fix(ios): pay for one toolchain in the snapshot bridge identity read (#2712)

[thymikee]

Summary

verify-installed-snapshot-bridge died on xcrun --sdk iphonesimulator --show-sdk-version, twice at the 30 s probe ceiling, and a cache-key field took the iOS smoke job down.

Triage disagreed with the reported cause. That script ran 11:15:44.964 → 11:16:45.862 (60.9 s) against a 120 s deadline, and the identity read runs before the cache lock, so remainingMs cannot reach 0 and the retry cannot have been skipped. The only reading that fits is two 30 s attempts on xcrun: the retry fired and stalled too — #2422's retry only recovers a stall that finished while the killed exec was going away.

The fix is at SnapshotSourceToolchainIdentity. The Simulator SDK ships inside the selected Xcode.app, so the version and build xcodebuild -version reports already pin it: the field left the key and the exec left the identity read. xcrun clang in the build pays that wait once, under the build's own ceiling, where it fits instead of exhausting the budget. A stall now reports toolchain-probe-stalled / native-build-stalled with the command, each attempt's budget and the exec kill as cause, instead of an unattributed xcrun timed out after 30000ms that reads like a device failure. 6 files.

Cache keys change, so existing entries rebuild once; old manifests hash differently and cannot be misread, so no schema bump. Untouched: the sibling runner prober's policy, and the generation circuit a preparation stall still opens.

Validation

4b1cc67ff: pnpm check:affected --run (all runnable gates, 1712 tests), check:fallow, check:layering, check:bundle-owner-files, build, lint, typecheck clean. Six tests fail with the two source files reverted to the merge-base.

Host run the review asked for (macOS 27.0, Xcode 27.0): prepared from source, exit 0 in 1.9 s, second run reused the entry.

$ pnpm --filter @agent-device/platform-apple run \
    verify-installed-snapshot-bridge <checkout> /tmp/sb3
"toolchain": { "xcode": "Xcode 27.0\nBuild version 27A5252f", "macosBuild": "26A428",
  "macosProductVersion": "27.0", "architecture": "arm64",
  "simulatorRuntime": "installed-package-verification" }
$ xcrun --sdk iphonesimulator --show-sdk-version   # the dropped field, on this host
27.0

A cold stall cannot be reproduced on demand; the attribution's real proof is the next cold CI host. ios.yml filters on this directory, so this PR runs the job it fixes.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.80 MB	4.80 MB	+639 B
Package (unpacked)	4.80 MB	4.80 MB	+639 B
Package (download)	1.43 MB	1.43 MB	+227 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.5 ms	26.4 ms	-0.1 ms
CLI --help	77.4 ms	76.5 ms	-0.9 ms

[thymikee]

Reviewed at 6f458ca: the code looks good. Dropping the Simulator SDK version from the cache key is sound, because the selected Xcode version and build already pin that SDK, and old manifests cannot match the new key. The retry, deadline and cancellation behavior is unchanged, and a cancel cannot turn into native-build-stalled.

One small note, not blocking: the compile-stall test builds the timeout error by hand instead of through exec.ts (https://github.com/callstack/agent-device/blob/6f458ca/packages/platform-apple/src/snapshot-source/cache.test.ts#L289), so it would not catch a later change to that error's shape.

I did not see the output of the reported host run. A cold xcrun stall cannot be reproduced on demand, so the real proof comes when CI hits a cold host.

Two Smoke Tests jobs are still running. The iOS one runs the snapshot bridge preparation step that this PR changes, so it should finish green before merge.

[thymikee]

Thanks — fixed at 4b1cc67ff, at the fixture rule rather than the one line.

__tests__/exec-timeout-fixture.ts now runs a command the exec layer really kills at its timeoutMs, and both stall fixtures throw whatever exec.ts raises. blockForWholeTimeout keeps advancing the fake clock by the budget the probe was handed, so the deadline arithmetic stays deterministic while the error stops being hand-shaped. cache.test.ts's compile fixture does the same. Both sides still assert through isCommandTimeoutError(error.cause), never a message.

Host run output added to the Validation section. Noted that a cold stall cannot be summoned: the attribution's proof is the next cold host, and it should now read toolchain-probe-stalled rather than a bare xcrun timed out after 30000ms.

Smoke Tests: all green on 6f458ca34; running on this head.

[thymikee]

CI on 4b1cc67ff: all checks green (17 pass), and the iOS Smoke Tests job ran the step that was failing — pnpm check:package -- --verify-snapshot-bridge-preparation — and logged Prepared the Simulator snapshot bridge from the clean-installed package. (job 106921424531). Linux/macOS/Android smoke jobs do not enable that flag, so they are unchanged.

[thymikee]

4b1cc67 looks good. The delta since 6f458ca is test-only: the fixture now takes its timeout error from the real exec.ts kill path instead of a synthetic one, so the snapshot bridge identity timeout test exercises the same failure the runner produces. No production code changed, and I found nothing left to fix.

CI is green on all 18 checks, including the smoke-test job that runs --verify-snapshot-bridge-preparation, which covers this path.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-09-23 05:40 UTC