add ToolCallLimits capability on PydanticAIBaseAgent via scalar wiring

[sanzog03]

Summary

Fills in PydanticAIBaseAgent._build_capabilities_from_scalars() — Consumers who set max_tool_iterations or max_tool_calls on the config now get a hard runtime cap enforced during the model-request loop and tool-execution
path, with AKD's existing UsageLimitExceeded family of errors surfacing cleanly through pydantic_ai's UnexpectedModelBehavior.

What it does

• Agents built on PydanticAIBaseAgent respect max_tool_iterations and max_tool_calls from the inherited BaseAgentConfig — the ReAct loop no longer runs unbounded.
• When a cap is hit, the run raises MaxToolIterationsExceeded / MaxToolCallsExceeded (both subclasses of UsageLimitExceeded), mirroring the semantics the legacy OpenAI-basedBaseAgent already exposes.
• Subclasses can still extend the capability list by overriding _build_capabilities_from_scalars, calling super(), and appending (the extension pattern is now pinned by a test).
• No new config fields on PydanticAIBaseAgentConfig — both scalars come in from BaseAgentConfig unchanged.

Files changed

• New files:
  • akd_ext/agents/_base/pydantic_ai/_capabilities.py — holds ToolCallLimits (Hooks factory, closure state, two conditional lifecycle callbacks).
• Modified files:
  - akd_ext/agents/_base/pydantic_ai/_base.py — imports ToolCallLimits; _build_capabilities_from_scalars body replaced with the max_tool_iterations / max_tool_calls.
  - tests/agents/test_base_pydantic.py — adds four tests under ToolCallLimits" section; restores one field on _EchoConfig (enable_extra_capability: bool) that the subclass-extension test needs.

Testing

Run the test suite (hermetic; no network, no model keys):

  uv run pytest tests/agents/test_base_pydantic.py -n=3