Phase 15 — Performance and scale

Goal: be the fastest option for every common operation and handle structures from one atom to millions, with published, reproducible benchmarks against spglib, gemmi, ASE and pymatgen.

Release gate: v1.8.0.

---

Task 15.1 — Benchmarks as a contract

• [ ] 15.1.1 Comparative benchmark suite Description: benches/compare/ runs the same tasks (CIF read, P1, symmetry, neighbors, PXRD, supercell 10^6 atoms, porosity) in xtalforge and the reference libraries; results published on the docs site per release.
• [ ] 15.1.2 Regression gate Description: scripts/check-bench.sh fails the local CI gate when a tracked benchmark regresses more than 10 %.

Task 15.2 — Algorithms at scale

• [ ] 15.2.1 Structure-of-arrays storage for P1 cells and neighbor lists.
• [ ] 15.2.2 Streaming readers and writers for very large CIF/XYZ/LAMMPS files (constant memory per frame).
• [ ] 15.2.3 Parallel UFF energy/forces with a fixed-order reduction.
• [ ] 15.2.4 Lazy symmetry expansion: operations over the asymmetric unit without materialising P1 where the result allows it.
• [ ] 15.2.5 Configurable limits presets for million-atom work (huge).

Task 15.3 — Binary size and startup

• [ ] 15.3.1 Split WASM packages (core, symmetry, analysis, build) so a viewer loads only what it uses; size budget per package in scripts/check-wasm-size.sh.
• [ ] 15.3.2 Lazy data bundles (download on first use in WASM, compressed in the crates).
• [ ] 15.3.3 CLI cold start under 20 ms for info.

Task 15.4 — Surfaces and docs

• [ ] 15.4.1 Performance page with methodology, hardware and numbers.