[lachlangrose]

Draft, stacked on #375 (stage 1), which is stacked on #374. Its diff against master includes those commits until they merge; the change here is the last commit. Stage 2 of the plan in #374's description.

Recipes now rebuild faults and unconformities as well as foliations.

Change

• Faults. A fault is stored with its creation parameters (displacement, axes, buffer, fault_center, ...) and its explicit faults list by name. faults=[] (no faults, what create_and_add_fault passes by default) is kept distinct from faults=None (every fault in the model), because they build different models. The reader resolves the names against earlier entries.
• Unconformities. add_unconformity / add_onlap_unconformity record the feature, the value and whether it is onlap, on the object that actually lands in model.features (for onlap that is the inverse, named __<feature>_unconformity_inverse). The reader replays add_unconformity / add_onlap_unconformity in recipe order, which reproduces the regions they add to earlier features.
• Faults assigned to a foliation after it was built (the old post-hoc restore) are applied only to foliations; faults and unconformities are built with theirs.
• An entry that refers to a feature that is not an earlier entry is refused with a clear error.
• Still refused when the recipe is written: fold frames, folded foliations, domain faults, intrusions, custom data=, parameters JSON cannot hold.

Evidence

• Fault model (the gallery's fault example): rebuilt strata and fault displacement are identical (difference 0.0; before this PR 0.21 on a field of range 10, and 0.49 for the displacement). The rebuilt fault is a FaultSegment.
• Two-group model with an erode unconformity, and the same with an onlap unconformity: identical fields for both groups, identical NaN patterns, identical evaluate_model unit ids, identical feature list and feature types.
• faults=[] vs the default: the two models differ, and each rebuilds to itself (the test checks they differ, so it is not vacuous).
• New test_recipe_faults_unconformities.py (11 tests, including a fault + unconformity model). Against the stage 1 code the 10 tests that existed when I checked all failed; the combined-model test was added afterwards and I did not run it against stage 1. test_recipe_limits.py updated: the stage 1 "fault/unconformity is refused" tests are replaced by a domain-fault refusal.
• tests: 757 passed, 6 skipped before the combined-model test was added (it passes on its own, together with the rest of its file). ruff: no new findings.

Things to know

• A domain fault is refused, but under "no record of how it was created" rather than "created as domain_fault": the object in model.features is not the one create_and_add_feature returns, so it carries no record. The behaviour (refused, naming the feature) is what matters; I did not chase the cause.
• Unconformities are replayed by add_unconformity on the rebuilt base feature, so they depend on feature order. The recipe keeps model.features order, which I verified for the cases above. A model with a fault, two groups and an unconformity also rebuilds exactly (tested). I did not test several unconformities in one model.
• Not done (stage 3): folds. Their fold functions are callables, which JSON cannot hold.
• COMPAT.md row updated.

🤖 Generated with Claude Code