fix(vllm): validate persistent container configuration

[nvddr]

Summary

• label repository-owned vLLM containers with a deterministic launch fingerprint
• reuse healthy or stopped containers only when their launch contract matches
• recreate stale or unlabeled containers before model startup

Why

A persistent container could be restarted with model arguments, GPU assignment, or memory limits from an older profile even after the active YAML changed.

Validation

• uv run --project tests pytest -q tests/test_vllm_docker.py
• ruff check on changed Python files
• SPDX header check

[nvddr]

Review outcome: ready for review

The branch is rebased on current main, and the earlier blockers are resolved:

• Healthy ports are reusable only when the expected named Docker container is running and carries the matching launch fingerprint; an unowned listener now fails without being stopped.
• The fingerprint includes a versioned launcher contract, so changes to launcher-controlled Docker behavior (entrypoint, bootstrap, fixed flags) invalidate older immutable containers.
• Matching stopped containers restart, matching containers still starting are monitored instead of receiving a conflicting second docker run, and stale containers are recreated.

Validation:

• Focused isolated suite: 42 passed (tests/test_vllm_docker.py)
• Ruff: passed
• Fresh GitHub CI: all checks passed (Python 3.11/3.12, docs build/publish, Ruff, lock, SPDX, DCO, and CodeQL)

No unresolved review threads remain. The PR is ready for maintainer review and merge.

[wenxind-nvidia]

Follow-up review

Blockers

None. The update at 724efff resolves the previous blockers:

• Healthy listeners now require the expected running container and matching fingerprint.
• The fingerprint includes a versioned launcher contract.
• The branch is current with main, conflict-free, and mergeable.

Suggestion

[P2] Account for credential rotation in the container fingerprint

HF_TOKEN remains outside the fingerprint even though it is stored in the immutable container environment. A matching stopped container can therefore restart with an obsolete token after credential rotation. Consider hashing it into the fingerprint or documenting the intentional exclusion.

Nits

• Remove the unused existing and running assignments at lines 640–641 and 662–663, already flagged by the code-quality bot.

Validation

• tests/test_vllm_docker.py: 42 passed
• Ruff, SPDX, and diff checks passed
• GitHub CI: 14/14 checks passed