[kitlangton]

Why

The CLI's managed-service tests spawn real serve --service processes and are among the slowest in the CLI suite. Much of that time is avoidable test overhead: every spawned process transpiles the CLI from a cold cache, and three tests wait out a server's real 5 s registration check to prove decision logic that doesn't need a real process.

What Changes

Test-only. No production code changes.

• Shared transpiler cache: each isolated test home sets XDG_CACHE_HOME, which also moved Bun's transpiler cache into it, so every spawned CLI process started cold. isolatedEnv and serviceEnv now point BUN_RUNTIME_TRANSPILER_CACHE_PATH at one shared directory.
• In-process registration checks: deleting, corrupting, or replacing a managed registration now runs ServiceRegistration.register in-process under TestClock, instead of starting a real server and waiting 5 s. The real-process test for a server whose boot failed stays, because it proves the failed server is still watching its registration. service-lifecycle.test.ts also still shows a real server exiting after its registration is corrupted.
• One duplicate removed: the occupied-port test in service.test.ts repeated the case in service-lifecycle.test.ts. That case now also asserts the exact conflict message, the opencode service set port <port> hint, and that no registration was written.

Unchanged on purpose: the 8 s wait in the bind-race incumbent test, which must exceed the earlier 6 s incumbent window, and the server's 15 s bind-recovery window that the occupied-port cases wait out.

Verification

cd packages/cli
bun typecheck
bun test --timeout 60000 test/service.test.ts test/service-lifecycle.test.ts

• All tests pass: service.test.ts 21/21 and service-lifecycle.test.ts 7/7.
• Back-to-back on the same heavily loaded machine (load average around 35), the two files went from 244 s on v2 to 196 s. service.test.ts alone went from 114 s to 70 s.