schema: hoist inline definitions into named $defs for codegen

[ad-cqc]

Motivation

The current schema is correct for validation, but its pervasive use of inline definitions makes it hostile to code generation (Python, Julia, or any target that turns a JSON Schema into a set of named types in a single module):

• Inline object schemas inherit the surrounding key name. Two unrelated inline objects under different parents but with the same key (e.g. Attributes, Direction, CoordinateSystem) collide as soon as the generated types share a namespace.
• Inline schemas inside arrays produce synthetic …Items types. SurfaceCurrent.items was an inline object, which generators render as SurfaceCurrentItems. The same was true for Elements, producing SurfaceCurrentElementsItems — a name with no relationship to anything in the source of truth.
• oneOf-by-required discriminators are anonymous. LumpedPort and SurfaceCurrent items used oneOf: [{required:[Attributes]}, {required:[Elements]}] over a single inline object. Generators can't see two distinct shapes, so they emit one type with every field optional and lose the discriminator.
• Inline string enums collide. Repeated literals like ["Cartesian", "Cylindrical"] or ["MGS", "CGS", "CGS2"] show up at multiple call sites; without a named definition, each call site produces a separate, name-clashing enum.

What this PR does

Commit 1: schema refactor (scripts/schema/config/*.json)

Pure refactor — no semantic changes. The same documents that validated before still validate, and the same documents that failed still fail.

For each of boundaries.json, domains.json, model.json, problem.json, solver.json:

1. Hoist inline object schemas into \$defs with unique, source-traceable names. Examples:

   • boundaries: PEC, PMC, Absorbing, Conductivity, WavePort, WavePortPEC, Ground, ZeroCharge, Terminal, Periodic, BoundaryPostprocessing
   • domains: Material, DomainPostprocessing, DomainEnergy, Probe, CurrentDipole
   • model: Box, Sphere

2. Split anonymous oneOf discriminators into named variants:

   • LumpedPort.items → oneOf: [LumpedPortAttributes, LumpedPortElements]
   • SurfaceCurrent.items → oneOf: [SurfaceCurrentAttributes, SurfaceCurrentElements]

   This gives codegen two distinct, well-named types plus a tagged union, instead of one type with everything optional.

3. Promote inline string enums to named \$defs: CoordinateSystem, ProblemType, GSOrthogonalization, AdaptiveGSOrthogonalization, AdaptiveCircuitSynthesisDomainOrthogonalization, MaterialAxes, Direction, DipoleDirection.

4. Name polymorphic value shapes that previously appeared as anonymous oneOf fragments:

   • Samples → SamplesPoint | SamplesLinear | SamplesLog
   • Port Direction → PortDirection | Vector3
   • Port Excitation integer → ExcitationIndex

Commit 2: validator follow-ups (palace/utils/jsonschema.cpp, test/unit/test-schema.cpp)

Caught by running the schema unit tests locally; required to keep behavior consistent under the named-\$ref schema.

1. FindAllSchemasByKey now resolves internal #/\$defs/ refs instead of skipping them. Previously the function continued past any \$ref into \$defs, which meant schemas only reachable through a named ref (e.g. BoundaryPostprocessing.FarField) were no longer findable via ValidateConfig(data, "FarField"). Adds a depth guard against pathological cycles.
2. SchemaErrorHandler enhances type-mismatch and enum errors using substring matching. With LumpedPort.items now a oneOf, the validator wraps simple errors as [combination: oneOf / case#0] unexpected instance type. The exact-string check stopped matching, so the (got string) / valid-enum hints disappeared. Switching to find() makes the enhancement fire regardless of oneOf/anyOf wrapping.
3. Two exact-string assertions in test-schema.cpp relaxed to substring checks. The validator's behavior is unchanged, but the wording for Index failures shifted from a flat single-message form to a oneOf-enumerated form. Asserting on the underlying message keeps the test resilient to validator-output formatting.

Why this is safe

• Every schema change replaces an inline schema with a \$ref to a \$defs entry whose body is byte-equivalent (modulo whitespace) to the inlined version.
• additionalProperties: false, required, and oneOf/allOf constraints are preserved at the same nesting level.
• For LumpedPort / SurfaceCurrent, the two split variants carry the full property set; only the required differs, so no document that was valid before is rejected now.
• FindAllSchemasByKey continues to return the same results for top-level property lookups; it now also finds keys reachable only through internal \$defs refs (strictly more permissive).

Test plan

• Local: cmake --build build/palace-build --target unit-tests && build/palace-build/test/unit/palace-unit-tests "[schema]" — 10 passed, 1 skipped (the Embedded Schema Matches Source SECTION skips by design when its source path isn't present), 0 failed, 155/155 assertions.
  • Sub-schema lookup by key for LumpedPort / WavePort / SurfaceCurrent / CurrentDipole / Materials / FarField.
  • LumpedPort / SurfaceCurrent oneOf: [Attributes, Elements] discrimination.
  • PEC/Ground and PMC/ZeroCharge mutual exclusion.
  • All 24 example configs validate.
  • Problem.Type → Solver requirements (Driven/Eigenmode/Transient reject if matching solver section absent; Electrostatic/Magnetostatic accept defaults).
  • Error-message format: enum hints, type-mismatch hints, oneOf-wrapped messages, additional-property failures.
• CI style.yml "Check JSON Schema" — scripts/validate-config runs over every examples/**/*.json against the refactored schema.
• CI unit-test job (test-schema.cpp).

[hughcars]

Generally LGTM, one small request on error/warning handling. I think we should probably hold off merging this until #719, and #716 go in.