feat(tii)!: interpret the full complex param-type model

[scarmuega]

Summary

Reworks ParamType to interpret every shape tx3c emits into the TII params schema, per the canonical model in the SDK spec's api-surface/args.md. The python-sdk slice of a cross-fleet complex-type parity effort (rust/go/python/web).

Changes

• Flat str Enum → ParamKind + ParamType dataclass carrying element/field types (inner/elements/fields/cases/schema); adds UNIT/UTXO/ANY_ASSET/TUPLE/MAP/RECORD/VARIANT/UNKNOWN.
• Trailing-name $ref matching across tii#/$defs/ and legacy core# forms (previously the tii# form degraded to a typeless custom); drops the latent string→ADDRESS default.
• Never raises — unrecognized shapes become UNKNOWN; resolves #/components/schemas/<Name> refs (threaded through Protocol.invoke).
• Protocol.invoke now also exposes party (Address) and environment-schema params, matching the reference SDK.

Tests

pytest — 62 pass (new test_param_type.py over the full canonical table incl. both ref forms, nested compounds, component resolution, never-raise fallbacks; existing tii suite green). ruff clean.

Validated against the shared fixture sdk-spec/test-vectors/complex-types/complex.tii via the full Protocol.invoke path (parties + env + tx params + component resolution).

Breaking change

ParamType changes from a str Enum to a dataclass; the kind discriminant moves to the new ParamKind enum. Needs a version bump per sdk-spec/release-policy.md.

🤖 Generated with Claude Code

[scarmuega]

Pushed a behavior-preserving readability pass (1afc17d) — the same review we applied to rust-sdk #41 and go-sdk #14, against the four code-quality dimensions. No change to the kind map; all existing tests still pass.

1. Module cohesion — already good, no change. Unlike rust (where the interpreter was buried in mod.rs), Python already has a dedicated tii/param_type.py. Nothing to extract.

2. Function decomposition. param_type_from_schema was a ~90-line function with the $ref, array, and object arms inline. It's now a thin dispatcher delegating to _ref_type / _array_type / _object_type (joining the already-extracted _variant_case / _core_kind_from_ref). Same fallbacks, recursion, and kind map.

3. Dead-code removal. Once the interpreter became never-failing it stopped raising InvalidParamTypeError, leaving it dead — only the class def + export survived, referenced by no test. Removed it (and its two __init__ export lines). InvalidParamsSchemaError stays — protocol.py still raises it.

4. Integration coverage. test_param_type.py covers the interpreter well in isolation, but the Protocol.invoke path this PR added — threading components, party→Address, environment params — was untested against a real complex TII (test_tii.py used all-primitive transfer.tii). Added test_invoke_interprets_complex_params + the shared complex.tii fixture, asserting the compound kinds plus a resolved component-$ref Record (asset.policy → Bytes) and the 2-case Variant — the assertion that actually guards the components threading.

pytest -m "not e2e" → 68 passed; ruff check clean.