fix(mcp): preserve assistant messages before tool results

[rksharma-owg]

Summary

• append each assistant message containing tool_calls before its tool results
• preserve tool-call ordering and IDs across successful, failed, and multiple calls
• add isolated async regression tests for the MCP chat-completion loop

Problem

The MCP mock forwarded tool results without first retaining the assistant message that requested them. A follow-up completion therefore received roles user, tool instead of the protocol-valid user, assistant, tool, which compatible OpenAI endpoints reject.

The focused reproducer exercised the real chat_completions function with mocked transports and observed the missing assistant message. No matching issue, pull request, or branch was open when I audited the repository.

Validation

• uv run python -m unittest discover -s tests -v — 3 passed
• uv run black --check app/ client/ tests/
• uv run isort --check-only app/ client/ tests/
• uv run python -m compileall -q app client tests
• git diff --check
• uv run mypy app/ client/ — retains the same 7 pre-existing errors reproduced on pristine main; this patch adds none

The repository's make test target was not run because it pulls the 20B Ollama model and provisions the full Podman stack; the changed loop is covered without external services by the new regression suite.

AI assistance

AI assistance was used during the audit and test drafting. I reproduced the defect against the real function and reviewed and ran every submitted change locally.