Add third model family

[Danny1218]

Extend the calibrated transfer benchmark to a third model family.

Goal:

• Test whether the Qwen -> SmolLM2 fragility divergence appears again, narrows, or reverses.

Acceptance criteria:

• Add a documented model key or local registry entry.
• Run FP16 calibration and targeted transfer levels only after calibration passes.
• Add comparison artifacts and update the final evidence summary without mutating existing raw outputs.