What changed

• select a representative S1 trajectory with medoid, dispersion, depth-clearance, and smoothness scoring
• cancel stale action blocks after collision or low GPS displacement and bound repeated-turn recovery
• restore the camera after lazy low-head RGB/depth capture
• track duplicate pixel goals, instruction state, stair-height evidence, and diagnostic context
• keep uncertain semantic context disabled in the evaluated conservative configuration

Why

The original evaluator averaged multimodal S1 candidates, continued stale actions after collisions, and could repeatedly retry unproductive directions. Camera-only depth capture also needed an explicit restore path.

Impact

The seed-42 regression preserved all nine previously successful holdout episodes and added one metric success. Recovery remains lightweight, while suspected rendering and high-level semantic failures stay visible in diagnostics.

Validation

• python -m pytest tests/unit_test/test_minimal_s1_safety.py tests/unit_test/test_navigation_state.py -q — 29 passed locally and on the cloud evaluation host
• two-GPU seed-42 regression: 20 holdout episodes plus selected prior failures