[STORY-020] Verify both ornith models on the RTX 2060 (sm75) without the FA NaN crash

[dogkeeper886]

Context

Part of STORY-020 · plan #416

Goal

Verify both ornith models run coherently on the RTX 2060 (sm75) WITHOUT the flash-attention
NaN crash. Because ornith runs on the qwen35 arch (in the FA allowlist), FA enables on Turing —
so this run is only valid on an image that carries the #385/#395 Volta-only FA gate.

Done When

• The sm75 run uses a gate-carrying image (confirmed in the prereq task); ornith does NOT
  panic: failed to sample token on Turing.
• ornith:9b and ornith:35b pre-pulled on the sm75 runner.
• test-models.yml run with runner_label=sm75; both models generate coherently.
• Where ornith:35b exceeds the 2060's 6 GB, the offload split + throughput are reported
  as observed facts, not a correctness failure.
• Results captured on this issue.

[dogkeeper886]

Dependency note (from dw-review-tasks): the sm75 run must use the same ported image built in #424 — i.e. the image that carries both the ornith renderer/parser port (#422) and the #385/#395 Volta-only FA gate. #421 confirms the gate is present; #424 produces the actual image. Order: #421 → #422 → #424 → #425.