[Validation] Fix 'Mirror Bug' Blindspot in Round-trip Verification

[DsThakurRawat]

Description (Serious Architectural Debt)

The validate_roundtrip.py utility currently validates the exported ONNX model against a 'decomposed' PyTorch reference implementation.

The Vulnerability

If the logic used to decompose the PyTorch model for export (e.g., the VLM KV-cache slicing) contains an error, the exporter will bake that error into the ONNX graph. Since the validation suite compares the ONNX graph against that same decomposed code, they will match numerically, and the bug will be validated as 'Correct' while the robot fails in the real world.

Task

• Implement a 'Truth Bridge' that validates against the monolithic (non-decomposed) teacher model for at least N samples.
• Add a STRICT_VERIFY mode that fails if the decomposed reference and monolithic teacher disagree, even if the ONNX matches the reference.