Correct hypothesis logic and opportunity evidence handling

[lullu57]

Summary

The current hypothesis evaluation logic and opportunity mining path can misstate evidence. H1 rewards instability, H2 is encoded more strictly than the research text, the measurement gate is not enforced before hypothesis decisions, and opportunity mining counts many unevaluated compile-space rows as evidence.

Evidence

• src/kernel_tuner/analysis/comparison.py:408
• src/kernel_tuner/analysis/comparison.py:450
• src/kernel_tuner/analysis/opportunities.py:97
• src/kernel_tuner/analysis/opportunities.py:261
• configs/studies/validation_phase.yaml:23
• docs/research/06_hypotheses_and_ablation_plan.md:11
• docs/research/06_hypotheses_and_ablation_plan.md:55

Why This Matters

This is the last step before benchmark evidence becomes paper claims. If this layer is wrong, the project can produce polished but misleading conclusions.

Expected Fix

• Correct H1/H2 study logic to match the intended hypothesis meaning.
• Enforce the measurement-validity gate before substantive hypothesis decisions.
• Restrict opportunity evidence to measured/decision-relevant cases rather than all cross-joined compile-space rows.
• Add tests covering study logic and opportunity aggregation.

Acceptance Criteria

• Hypothesis evaluation matches the documented research plan.
• Invalid measurement contexts produce inconclusive outcomes rather than premature support/unsupported states.
• Opportunity mining uses decision-relevant evidence only.