[rominf]

Follows up #649.

write_fake_executable's errors now name the path and the step that failed (spawn, send, wait or chmod) and keep the original error kind, so each call site's single .expect still shows whether the write or the chmod failed. The two rocm-smi stubs in xtask/src/workflow_contract.rs now say which preflight (bash/sh or pwsh) they belong to. This change also drops the dead stdout(Stdio::null()) and documents why the module is a plain pub module rather than gated behind a feature: its callers are test code in other crates, which can't see cfg(test), and a feature would have to be enabled in every consumer's dev-dependency.

A new test, a_failed_chmod_names_the_path_and_the_step, makes a real chmod fail. Removing the chmod error context fails it.

Verified locally:

• fmt
• tests for rocm-core (623), xtask (247) and e2e-cucumber lib (142), plus feature_naming
• both clippy invocations, on Linux and for x86_64-pc-windows-gnu

No scenario: this is test-support code, and nothing a CLI user sees changes.

[juhovainio]

Reviewed this (test-only, ~40 lines across 2 files) with two independent passes. No blocking issues.

Verified: all four step_error(...) call sites correctly map to their failure modes (spawn/wait/send/chmod) with the right ErrorKind preserved; the two xtask .expect(...) edits line up with run_preflight/run_preflight_pwsh; the new chmod-failure test is a real, reliable repro (Unix-only via cfg(all(test, unix)), relies on real EPERM, correctly guarded by the existing running_as_root() convention); the stdout(Stdio::null()) removal has zero behavioral effect since the spawned command's own shell redirect already controls where output goes.

Two minor non-blocking notes, neither worth holding this up for:

• The stdout(Stdio::null()) removal is a touch of a behavior-adjacent change riding along in a test(...)-typed commit without being called out in the new doc comment. Harmless since it's confirmed dead code, but a skimming reviewer could miss it.
• running_as_root() unconditionally returns false on non-Linux unix (e.g. macOS), so this new test would misbehave if ever run as root on a hypothetical macOS runner. Pre-existing, not introduced here, and there's no macOS CI lane today, so purely latent.

Approving - nothing here blocks merge.

[rominf]

Thanks. On the two notes:

• stdout(Stdio::null()): the description calls the removal out ("drops the dead stdout(Stdio::null())"). The squash merge uses the PR title and description, so the merged commit will record it too. The spawned command already redirects its own output, so removing the call changes nothing at run time.
• running_as_root() on macOS: that platform is out of scope. AGENTS.md lists Windows and Linux as the only supported hosts and asks not to change logic for other platforms, and there is no macOS lane. I am leaving it as it is.

[rominf]

The merge queue ejected this on engine-shell-01 (merge-group run 38473081629). #656 does not touch the engine shell. The step checks for a prompt line before bash has drawn one, a race #663 fixes. Re-queued unchanged.