feat(server): typed handler params via Pydantic annotation (closes #214)

[bokelley]

Summary

Closes #214 — the biggest structural boilerplate complaint across rounds 4–7 of DX validation. Handlers declare `params: GetProductsRequest` instead of `dict[str, Any]`; the dispatcher deserialises at the boundary.

```python
class MySeller(ADCPHandler):
async def get_products(
self,
params: GetProductsRequest,
context: ToolContext | None = None,
) -> GetProductsResponse:
# Typed attributes — no params.get("buying_mode") anywhere.
if params.buying_mode.value == "refine":
...
```

What's in

• Signature inspection (`_resolve_params_pydantic_model`). Reads the handler override's `params` annotation. Returns the Pydantic class for direct annotations and unions with dict (the specialized-base shape). Returns `None` on dict / missing / forward-ref failure — dispatcher falls back to dict path.
• Dispatcher (`create_tool_caller`). Inspection runs once at setup. On each call, `model_validate` the raw dict before calling the handler. `ValidationError` → structured `ADCPTaskError(INVALID_REQUEST)` with field path and Pydantic error details.
• Context echo preserved. `inject_context` still reads from the raw dict, so ADCP context echo works under typed dispatch (the wire `context` field isn't part of typed request models).
• Back-compat is automatic. Legacy `params: dict[str, Any]` handlers route unchanged. Sibling methods with mixed typed/dict signatures coexist.

What's out

• Response-side typed validation — current response builders return dicts; breaking that would require coordinated migration.
• Full seller-example rewrite — left to a follow-up once skills are updated to teach the typed pattern as primary. `examples/typed_handler_demo.py` is the minimal reference for now.

Test plan

• 13 unit/integration tests in `tests/test_typed_handler_params.py`
• Signature resolution: direct Pydantic, union with dict, missing, non-Pydantic class, plain dict
• Dispatch: typed handler receives Pydantic instance, legacy dict handler unchanged
• Error surface: Pydantic ValidationError → ADCPTaskError(INVALID_REQUEST) with details
• Mixed typed/legacy on same handler
• Context echo works under typed dispatch
• Existing `Model.model_validate(params)` pattern inside specialized bases still works (no double-validation crash)
• Custom Pydantic model (not SDK-generated)
• Second tool (list_creative_formats) — plumbing is tool-agnostic
• Full suite: 1863/1863 pass
• mypy clean (672 source files)

🤖 Generated with Claude Code