Research: node id scheme and scope mapping (issue #571)

Feeds ADR-0026. Scopes a node id convention for ADK's own real app_name/user_id/session_id/event_id hierarchy, read against all four existing adapters' own real precedent (ADR-0016, ADR-0020, ADR-0021, ADR-0023) rather than inventing a fifth scheme from a blank page.

The real scoping key is flat, not a tree: (app_name, user_id)

Checked directly against search_memory's own real signature (#569): search_memory(*, app_name, user_id, query) - no session_id parameter at all. add_events_to_memory's own docstring confirms session_id is optional and implementation-defined ("Implementations may ignore session_id if it is not applicable"). ADK's own guide states the scoping rule in plain language (#570): "Memory is scoped by the (app_name, user_id) pair, so one user never sees another's memories." There is no real tree-walk operation anywhere in BaseMemoryService the way CrewAI's list_scopes/get_scope_info are real one-level walks (ADR-0021) - session_id is a property of where a memory came from, not a retrieval key ADK itself ever searches by.

This means: the scope prefix should be flat, :-joined ({app_name}:{user_id}:), matching LangGraph's, the OpenAI Agents SDK's, and AutoGen's own precedent (ADR-0016/0020/0023) - not CrewAI's /-delimited real hierarchy (ADR-0021), which earns its own path syntax specifically because CrewAI has real nested-scope operations to support. ADK has none, so fighting for path semantics here would be convention for its own sake, the exact trap ADR-0021's own "alternatives considered" section already named and rejected for the opposite reason.

A genuine hybrid: real ids exist for events, optional for direct writes

Two different situations, not one:

• Events, via add_session_to_memory/add_events_to_memory: Event.id is a real, always-present field (#569) - the same situation CrewAI's MemoryRecord.id put ADR-0021 in: a real id already exists, so honor it literally rather than inventing a parallel scheme. Node id: {app_name}:{user_id}:{session_id}: {event.id}.
• Direct writes, via add_memory: MemoryEntry.id is Optional[str] = None (#569), and #570's own research found no real guide or sample that ever populates it - the same situation the OpenAI Agents SDK's transcript items and AutoGen's MemoryContent put ADR-0020/ADR-0023 in: no real id to honor, so mnem-adk needs its own.

Recommendation for ADR-0026

Reuse ADR-0020/ADR-0023's own reserved-sequence-counter idiom for the one real gap (add_memory entries with no id), rather than inventing a sixth idiom for a problem two adapters already solved:

• Scope prefix: {app_name}:{user_id}:, flat and :-joined.
• An ingested event's node id: {app_name}:{user_id}:{session_id}: {event.id} - the real id, honored literally.
• An add_memory entry that already carries MemoryEntry.id: {app_name}:{user_id}:_memory:{memory_entry.id} - still honored literally, placed under a reserved _memory: sub-prefix so a caller's own session_id can never collide with it.
• An add_memory entry with no id: {app_name}:{user_id}: _memory:{seq:010d}, backed by a reserved {app_name}:{user_id}: _memory:_seq counter node, zero-padded for the same reason ADR-0020 zero-pads - BTreeMap lexicographic order then matches insertion order with no scan needed.
• _memory:_seq is excluded from search_memory and any content iteration, the same treatment ADR-0020's _seq and ADR-0016's _meta already get as adapter-internal bookkeeping, not a real memory.

Related: #568 (map), #569 (the real field types this scheme is built on), #570 (the scoping-by-pair finding this directly follows from), #574 (the delete-capability decision, which will need this same scope prefix to address a node by).