fix(drivers): validate vendor-agnostic CDI device names

[alangou]

Summary

Reject host-device paths and malformed CDI selectors before Docker or Podman container creation, while preserving valid selectors from any vendor. Both drivers use shared CDI qualified-name validation for explicit requests and resolved device IDs.

Related Issue

No issue required: localized validation fix for the existing cdi_devices input.

Administrator-controlled CDI admission is a separate follow-up: #4327. Podman server inventory discovery is also being handled separately.

Changes

• Add shared validate_cdi_device_names in openshell_core::gpu, using a local is_qualified_name implementation aligned with upstream CDI grammar.
• Validate explicit selectors through a shared CDI-specific GPU-request wrapper for Docker and Podman, and validate resolved selectors at both drivers' production container-spec construction boundaries. Keep the generic GPU-request validator identifier-agnostic so VM PCI addresses remain valid.
• Preserve NVIDIA index, UUID, MIG, and all selectors, and accept syntactically valid non-NVIDIA names without claiming additional GPU execution support.
• Retain Podman create-preflight coverage and add shared parser and Docker/Podman explicit/resolved selector regressions.
• Update published docs and the cluster debugging skill.
• Track replacement of the local parser with the public upstream API in container-device-interface-rs PR #177.

Testing

• mise run pre-commit (repository-wide lint, formatting, and license checks).
• Changed-crate library tests: 900 passed, 1 ignored across openshell-core, openshell-driver-docker, and openshell-driver-podman.
• Full VM library suite: 221 passed with e2fsprogs supplied through a temporary Nix shell, including all PCI-address selection regressions.
• Changed-crate Clippy with all targets and -D warnings.
• Rust formatting and git diff --check.
• mise run docs: 0 errors and 3 Fern warnings.
• Podman GPU E2E: not run locally for this update. Existing PR CI covered Podman E2E and Docker GPU E2E; the pushed revision requires fresh CI results.

Checklist

• Follows Conventional Commits.
• New commit is signed off for DCO.
• Relevant user documentation and skill guidance updated.

[github-actions]

Label test:e2e-gpu applied for 20e5f49. Open Branch E2E Checks, find the run for commit 20e5f49, and click Re-run all jobs to execute with the label set. The run will execute GPU E2E after building the required supervisor image once. The matching required CI gate status on this PR will flip green automatically once the run finishes.

[github-actions]

Label test:e2e applied for 20e5f49. Open the existing run and click Re-run all jobs to execute with the label set. The run will execute the standard E2E suite after building the required gateway, sandbox, and supervisor images once. The matching required CI gate status on this PR will flip green automatically once the run finishes.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-4306.docs.buildwithfern.com/openshell

[elezar]

One thing to note is that CDI is not NVIDIA specific, so we may have to check whether there is another way to perform the required validation or limit device names in some way.

[elezar]

The shared CDI qualified-name check is now implemented for Docker and Podman. Both drivers use validate_cdi_gpu_device_request for explicit selectors and validate resolved selectors at their final container-spec construction boundary. The generic validate_specific_gpu_device_request remains identifier-agnostic so VM PCI selectors stay valid. Shared parser and driver regression tests cover host-path rejection and valid vendor-agnostic names.

The local implementation matches upstream's is_qualified_name until its public API is released. I've opened container-device-interface-rs PR #177 to expose is_qualified_name and parse_qualified_name and remove the validation's stdout output.

The follow-ups remain separate from this PR:

• Draft PR #4329 adds Podman server CDI inventory discovery through the existing /v5.0.0/libpod/info endpoint. It distinguishes authoritative empty inventory from unsupported discovery, preserves legacy path-based fallback, and refreshes before default GPU validation and creation.
• Issue #4327 covers designing and implementing administrator-controlled CDI device admission.

Podman 6's CDI directories and discovered-device fields were verified live in tmachine with Podman 6.1.1 and a synthetic non-NVIDIA CDI spec; see Podman PR #28712. That verifies discovery, not non-NVIDIA GPU execution. Automatic GPU selection retains its existing NVIDIA policy.

[elezar]

Thanks @alangou. Since this was a shared effort, please merge if you're happy with my adjustment.