Format report should produce a non-empty string.

Format report should produce readable output.

Format report should produce readable output.

End-to-end: all MoE components should be modified together.

Fused 3D projections should preserve norms when requested.

Fused 3D parameter patterns (GPT-OSS style) should project both directions.

Should raise RuntimeError (not silently corrupt) when offload dir is missing.

Actual compliant responses should not trigger semantic detection.

GPT-OSS is MoE. Without config, defaults to small (conservative).

GPT-OSS with config providing expert count → small MoE.

Should handle 3D tensors by reshaping.

Should handle 3D tensors by reshaping.

Should handle activations with an extra batch dimension.

Should handle non-unit directions.

Selective head projection should only modify targeted head rows.

Head surgery on GQA non-square o_proj with norm_preserve.

Head surgery should work for GQA models with non-square o_proj (attn_dim != hidden_dim).

Head surgery with norm_preserve should maintain per-head norms.

Calling main(['--help']) should print help and exit 0.

The comparison table should include Heretic's published reference numbers.

Sparse surgery with large weight matrix.

Test with realistic hidden dimension (4096).

Higher z_threshold should mask fewer neurons.

Test install/remove lifecycle without a real model.

Test steering on a simple nn.Sequential model.

Should install hooks on MoE router modules.

Hooks should record router logits during forward passes.

Identical directions across layers should give persistence = 1.

KL divergence of identical distributions should be 0.

CKA of identical representations should be 1.0.

Identity matrix should have effective rank equal to dimension.