #58602 measured);
   • git executable lookups: 0, and every spawn uses the resolved path;
   • catalog lookups per step: 1;
   • Azure model edits on a model rebuild, with Azure enabled but no resource name: 0.
4. On failure, it prints each count that moved and by how much, followed by every git spawn. Changing a count on purpose is fine. Update the number in the test and say why in the PR.

Risk

Test only, no src changes. A legitimate new git spawn or catalog lookup fails this test until the expected number is updated, which is the point.

Proof

• Reverting all three speedups locally fails the test with: git lookups 0 → 17 (+17), catalog lookups per step 1 → 6 (+5), Azure model edits 0 → 78 (+78).
• Passes with bun run script/test.ts test/boot-work.test.ts and bun test, with the same counts across 6 repeated runs.
• bun typecheck (core), oxlint and prettier are clean.

[opencode-agent]

The new test runs the same way every time and catches the regressions it guards. I ran it 12 times (some under CPU load) and the counts never changed. With #58601 and #58600 reverted locally it fails with a clear drift report (catalog lookups per step 1 → 6, Azure model edits 0 → 156) and lists every git spawn. tsc -b on core is clean.

Public API is unchanged; the change is test-only.

No blocking concerns. One note: on Linux CI, git is never looked up on PATH, so the git-lookup and resolved-path checks always pass there. Only the macOS and Windows runs exercise #58602's part, which the test's comment already says.

[opencode-agent]

The follow-up is fine. It allows one git PATH lookup, for the Windows Git Bash lookup. A regression back to a lookup on every spawn would still fail the test, since that adds about 17 lookups. The test still passes repeatedly on Linux and tsc -b on core is clean.

Public API is unchanged; the change is test-only.

No blocking concerns. Minor: the extra lookup is allowed on every platform, not only Windows, so a single new git lookup on macOS would go unnoticed.