CRAP follow-ups: test-quality fixes from child PR reviews #3461

Purpose

The maintainer's reviews of the merged CRAP burn-down child PRs (#3414) left take-or-leave test-quality suggestions. This issue keeps the ones that catch a real regression or remove flakiness, and drops the purely cosmetic ones (casts, PR-body counts, import style). Each retained item must name the regression it would catch.

Area 1: client and perf-capture tests (#3429, #3432) — #3463

• Ineffective assertion: the close-code test in client-companion-tunnel-worker.test.ts only pins the final wire code. normalizeOutgoingCloseCode re-maps 1011/1001/1005 to 3001, so changing normalizeCloseCode doesn't fail it. Pin normalizeCloseCode's own output through the public seam, or drop the redundant first stage. (test(client): cover the out-of-process CRAP functions through their public seams #3429)
• Shared scenario state: the test in client-metro-companion.test.ts that chains seven scenarios lets case 7 reuse case 6's state file, so one failure hides the later ones. Use test.each, or one test per case with its own state. (test(client): cover the out-of-process CRAP functions through their public seams #3429)
• Wrong path in a comment: the EPERM comment in client-metro-companion.test.ts cites src/daemon/process/daemon-stop.ts; the file is src/daemon/daemon-stop.ts. (test(client): cover the out-of-process CRAP functions through their public seams #3429)
• Message-string assertion: the "rejects a non-xctrace kind" test in platform-runtime-perf-capture-host.test.ts asserts a message string. Assert a typed error code if wrongNativeKind has one. (test(platform-runtime): cover Apple perf capture and runner transport branches #3432)

Area 2: Android, Linux and Apple tests (#3433, #3434, #3435) — #3462

• Wall-clock flakiness: the boot-timeout test in emulator-lifecycle.test.ts runs 80 ms polls against a 150 ms deadline and asserts exactly 2 calls, so 20 ms of overshoot fails it. Use fake timers or a lower-bound assertion. (test(android,harmonyos,linux,provision-kit): cover boot-failure and toolchain probe branches #3433)
• Assertion that can't fail: the notEqual(BOOT_COMMAND_FAILED) right after the ANDROID_BOOT_TIMEOUT equality can never fail. Remove it. Keep "keeps original details" only if the reason override is intentional. (test(android,harmonyos,linux,provision-kit): cover boot-failure and toolchain probe branches #3433)
• Linux negative controls: the degenerate-rect snapshot test pins nothing observable. The missing-depth test would still pass with a default depth of 1. Add a depth-0 sibling with a smaller area after it. (test(android,harmonyos,linux,provision-kit): cover boot-failure and toolchain probe branches #3433)
• Host-binary fixture: the IPA fixture in install-artifact.test.ts shells out to the host zip. Use host-kit's writeZipFixture. Also, the spent-deadline test in physical-device-files.test.ts should assert details.step. (test(platform-apple): cover core launch diagnostics, CoreDevice normalization, runner-file copy, and IPA hint resolution #3434)
• Mock scope: host-provider.test.ts returns a cast partial fs.Stats, and its scratchHome() tmp dir is unused. Use a real tmp directory for the user-Applications case and mock only the system roots. The "only needs to be a record" test in runner-xctestrun-products.test.ts is weaker than its neighbors; make it pin an observable outcome. (test(platform-apple): cover macOS host app discovery and xctestrun product collectors #3435)

Dropped as cosmetic: inline import(...) types and the double cast in the perf-capture test, fallow-crap.json comment wording, createCompanionBridgeServer migration of older tests (separate refactor), and PR-body line counts.

Completion

• Tests only, through public entry points, with no production changes, no new exports and no test-only DI. Each PR ≤ 1,000 gross lines.
• pnpm check:affected --run passes on the pushed head.

Part of #3414.