Problem:
When a Runner is configured with auto_create_session=False, the A2A executor
layer completely bypasses this flag. Both A2aAgentExecutor._prepare_session
(legacy executor) and _A2aAgentExecutor._resolve_session (new executor impl)
call runner.session_service.get_session() and runner.session_service.create_session()
directly, instead of going through runner._get_or_create_session() (the method
that actually checks the auto_create_session flag). As a result, sessions are
always silently auto-created regardless of the flag value, making it impossible
to enforce explicit session management in A2A workflows.

Additionally, to_a2a()'s internally created Runner did not set
auto_create_session=True, which would cause every incoming A2A request to fail
with SessionNotFoundError since A2A clients never pre-create sessions.

Solution:

• Replace direct session_service calls in _prepare_session and
  _resolve_session with runner._get_or_create_session(). This single method
  already encapsulates the auto_create_session logic and raises
  SessionNotFoundError when the flag is False and the session is missing.
• Set auto_create_session=True in to_a2a()'s internal create_runner, to
  preserve backward-compatible behavior for users who rely on to_a2a() to handle
  session lifecycle automatically.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Updated tests in test_a2a_agent_executor.py and test_a2a_agent_executor_impl.py
to mock runner._get_or_create_session instead of runner.session_service.get_session
/ create_session. Rewrote test_resolve_session_creates_new_session to verify
the correct method and arguments are used.

Fixed tests/unittests/a2a/integration/server.py to pass auto_create_session=True
to FakeRunner, restoring the pre-fix behavior for integration tests where sessions
are always created on-the-fly.

pytest tests/unittests/a2a/ -v
...
56 passed, 139 warnings

pytest tests/unittests/ -v
...
5456 passed, 1 skipped, 2240 warnings in 130.24s

Manual End-to-End (E2E) Tests:

Created a minimal A2A server using to_a2a() with a LiteLLM-backed agent and
tested three scenarios:

Scenario 1 — auto_create_session=True (default via to_a2a):

curl -s -X POST http://localhost:8000/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "message/send",
    "params": {
      "message": {
        "messageId": "msg-1",
        "role": "user",
        "parts": [{"kind": "text", "text": "What time is it?"}]
      }
    }
  }' | python3 -m json.tool

{
    "result": {
        "status": { "state": "completed" },
        "artifacts": [
            { "parts": [{ "kind": "text", "text": "It's currently 2026-04-15 22:51:30 UTC." }] }
        ]
    }
}

Scenario 2 — auto_create_session=False with missing session:

{
    "result": {
        "status": {
            "state": "failed",
            "message": {
                "parts": [{ "kind": "text", "text": "Session not found: <session_id>" }]
            }
        }
    }
}

Scenario 3 — auto_create_session=False with pre-created session:
Session pre-created via lifespan hook → request succeeds with state: completed.

curl -s -X POST http://localhost:8000/ \   
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "message/send",
    "params": {
      "message": {
        "messageId": "msg-1",
        "role": "user",
        "contextId": "my-test-session",
        "parts": [{"kind": "text", "text": "What time is it?"}]
      }
    }
  }' | python3 -m json.tool

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

The runner._get_or_create_session() method already existed and was used
correctly by all other runner entry points (run_async, run_live,
rewind_async). The A2A executor was the only layer that bypassed it, making
auto_create_session unreliable in A2A-based deployment.