fix(ci): halve smoke-lite concurrency, which #715 made too high for podman

[pofallon]

test_compose_override_command_lifecycle_runs has failed 2 of the 3 Podman runs since #715, always the same way — podman cannot start an exec session:

container create failed (no logs from conmon): conmon bytes ""

which surfaces as a bare deacon up failed and reads like a lifecycle defect. It isn't one.

It's contention, and #715 caused it

Compose calls used to run on a hardcoded docker even under DEACON_CONTAINER_RUNTIME=podman. So on the Podman lane, four concurrent smoke-lite tests were four concurrent docker compose projects. Now they're four concurrent rootless podman ones, each with its own conmon and network helper, on a shared runner.

Evidence

	Result
Podman CI since #715	2 of 3 runs failed, same test, same conmon error
Same test locally, serial, real rootless podman 4.9.3	5/5 pass
Full integration suite locally under podman	433/433 pass

Passes serially, fails under concurrency — that's contention, not a defect in the test or the code.

This is the blunt lever, deliberately

[test-groups] has no per-profile form, so 4 → 2 slows smoke tests on every lane, not just Podman. I'm taking that trade because a lane failing two runs in three is worse than a slower one — but flagging it rather than burying it.

The narrower alternative, if the wall-clock cost shows up: move only the compose smoke binaries into their own group and leave the rest of smoke-lite at 4. Happy to switch to that instead.

Not a re-run

I re-ran the first occurrence and it passed, which is why I didn't act then — one data point with a plausible story isn't evidence. Three data points, a failed reproduction attempt, and a mechanism that explains why it passes locally and fails in CI is a different situation.

🤖 Generated with Claude Code

https://claude.ai/code/session_016SFzA2sTh2EpX8MZU3TWNS