fix(doctor): warm the runner cache only for simulators this host discovered

[thymikee]

Summary

On macOS, three provider-scenario doctor tests started a real background xcodebuild build-for-testing. Doctor's warmup read inventory without its discovery source. The scenario worlds supply their simulator through a provider inventory source, so doctor treated it as a host simulator.

Doctor now keeps the inventory source and picks a warmup device only from host-discovered devices. Device selection uses the same rule before local Apple selection (source === 'local'). The session-device path does not change. In production, provider inventories list only provider-owned devices, which isProviderDevice already skipped.

Rejected: Apple tool providers on the worlds, so #3225's gate applies. Doctor gets a provider scope only when one device resolves. With #3225 applied, 3 of 5 warmups still started.

4 files. hermetic-signal-setup.ts is not changed.

Validation

Tested commit fb0c11a013, macOS:

• New unit tests: the 3 provider-reported cases fail without the fix and pass with it.
• Instrumented prewarmAppleRunnerCache count, vitest run --project provider-integration test/integration/provider-scenarios/:

Tree	doctor	keep-hot
main	6	51
main + this PR	1	51
main + #3225 + this PR	0	0

The remaining doctor warmup is the apple-platform-output-guard session path, inside the Apple tool scope #3225 gates. With both PRs, 70/70 files pass. remote-proxy-parity fails on main before and after (#3225 fixes it).

• Linux (node:22-bookworm): base 939a4ce070 and head have the same status for all 226 provider-integration tests, and all doctor tests pass. The same 34 tests fail at both: 33 because the image has no zip, and 1 on lease-upload timing.
• pnpm check:affected --run: all runnable checks pass.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.06 MB	5.06 MB	+241 B
Package (unpacked)	5.06 MB	5.06 MB	+241 B
Package (download)	1.52 MB	1.52 MB	+71 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.3 ms	26.8 ms	-0.5 ms
CLI --help	86.2 ms	83.6 ms	-2.6 ms

[cubic-dev-ai]

No issues found across 4 files

_{Re-trigger cubic}

[thymikee]

This PR is ready as of fb0c11a. All 19 checks pass, and there are no conflicts. I ran no tests locally; I read the diff and the pre-change code, and I did not reproduce the prewarm counts or the Linux parity run from the PR body. No live agent-device doctor run against a real host simulator is attached, so real-host behavior rests on the source === 'local' path and unit tests that use a fake local inventory. Not blocking, and fine to take or leave: hostDevices is derived by object identity after filterInventoryForSelector. It works today because every step keeps the same objects. If a later step clones devices, hostDevices becomes empty and doctor stops warming the runner on real hosts, with no failing signal beyond the host-source unit test. Filtering the combined host list with the same selector predicate would remove that coupling. The multi-line doc comment above resolveHostWarmupDevice is also a note only.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-05 11:36 UTC