fix(mcp): preserve Snowflake ID precision across MCP tool boundary

[lightzt99]

Summary

Closes #1098.

Snowflake IDs (64-bit integers, e.g. 725653563933458432) exceed JavaScript's Number.MAX_SAFE_INTEGER (2^53 − 1 = 9007199254740991). The MCP tool layer declared memory_id as int and serialized id fields as JSON numbers, so V8-based clients (including Claude Code) silently rounded them — e.g. 725653563933458432 → 725653563933458400 — causing update_memory / delete_memory / get_memory_by_id / delete_memory_with_profile to 404 on any memory added in the same MCP session.

The REST API layer already handles this via @field_serializer on memory_id; this PR brings the MCP layer to parity.

Changes

• Input: memory_id: int → memory_id: str in the 4 affected tool signatures. FastMCP now emits "type": "string" in the JSON schema, so clients treat the ID as an opaque string. The function body coerces to int(memory_id) before calling the Python SDK, so the SDK contract is unchanged.
• Output: new _stringify_ids helper recursively converts id / memory_id int fields to strings in MCP responses, applied in _fmt. metadata sub-dicts are passed through unchanged so user-provided numeric fields are not silently rewritten.
• Error path: non-numeric memory_id (e.g. "abc") returns {"success": False, "error": "memory_id must be a numeric string"} instead of raising — mirrors add_memory's error style.

Test plan

New tests/unit/test_mcp_snowflake_id.py (23 tests) covers:

• Schema: 4 tools declare memory_id as "type": "string" (not integer / anyOf); limit stays integer, delete_profile stays boolean.
• Input coercion: each of the 4 tools receives a string Snowflake ID and passes int(memory_id) to the underlying SDK with exact precision (mock assertion).
• Invalid input: "abc" returns {"success": False, ...} and does not call the SDK.
• Output serialization: top-level id, nested results[].id, memory_id key all stringified; non-ID ints (count, limit, threshold) and string IDs (user_id, agent_id) untouched; bool not stringified (it is a subclass of int).
• Metadata skip: {"id": 725..., "metadata": {"id": 42}} → top-level id stringified, metadata.id preserved as int.
• End-to-end: add_memory returns string id; passing it back to update_memory calls memory.update(memory_id=725653563933458432) (mock assertion).
• delete_memory_with_profile response: memory_id key stringified.

.venv/bin/python -m pytest tests/unit/test_mcp_snowflake_id.py -q  # 23 passed

Existing MCP tests (tests/unit/server/test_main_mcp_optional.py, tests/unit/test_optional_dependency_imports.py) continue to pass — no regression in import / FastMCP registration paths.