fix(codebuild): sweep only build containers whose owning process is gone

[vieiralucas]

Summary

With #2520's retried-failure output, #2523's E2E run captured why codebuild_real_execution::cross_phase_shell_state_persists still failed intermittently after #2519:

Provisioning: FAILED  CLIENT_ERROR  "failed to create /codebuild/build: "

That is the docker exec mkdir into the build container right after docker run -d -- the container had already been removed.

Cause

At startup, CodeBuild swept leaked build containers: every fakecloud-codebuild container whose fakecloud-instance label was not the current process. That includes containers of other fakecloud processes that are still running on the same daemon. When a second server starts -- parallel e2e test servers, or two installs side by side -- it removes the first server's in-flight build container.

The shared startup reaper (fakecloud-server/src/reaper.rs) already does this correctly: it removes an object only when its owning PID is dead. CodeBuild's own sweep skipped that check.

Reproduced with a new e2e: start a build on one server, start a second server while it runs. Old sweep: the first build ends FAILED. New sweep: SUCCEEDED.

Fix

• pid_alive moves from the server reaper to fakecloud_core::container_net, plus owned_by_dead_process(label, is_alive): an object is orphaned only when its fakecloud-<pid> owner is neither this process nor alive, and an unparseable label is never treated as orphaned.
• The server reaper and CodeBuild's sweep both use it, so the two cannot diverge again.
• libc becomes a cfg(unix) dependency of fakecloud-core (already in the lockfile).

Also, test-only: ec2_instance_runtime's container should be running assertion now reports the container's status, exit code, OOM flag, error, start/finish time and logs. That test also failed once in #2523 (container present but stopped, not a pull failure) and the cause isn't identified yet; the next occurrence will name it.

Surfaces

No API, SDK, docs, conformance or count change: startup cleanup behavior only.

Test plan

• owned_by_dead_process unit test (live other owner kept, dead owner swept, self kept, unparseable labels kept) and a pid_alive probe test in fakecloud-core; the reaper's existing tests still pass against the moved function.
• New e2e another_server_starting_does_not_kill_a_running_build: passes with the fix; with the old sweep restored the build ends FAILED.
• cargo clippy -p fakecloud-core -p fakecloud-codebuild -p fakecloud --all-targets -D warnings, e2e test clippy, fmt clean; fakecloud-codebuild lib 67/67.

---

Summary by cubic

Fixes CodeBuild's startup sweep to only remove build containers whose owning fakecloud process is dead. Previously, any container not owned by the current process was removed, so a second server starting on the same daemon would delete another server's in-flight build and fail it in PROVISIONING.

Bug Fixes

• Moves the ownership check (pid_alive + owned_by_dead_process) into fakecloud_core::container_net and reuses it in both the server reaper and CodeBuild's sweep.
• Improves the ec2_instance_runtime "container should be running" assertion to include container state and logs on failure.

^{Written for commit 774b459. Summary will update on new commits.}