refactor(adapters/langgraph): align ManagerWorkers code and tests with codebase style

[spichen]

Cleanup pass over oracle#211, plus one bug fix its own test exposed.

Bug fix: ManagerWorkers as a flow step could not produce its declared output

test_managerworkers_runs_as_a_flow_step_with_data_edge_inputs fails on the PR head: the ManagerWorkers parent graph runs over MessagesState, so the manager subgraph's structured_response never propagates outward, and extract_outputs_from_invoke_result finds nothing — any manager with declared outputs raised ValueError: Expected node ... to have a value for property .... (The fallback the test relies on lives in the unmerged feat/agent-single-string-output branch.)

AgentNodeExecutor._format_agent_result now answers a ManagerWorkers node's single string output with the manager's final message content, and raises NotImplementedError for other output shapes.

Test cleanup (797 → 421 lines)

• Removed the box-drawing section dividers and module docstring; no other test file uses them.
• Module-level pyagentspec imports and a shared _agent() helper instead of ~10 lines of Agent(...) boilerplate repeated per test; langchain/langgraph imports stay in-function, matching test_swarm.py / test_agentnode.py.
• pytestmark = pytest.mark.usefixtures(...) instead of an empty-bodied autouse fixture.
• Dropped micro-tests of trivial private helpers (_messages_of, _surface_to_parent_command, tool name/description, echo-worker wrapper) — all covered by the end-to-end delegation tests — and merged duplicates (test_normalize_identifier duplicated test_safe_node_name; the roster and routing tests are consolidated).
• The astream_events-unwrapped test asserts on compiled.__dict__ instead of patched-function __name__s.

Source cleanup

• Dropped the @lru_cache on the delegation-tool factory, replaced the __slots__ class _WorkerSubgraphNode with closures, removed the _tc_get dict-or-object shim (langchain tool calls are always dicts), and inlined _surface_to_parent_command into the delegation tool.
• Trimmed decision-justifying comments to codebase density while keeping the load-bearing ones (no-goto invariant, one-Send-per-delegation, ambient-config namespacing).
• black formatting fix for one over-long line in _langgraphconverter.py.

Net: −452 lines against the PR head. Verified offline: 17/17 ManagerWorkers tests pass, plus the langgraph flows suite, swarm, tracing/serialization/validation core suites.