[niklashoelter]

Pins aimnet[ase,torchsim]==0.2.6, which includes our upstream fix isayevlab/aimnetcentral#150: the batched gas-phase Coulomb neighbor list is now sized by the largest molecule instead of the batch's total atoms. The lock only moves aimnet (0.2.5 → 0.2.6); its dependency ranges are unchanged.

With that, AIMNet2 batches no longer fill the GPU, so the temporary AIMNet2 L-BFGS warning and its docs, skill and AGENTS.md lines are removed. The general BFGS warning stays (its dense Hessian is a torch-sim limit for every backend). The n_atoms memory-metric rationale is reworded, since memory no longer grows with the square of the batch.

Testing

• Full suite: 833 passed, 1 skipped.
• Large L-BFGS jobs that ran out of memory on 0.2.5, now without gpuma's per-structure fallback:
  • aimnet2, 9.8k structures including 140–423-atom molecules: 105 s, all converged.
  • aimnet2, 29k structures: 243 s, all converged.
  • aimnet2-nse, 9.8k mixed: 114 s, all converged.