Order Moire C4 interactions by coupling magnitude

[changkai-zhang]

Summary

• Sort Moire C4 interactions by descending absolute coupling before AutoMPO builds the Hamiltonian. Python's stable sort preserves their original order at equal magnitude.
• Update the model's SHA-256 in registry.yaml. The interaction map and coefficient calculations are unchanged.

Motivation and limits

At N=48, the interaction-map order can move from coefficients around 1e-10 to around 1e-2 when an index family changes. Ordering by magnitude avoids those large scale jumps in successive AutoMPO additions.

This ordering is already present in our campaign-local model override, including its Paratera deployment. It is a numerical mitigation, not a complete cure: a Paratera A6 N=48 continuation using this ordered model still encountered Torch SVD nonconvergence during AutoMPO compression and needed a separately guarded SVD fallback. I am contributing the ordering independently of that runner-level fallback.

Verification

• python tools/verify_registry.py: all four registry entries OK.
• git diff --check: clean.
• Updated model matches the deployed campaign-local source byte for byte (SHA-256 809fa874fe7b1972ab3f38c89e9b1a591227acf667820c029c8effc41df03baa).
• No new scientific calculation or Slurm job was launched for this PR.