[Epic] Qualify and pin Nemotron 4 as a NemoClaw runtime profile

## Outcome

NemoClaw provides a qualified, reproducible Nemotron 4 runtime profile on at least one approved hardware target. A user can select the pinned profile and complete setup, first run, inference, diagnostics, recovery, and cleanup with evidence tied to the exact model, runtime, hardware, and NemoClaw revision.

## Scope

- Select the supported Nemotron 4 model profile and the first hardware target to qualify.
- Define a repository-owned profile that pins the model revision, runtime image digest, serving settings, capabilities, and readiness contract.
- Integrate the profile through the existing serving-profile and onboarding experience.
- Validate setup, cold and cached startup, inference, status, diagnostics, recovery, and cleanup on physical hardware.
- Publish the supported configuration, limitations, troubleshooting guidance, and reproducible qualification evidence.

## Acceptance criteria

- [ ] The exact Nemotron 4 profile and first hardware target are accepted before qualification begins.
- [ ] The profile pins the model revision, runtime image digest, serving configuration, and capability contract.
- [ ] The profile is selectable only on hardware that satisfies its declared readiness requirements.
- [ ] Setup, first run, inference, status, diagnostics, recovery, and cleanup pass on the approved physical target.
- [ ] Evidence binds the NemoClaw commit or tag, host and GPU configuration, model revision, image digest, launch configuration, observed performance, and relevant system errors.
- [ ] Documentation and the support matrix advertise only the configuration supported by the accepted evidence.
- [ ] Unsupported profiles and hardware fail before model download or runtime mutation with actionable guidance.

## Readiness and boundaries

- Qualification begins only after the supported model profile and first hardware target are accepted.
- This Epic qualifies Nemotron 4 as a runtime; it does not own model training, fine-tuning, or evaluation research.
- The [Post-Training Feedback Loop Blueprint](https://github.com/NVIDIA/NemoClaw/issues/8568) remains a separate product workstream.
- Qualification of one approved target does not imply support for every NemoClaw hardware platform.
- Changing an automatic model default requires a separate product decision.