Generate a crystal structure with GPSK-300

@will, this route doesn't get enough attention for how interesting it is. A reciprocal-space DiT that conditions on formation energy, e_above_hull, band gap, and magnetic ordering is a different paradigm from the screening pipeline we've been building. Instead of filtering known structures for desired properties, you're generating new ones with those properties baked in. That's the generative complement to everything else on the platform.

The FeCoNiPt structure @mmoderwell generated here is a useful test case for what this route can and can't do yet. The min interatomic distance came out at 0.634 Å, which is well below any physical Fe-Co, Fe-Ni, or Co-Pt bond. That's not a knock on the model — it's the expected failure mode for multi-component generative models where the diffusion process hasn't fully converged on physically valid configurations. The "selected from 4 candidates" selection step helps, but it also means the other three were worse.

What would make this route more useful for the screening pipeline is a post-generation validation step. Right now the natural workflow is: generate with GPSK-300, then relax through Orb v3 or MACE-MP to see if the structure survives. The ones that relax to a reasonable min distance (>2 Å for transition metal alloys) with small energy change are the keepers. The ones that collapse or keep unphysical distances are filtered out. That's the same validation logic we're already applying to literature structures in the MLIP failure mode benchmark.

This also connects directly to the Grau-Crespo conversation. CrystaLLM and GPSK-300 are solving the same problem from different angles: CrystaLLM generates structures from text (composition or formula), GPSK-300 generates from composition plus property conditioning. A side-by-side comparison on the same composition set would be a genuinely interesting benchmark, and we now have the infrastructure to run it: GPSK-300 generates, cifkit analyzes coordination geometry, Orb v3 validates through relaxation, and the convex hull route checks stability. The full stack, end to end, on the same platform.

One concrete suggestion: adding a min_distance filter to the candidate selection (discard any generated structure with min distance below a chemistry-aware threshold before returning) would catch the 0.634 Å cases at the source and save the user a relaxation step.