Dead code path in MCPServer._handle_call_tool and incorrect call_tool return type

[SHAI-Akshay-Tripathi]

Summary

MCPServer._handle_call_tool in src/mcp/server/mcpserver/server.py
contains a dead isinstance(result, dict) branch, with an inline TODO
that already documents both the dead-code issue and the related
incorrect return-type annotation on MCPServer.call_tool.

I'd like to clean this up. Posting the diagnosis here first per
CONTRIBUTING.md.

Diagnosis

FuncMetadata.convert_result in
src/mcp/server/mcpserver/utilities/func_metadata.py can return exactly
three shapes (lines 105-123):

1. CallToolResult (when the tool function returned one directly)
2. Sequence[ContentBlock] (no output schema — via _convert_to_content)
3. (unstructured_content, structured_content) tuple (with output schema)

Never a raw dict. So:

• The isinstance(result, dict) branch in _handle_call_tool (lines
  325-332 on main) is unreachable. It's marked # pragma: no cover
  because tests cannot reach it.
• MCPServer.call_tool's return-type annotation
  Sequence[ContentBlock] | dict[str, Any] (line 402 on main) is wrong:
  it advertises a dict return that cannot happen, and is missing the
  CallToolResult and tuple shapes that can happen.

Proposed fix

1. Remove the dead isinstance(result, dict) branch.
2. Drop the now-unused import json.
3. Change call_tool's return type to
   CallToolResult | Sequence[ContentBlock] | tuple[Sequence[ContentBlock], dict[str, Any]]
   and document all three shapes in the docstring.

The return-type change is a breaking type-signature change for
subclasses, so this targets main (v2) per CONTRIBUTING.md.

I have the patch ready and CI-clean locally (pytest + ruff + pyright).
Happy to open the PR once this is acknowledged.

AI assistance disclosure

I used an AI assistant to help trace the call path and draft the diff.
I've reviewed every changed line, validated the diagnosis against the
source, and run the local checks myself.