Python SDK: PydanticAI adapter - #131

Adds a PydanticAI adapter to the Python SDK so a PydanticAI agent can be bound to a governed run with no change to the agent beyond wiring the adapter:

from pydantic_ai import Agent
from riskkernel.adapters.pydantic_ai import govern

agent = Agent(govern("anthropic:claude-sonnet-4-5", run, gate_tools=True))
agent.run_sync("...")   # halts with BudgetExceeded when the budget is spent

How it works

I wrap the model with PydanticAI's WrapperModel (which forwards every model method to the wrapped model) and override only request/request_stream. One model request counts as one governed step, so the deterministic loop/time budget is enforced over the agent's outer loop and a halt surfaces as BudgetExceeded. With gate_tools=True, each tool call the model proposes is routed through the human-approval gate before the agent executes it (a denial raises ApprovalDenied).

This keeps the adapter on a single stable surface — the Model.request contract, which hasn't changed across the post-1.0 line — for both loop/time enforcement and tool gating, rather than depending on a newer hooks API. The proposed tool calls are present in the model's response before the agent runs them, so I gate them right after the response comes back and before they execute.

Propagation

PydanticAI only retries a model request on its own ModelRetry signal and re-raises every other model-request error by default, so I deliberately raise plain BudgetExceeded/ApprovalDenied (never ModelRetry): they propagate out of agent.run()/run_sync() and stop the agent rather than being retried into another paid request. This matches the propagation discipline of the LangChain (raise_error) and CrewAI (step_callback vs event bus) adapters.

pydantic-ai is lazily imported and added only as an optional extra (pip install riskkernel[pydantic-ai]), so the core SDK still installs and imports with no third-party dependencies. Supported against pydantic-ai (pydantic-ai-slim) >= 1, < 2.

Tests

tests/test_pydantic_ai.py — stdlib-only unit tests (step ticking, loop-budget halt surfaced-not-swallowed with the underlying request skipped, tool gating off/on/multi-call/final-answer/denied) plus two skipUnless(pydantic-ai installed) integration tests that drive a real Agent with a FunctionModel and prove a runaway agent halts at its loop budget and a denied tool blocks. Ran the full SDK suite (passes; pydantic-ai-gated tests skip without it) and verified the integration tests against a real pydantic-ai 1.107.0 install in a throwaway venv — all green.

Closes #85