Fix/verification

[huseyincavusbi]

Description

Fixes three verification bugs found during Gemma4 adapter testing:

• Component benchmark false failures — Models using DelegatedAttentionBlockBridge (Gemma4) reported 81-96 component failures per model, dragging P1 from 100% to 50%. These are benchmark infrastructure failures. The component comparison can't call delegated attention/rotary/PLE modules in isolation because they require model-specific kwargs the benchmark doesn't provide. Added skip logic in component_outputs.py that detects DelegatedAttentionBlockBridge (via missing hook_q_input in hook aliases) and skips untestable components (attn, rotary_emb, per_layer projections). The gold-standard forward_pass_logits test was already passing. This just stops the benchmark from reporting false negatives.

• Phase 2 empty header — main_benchmark.py printed a "PHASE 2:" header with zero tests on every run regardless of whether Phase 2 was selected. Moved the header inside the should_run_phase(2) guard.

• Misleading "encoder-decoder" warning — The HF model loading log printed "for encoder-decoder model" for multimodal architectures using AutoModelForImageTextToText. Removed the incorrect label.

Verification (tiny-random/gemma-4-e):

• Dev branch: 15/64 component failures
• Fix branch: 0/31 component failures (all untestable components skipped)
• Unit tests: 167/167 passed,

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• My changes generate no new warnings
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Looks good! One small suggestion: _is_delegated_block() detects delegation by checking that hook_q_input is missing from the block's hook aliases, which is a bit indirect. The benchmark already skips attention via an explicit flag (maintain_native_attention / requires_position_embeddings), you could set maintain_native_attention=True on DelegatedAttentionBlockBridge, which would then reuse the existing skip for that case, rather than adding the hook-alias check?