Add evidence-backed LangChain migration skill

[adtyavrdhn]

Summary

• add a repository-owned migrating-langchain-to-pydantic-ai skill to the ai plugin and standalone skills catalog
• require every non-1:1 mapping to name a concrete Pydantic AI workaround, residual limitation, disposable parity probe, and evidence status
• include a deterministic LangChain/LangGraph inventory scanner, project-scale Open Deep Research guidance, and trusted semantic oracles
• bump the ai plugin to 0.2.0 and enforce byte-identical plugin/standalone copies

Relationship to #32

This is complementary to #32 rather than a second Deep Agents skill:

• this skill owns ordinary LangChain/LangGraph migrations: create_agent, StateGraph, LCEL, LangSmith, and Open Deep Research
• Add independent Deep Agents migration skill #32 owns create_deep_agent, Harness profiles, repository skills, planning/compaction, sandboxes, Open SWE, and Deep Agents deployments

The frontmatter explicitly excludes Deep Agents. The scanner detects Deep Agents usage and routes it to migrate-deep-agents-to-pydantic-ai; negative trigger cases and a cold forward test verify that split.

Evidence-backed workarounds

The bundled evaluator runs checked-in behavioral oracles against real LangChain/LangGraph and Pydantic AI APIs with deterministic models. It verifies:

• dynamic-prompt replacement versus instruction composition, including continued history
• same-handler retry through Hooks.on.tool_execute, preserving LangChain's two-model/two-tool-call trajectory
• interrupt re-entry versus authenticated deferred approval, including unknown, foreign, and consumed-ID rejection
• run_stream() early finalization versus complete run_stream_events() execution after retry

Other recipes cover explicit structured-output transport, return_direct, durable runtimes, fan-out/reducers, event contracts, and concurrency-safe persisted budgets. Unexecuted target-specific behavior remains integration-required rather than being called native.

Validation

• plugin/standalone byte-sync check
• shell syntax and JSON validation
• Ruff check and format validation using the Pydantic AI configuration
• 6 inventory scanner tests
• skill-creator quick_validate.py
• official skills-ref validate
• trusted semantic oracle execution against LangChain 1.3.14, LangGraph 1.2.9, and Pydantic AI 2.10 development sources
• low-effort forward tests for generic LangGraph ownership and Deep Agents routing
• repeated adversarial review with no remaining P1/P2 findings