MCPServer completion handler returning more than 100 values fails on 2026-07-28 sessions and violates the 100-item limit on 2025-11-25 sessions

Finding: In the MCP Python SDK, an MCPServer @mcp.completion() handler that returns more than 100 values is reported to behave inconsistently by protocol version. On a 2026-07-28 session the whole completion/complete request fails with a generic error, and on a 2025-11-25 session the oversize list is sent without pagination hints. The environment was mcp 2.3.0 and main @ 91941ed, Python 3.12.12/3.14.0 and Windows 11 (https://github.com/modelcontextprotocol/python-sdk/issues/3649). Reported symptoms: the issue author says the 2026-07-28 session returns an MCPError, 'Handler returned an invalid result' (JSON-RPC -32603), so the user gets no suggestions. On 2025-11-25, their script printed 150 values with total=None and has_more=None. The author's server log shows a pydantic too_long validation error for CompleteResult.completion.values (at most 100 items, got 150). The issue notes this is easy to trigger with the prefix-filter pattern from the docs, because an empty typed value matches every candidate (https://github.com/modelcontextprotocol/python-sdk/issues/3649). Reported cause (author analysis, not maintainer-confirmed): the 100-item cap exists only in the 2026-07-28 wire model. The runner validates handler results against it and converts the ValidationError into an INTERNAL_ERROR. The 2025-11-25 model lacks the constraint, and the MCPServer.completion wrapper passes the Completion through unchanged. The author also cites the TypeScript SDK as truncating to 100 and setting total and hasMore (https://github.com/modelcontextprotocol/python-sdk/issues/3649). Proposed mitigation (unverified): cap values at the first 100 in the MCPServer.completion wrapper, set total to the full count and has_more to True unless the handler set them explicitly, and leave the lowlevel Server untouched. A commenter restated this approach (https://github.com/modelcontextprotocol/python-sdk/issues/3649#issuecomment-6067800188) and later said they opened PR #3660 with regression tests for both modes (https://github.com/modelcontextprotocol/python-sdk/issues/3649#issuecomment-6068723043). Status: open, with no maintainer response in the supplied thread. No merged fix or released version is evidenced, and the PR was not supplied, so its content and state are unknown. Limitations: this is an external report and was not reproduced by KFA. The reproduction output is the author's own, and the author disclosed AI assistance. Cause, intended behavior and fix remain unconfirmed by maintainers.