…g skill

Model fixes:
- Cross-attention timestep: always use cross-modality sigma instead of
  conditional on use_cross_timestep (matching reference preprocessor which
  always uses cross_modality.sigma)
- This was the root cause of the remaining 3.56 pixel diff — the diffusers
  model used timestep.flatten() (2304 per-token values) instead of
  audio_sigma.flatten() (1 scalar) for cross-attention modulation

Pipeline fixes:
- Per-token timestep shape (B,S) instead of (B,) for main time_embed
- f32 sigma for prompt_adaln (not bf16)
- Audio decoder: .squeeze(0).float() to match reference output format

Parity-testing skill updates:
- Add Phase 2 (optional GPU/bf16) with same capture-inject methodology
- Add 9 new pitfalls (#19-#27) from bf16 debugging
- Decode test now includes final output format (encode_video, audio)
- Add model interface mapping as required artifact from component tests
- Add test directory + lab_book setup questions
- Add example test script templates

Result: diffusers pipeline produces pixel-identical video (0.0 diff) and
bit-identical audio waveform vs reference pipeline.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Adds the LTX-2.3 modular pipeline structure:
- modular_pipelines/ltx2/: encoders, modular_blocks, modular_pipeline
- Registration in __init__.py, auto_pipeline.py, modular_pipeline mapping
- Checkpoint utilities for parity testing
- Supports T2V with CFG guidance (pixel-identical to reference)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>