Assemble the parity snapshot set

[farrisric]

Child of #23.

Question

Which ~10 Ag(111)+O configurations form the committed parity set?
They should span the O-coverage range the ladder visits, from bare slab to oxide-like.
Decide the committed format (extxyz plus reference energies) and how references are produced (single-point and relaxed energies with pinned settings).
Produce the candidate set as a prototype to react to.

Blocked by: #25

[farrisric]

Decision half settled with the user (2026-08-18):

• Source: snapshots are harvested from the pilot-isotherm run of Calibrate production and smoke tier run lengths #25 - real sampled GCMC states across the 8 rungs, spanning the O-coverage range the ladder visits.
• Reference pinning: reference energies are generated with the same calculator settings as benchmark runs (compiled, cuEq on, default dtype) - the gate answers "did numerics change for the runs I actually do".
• Format: extxyz for configurations + JSON with single-point and relaxed reference energies.

Remaining: harvest the candidate set once #25 has run, produce it as a prototype to react to. Dependency is now #25, not #24.

[farrisric]

Correction to the reference-pinning decision above (user, 2026-08-18): the calculator is now checkpoint mh-1 (head mp_pbe_refit_add), compile on, cuEq OFF - cuEq conversion corrupts mh-1. References are generated with these settings. Everything else stands.

[farrisric]

Head correction (user, 2026-08-18): mh-1 head is now omat_pbe, not mp_pbe_refit_add. References generated with mh-1/omat_pbe, compile on, cuEq off. mu references still runtime-derived, so no numeric constants to update.