Instruction-column dominance probe: imperative vs non-imperative variants

Goal

Test whether the instruction-column dominance in #181's heatmap (51-80% marker rates across all training families) is driven by imperative surface form or is a deeper property of instruction-style prompts.

Hypothesis

Replacing the 9 instruction-family panel prompts with semantically equivalent non-imperative variants will drop marker rates by >10pp if surface form drives the effect, or maintain rates if it is a deeper property.

Setup

• 14 existing LoRA adapters from HF Hub (no new training)
• Same 36-prompt panel EXCEPT 9 instruction-family cells replaced with non-imperative rewrites
• Same scorer, vLLM params, 10 completions × 20 questions

Compute

~0.5 GPU-h on 1x H100 (inference-only)

Parent: #181 Clean result target: #207 (supplementary section)