fix(mcp): coerce string-encoded metadata, types, and tags in tool calls

[octo-patch]

Fixes #849

Problem

MCP tool bridges (Claude Code, Cursor, and others) sometimes serialize JSON arrays and objects as strings during transport. This causes Pydantic validation failures on retain and recall calls:

Error: 1 validation error for call[retain]
tags
  Input should be a valid list [type=list_type, input_value='["tag1", "tag2"]', input_type=str]

Error: 1 validation error for call[recall]
types
  Input should be a valid list [type=list_type, input_value='["world"]', input_type=str]

Error: 1 validation error for call[retain]
metadata
  Input should be a valid dictionary [type=dict_type, input_value='{"key": "value"}', input_type=str]

The agent retries with the same broken format, wasting tokens and silently failing to store memories.

Solution

Add defensive coercion at two layers — mirroring the pattern already used for MemoryItem.tags (added in #682):

HTTP API (http.py):

• RecallRequest: field_validator on types and tags (mode="before") that parses JSON-string arrays back into lists
• MemoryItem: field_validator on metadata (mode="before") that parses JSON-string objects back into dicts

MCP tools (mcp_tools.py):

• build_content_dict: coerce metadata from JSON string to dict, matching the existing tags coercion in the same function
• Both recall() variants (with and without bank_id param): coerce types and tags from JSON strings before passing to the memory engine

All coercions are backward-compatible — correctly-typed native arrays/objects pass through unchanged.

Testing

The existing coercion pattern for tags in MemoryItem (from #682) already has test coverage. The new validators follow the identical logic and can be validated by sending a JSON-string value to the affected fields.