notebooks: guard nvidia-smi for CPU-only runtimes (G1.2)

[mattdani21]

PR_SUMMARY — nvidia-smi guard for CPU-only runtimes

What

notebooks/colab-phase1-eval.ipynb and notebooks/phase1-pipeline-eval.ipynb called subprocess.run(['nvidia-smi'], ...) unguarded. On a runtime with no NVIDIA driver (CPU-only Colab/Kaggle session), that raises FileNotFoundError and the notebook dies at cell 0 — before any eval work.

Why

Observed in a smoke run (python 3.12.13 | scope: smoke | categories: bugfix,feature): traceback FileNotFoundError: [Errno 2] No such file or directory: 'nvidia-smi'. GPU presence should be reported and handled, not assumed.

Fix

• New nvidia_smi(*args) helper: returns '' on missing binary / timeout instead of raising.
• CPU-only runtime now prints nvidia-smi NOT FOUND + GPU count: 0 | total VRAM MiB: 0 + a warning directing to Runtime → Change runtime type → T4 GPU.
• smoke scope proceeds on CPU (slow but functional — llama.cpp CPU fallback); full scope still fails loudly via the existing >=20GB VRAM assert (vram=0 → assert fires with the clear message).

How tested

• Both notebooks patched, JSON re-validated, diff limited to cell 0 (30 insertions / 9 deletions).
• Patched cell-0 logic executed locally on a machine without nvidia-smi (exact repro of the failure): no crash; smoke path proceeds, full path asserts with the intended message.

Numbers

• Hardware floor (eval path): smoke eval no longer requires an NVIDIA driver to start — CPU-only runtimes can validate the env (slowly) instead of crashing. Runtime hardware-floor targets unchanged; full-scope guard intact.
• Quality parity / swap latency / cost per task: unaffected (notebook tooling only; no benchmark/task code touched — suite, grader, and baseline runner untouched per governance §2.3).