[tgockel]

a_global_images_own_entry_is_not_refused_through_its_stamped_label and its two siblings in the CLI's embedded_image.rs run outrig mcp over a probe, sh -c 'printenv STOLEN >&2', and assert that its stderr, quoted in the startup error, holds the variable. live-e2e failed on them in 5 of the 73 trunk-targeting runs since they landed on Oct 20, and in 4 runs on version/0.3.x, as exit: code 0 beside stderr (...): (empty) (#729).

The probe is an exec-stdio server that writes one line and exits within milliseconds, and podman exec -i can return its exit status without that line. #422 found the same loss under podman start --attach and traced it to the client's initialize message still being unread when the process exits; it restores the text from conmon's log, which an exec session does not have (#727). The probe now reads that line before it prints.

The library's embedded_image.rs has the same probe in probe_label, and launch_resolves_a_label_reference_the_spec_lists asserts on its stderr through McpStartupFailure::stderr_tail. It was not among the failures #729 counted, but it runs the same command through the same exec path, so it reads first too. The other copy of the command there, in a_spec_entry_replaces_the_label_entry_and_is_not_held_itself, only checks which error variant comes back and is left alone.

Reproduced here without outrig: podman exec -i of the old probe with an initialize line already in its stdin pipe and the pipe held open, as rmcp leaves it, 24 at a time pinned to four cores. The old probe lost its stderr in 24 of 400 runs and in 19 of 400 on a repeat, each time with exit status 0, the CI symptom. The probe that reads first lost it in none of 800. Both embedded_image binaries pass locally under --features e2e (15 CLI tests, 13 library tests). What does not follow from this is why the read helps: the effect is measured, and the explanation is #422's applied to exec.

Closes #729

Summary

What does this change, and why?

Checklist

• cargo fmt --all -- --check passes
• cargo clippy --workspace --all-targets -- -D warnings is clean
• cargo test --workspace passes
• Docs in doc/ updated if behavior changed (mdbook build still clean)
• CHANGELOG updated for user-visible changes

Notes for reviewers

Anything worth calling out -- tricky bits, follow-ups, or intentionally out-of-scope items.