arch: extract shared compute-driver discovery diagnostics

[kvnloo]

Context

Long-term follow-up to NVIDIA#3694 and NVIDIA#3211.

The CLI doctor patch intentionally does not duplicate schema-versioned gateway configuration or ComputeDriverRegistry selection logic. That keeps NVIDIA#3694 small, but means local doctor semantics and gateway startup discovery remain separate concepts.

Goal

Design a lightweight shared driver-discovery/diagnostic layer that can be consumed by gateway startup and diagnostic surfaces without making openshell-cli depend on openshell-server.

Desired properties

• Represent candidate drivers and probe outcomes structurally: selected, skipped, unavailable, permission denied, timed out, incompatible response, etc.
• Preserve gateway selection semantics, including explicit compute_driver and auto-detection ordering.
• Separate "runtime installed", "runtime service reachable", and "driver selected".
• Support actionable diagnostics without exposing credentials or sensitive endpoint details.
• Allow bounded probe timeouts and deterministic selection.
• Avoid duplicating schema-versioned gateway.toml parsing in the CLI.
• Be extensible beyond Docker/Podman to Kubernetes, VM, and MXC where appropriate.
• Give feat(gateway): explain compute driver discovery decisions NVIDIA/OpenShell#3211 a reusable diagnostic primitive rather than adding ad-hoc logging at each probe site.

Research questions

1. Which crate should own the shared candidate/probe types?
2. Which configuration-neutral inputs can be shared without importing server configuration?
3. Should gateway selection return a structured decision trace alongside the selected driver?
4. How should explicit configuration differ from auto-detection in diagnostic severity?
5. Which endpoint details are safe to log?
6. What timeout/concurrency policy should be common across drivers?
7. Which parts belong in runtime-specific driver crates versus a shared registry/discovery crate?

Non-goal

Do not block the scoped NVIDIA#3694 doctor patch on this refactor.

References

• feat(gateway): explain compute driver discovery decisions NVIDIA/OpenShell#3211
• feat(cli): support Podman in openshell doctor check NVIDIA/OpenShell#3694
• feat(cli): support Podman in doctor check NVIDIA/OpenShell#3748
• feat(cli): harden Docker and Podman doctor probes #4