…ssion, include_shared)

Before this PR, the LangGraph adapter only exposed raw `namespace=`
strings — multi-agent users had to hand-build `f"agent-runtime:{id}"`
on every call, and the multi-agent guide's advertised "LangGraph
adapter for multi-agent" was a thin wrapper around `start_session`
that hard-coded `namespace="default"`.

This PR threads a multi-agent semantic through the adapter:

* New `start_agent_session(agent_id, *, namespace=None)` derives the
  namespace from `agent-runtime:<agent_id>` (override available),
  records the session in storage, and binds `_current_agent_id` so
  subsequent search/add calls inherit the agent scope.

* New `search(..., include_shared: bool | None = None)` parameter
  encodes a documented 6-case state table:

    | include_shared | _current_agent_id | resolved namespace                  |
    |----------------|-------------------|--------------------------------------|
    | None (auto)    | "planner"         | "agent-runtime:planner,shared"      |
    | None (auto)    | unset             | caller's namespace= (legacy)        |
    | True           | "planner"         | "agent-runtime:planner,shared"      |
    | True           | unset             | raises ValueError (no silent leak)  |
    | False          | "planner"         | "agent-runtime:planner" (no shared) |
    | False          | unset             | caller's namespace=                 |

  `True` without a bound agent raises explicitly so a missing
  start_agent_session surfaces immediately rather than degrading to a
  silent un-pinned search.

* `add(..., namespace=None)` defaults to the bound agent's private
  bucket. Pass `namespace="shared"` to publish across agents
  mid-session — the explicit value always wins, so the caller can
  cross the bucket boundary without re-binding the session.

* `end_session` now resets both `_current_session_id` and
  `_current_agent_id`, so a stale agent binding can't survive into a
  new session. The legacy `start_session(agent_id, namespace)` is
  preserved as a low-level escape hatch with a docstring pointer to
  `start_agent_session`.

Explicitly out of scope: implementing LangGraph's `BaseStore`
(`aput` / `aget` / `alist_namespaces`) with the same multi-agent
awareness — tracked as a follow-up RFC. The adapter stays a
hand-rolled wrapper; multi-agent helpers live on `start_agent_session`
and `search(include_shared=)` only.

Test surface (test_langgraph.py):
- `TestMemtomemStoreInit.test_agent_id_none_initially` — fresh store
  reports no bound agent.
- `TestResolveSearchNamespace` — 6 cases, one per row of the state
  table above. The `True + unset` case asserts the explicit
  ValueError so silent fallback can never sneak in.
- `TestResolveAddNamespace` — 4 cases for the `add` namespace defaulting:
  no agent / agent + None / agent + explicit override.
- `TestStartAgentSession` — agent binding, derived namespace,
  explicit-namespace override, empty-agent-id rejection,
  end_session resets both fields.

Co-Authored-By: Claude <[email protected]>