Add Verlet-style neighbor list cache to calculators

[sirmarcel]

Summary

• Add NeighborListCache (src/lorem/neighborlist.py) implementing the Verlet displacement criterion: rebuild only when the combined position + cell deformation exceeds the skin budget
• Calculator builds neighbor lists with cutoff + skin, caches the batch, and on small geometry changes updates only sr.positions and sr.cell — the model recomputes R_ij from current positions/cell so forces/energy/stress stay correct
• Ewald k-vectors are recomputed automatically from sr.cell (pbc.k_grid stores integer frequency indices only), and Ewald parameters (lr_wavelength, smearing) are derived from the physical cutoff to keep the long-range decomposition unchanged
• skin kwarg (default 0.25 Å) added to Calculator, from_model, from_checkpoint, and LOREM_driver

Combined Verlet criterion (NPT-aware)

    Pair (i, j) with periodic image shift S:

    R_ij(t) = R_j(t) - R_i(t) + S · cell(t)

    ΔR_ij = (ΔR_j - ΔR_i) + S · Δcell
            ╰────────────╯   ╰────────╯
            ≤ 2 · d_max      ≤ max_shift · Σ|Δcell_A|

    Rebuild when:  2·d_max + max_shift · Σ|Δcell_A| > skin

max_cell_shift (max |S| in the neighbor list) is extracted from the batch at build time. Falls back to exact cell comparison when shift info is unavailable (backward compatible).

Test plan

Unit tests (test_neighborlist.py — 19 tests)

• needs_update returns True on first call (no reference)
• Returns False when atoms unchanged
• Returns False for displacement < 0.5*skin
• Returns True for displacement > 0.5*skin
• Boundary: displacement exactly at 0.5*skin (strict >)
• Cell change without shift info triggers rebuild (backward compatible)
• PBC change triggers rebuild
• Natoms change triggers rebuild
• Atomic numbers change triggers rebuild
• save_reference + reset lifecycle
• Cumulative displacement from reference (not per-step)
• Default skin is 0.25 Å
• Small cell change with max_cell_shift=1 stays within skin
• Large cell change with max_cell_shift=1 exceeds skin
• Cell change with max_cell_shift=0 has no effect (non-periodic)
• Combined position + cell change within skin
• Combined position + cell change exceeding skin
• Higher max_cell_shift amplifies cell contribution
• Reset clears max_cell_shift (reverts to exact comparison)

Integration tests (test_calculator.py — 10 tests)

• Default skin value
• Skin configurable via from_model
• Skin vs no-skin gives matching energy/forces on static structure
• Small displacement reuses neighbor list, results match fresh calculator
• Forces correct after position-only update (energy changes, shape correct)
• Large displacement triggers full rebuild
• NPT: small cell scaling reuses cache, energy/forces match fresh calc
• NPT: stress correct after cell-only update
• Large cell change triggers rebuild
• Combined position + cell change within skin, correct results

All 47 tests pass, lint clean.

🤖 Generated with Claude Code

[PicoCentauri]

Isn't this something the engine should do?