Fix: Add JSON string deserialization support for metadata parameter

[MyWWWChat]

Problem

The MCP server's metadata parameter schema doesn't match its implementation, causing validation errors when Claude (and other MCP clients) send metadata as JSON strings.

Current Issue:

• Schema: Only accepts object | null for metadata
• Implementation: Supports dict | str | None (can parse JSON strings)
• Result: Input validation error: '{"code": "..."}' is not valid under any of the given schemas

Solution

This PR fixes the schema-implementation mismatch by:

1. Updated type annotation from Metadata | None to Union[Metadata, str, None]
2. Added JSON deserialization logic with proper error handling
3. Added validation to ensure metadata is a dictionary after deserialization
4. Added comprehensive test suite with 13 test cases covering all scenarios

Technical Changes

src/mcp_server_qdrant/mcp_server.py

• Changed metadata type annotation to include str type
• Added JSON deserialization for string metadata (lines 121-128)
• Added type validation after deserialization (lines 131-132)
• Handles edge cases: "null", "none", empty strings, whitespace

src/mcp_server_qdrant/qdrant.py

• Added fallback to 'text' field when 'document' is missing in search results
• Prevents KeyError when retrieving search results with different payload structures

tests/test_metadata_deserialization.py

• New comprehensive test suite with 13 test cases
• Tests JSON strings, dictionaries, None, edge cases
• Tests error handling for invalid JSON and non-dict types

Testing

All tests pass:

• ✅ 13 new tests for metadata deserialization
• ✅ 37 existing tests continue to pass
• ✅ Backward compatibility maintained

Impact

This fix resolves the metadata validation error that prevents Claude and other MCP clients from using the metadata parameter as documented. The change is backward compatible - existing code using dict or None continues to work.

Related Issues

Fixes the error: Input validation error: '{"code": "def test(): pass"}' is not valid under any of the given schemas

This enables proper usage as described in TOOL_STORE_DESCRIPTION:

"Store code snippets with descriptions. The 'information' parameter should contain a natural language description of what the code does, while the actual code should be included in the 'metadata' parameter as a 'code' property."

[shigomany]

This PR fixes a critical schema-implementation mismatch that currently breaks the qdrant-store tool for all Claude (and other MCP) users.

The Problem: The MCP server's metadata parameter schema doesn't match its implementation. The schema only accepts object | null, but the implementation actually supports dict | str | None (JSON strings). This causes validation errors when Claude sends metadata as JSON strings:

Input validation error: '{"code": "..."}' is not valid under any of the given schemas
Why This Matters: According to the TOOL_STORE_DESCRIPTION documentation, code should be stored with: