feat: add real TypeChain Petstore agent factory

[sjungwon03]

Summary

Closes #21.

Add a real LangChain agent entrypoint over the existing read-only Swagger Petstore TypeMCP server.

• Add createPetstoreAgent({ model, server? }), which delegates to TypeChain createTypeMcpAgent() with an explicit resolver-backed PetstoreServer.
• Add a fixture runner and test using LangChain FakeToolCallingModel: the real agent loop selects and executes search_available_pets and returns the TypeMCP-derived fixture result.
• Document the distinction between the existing deterministic adapter workflow and the real agent factory.
• Keep model/provider selection, provider packages, credentials, authorization, retries, and observability application-owned.

Explicit boundaries

• No provider SDK or model identifier was added.
• No API key, credential, or environment-variable model configuration.
• No automated live LLM call in CI.
• No MCP HTTP/stdio transport, MCP client/session, or Petstore write operation.
• Existing Petstore client remains fixed-host and GET-only.

Verification

• npm ci --ignore-scripts --no-audit
• npm run lint
• npm run build
• npm test — 9 test files / 13 tests.
• npm run audit:prod — 0 vulnerabilities.
• npm run check
• npm run example:petstore:typemcp:fixture
• npm run example:petstore:agent:fixture
• npm run example:petstore:agent:real:fixture — observed available-pets-call and the two fixture pets.
• git diff --check

[sjungwon03-ai]

Approved at exact head 6934f3f. Real createTypeMcpAgent delegation with caller-owned model, genuine FakeToolCallingModel agent-loop proof against fixture data, TypeMCP stays read-only, README distinction is exact, and verify is green at this SHA with no unresolved threads.