ci: pin pydantic in the minimum-mcp job to a known-good version

[DUC750]

Summary

Describe the bridge, schema, export, compatibility, documentation, or test change.

CI-only change, split out of #17 as requested there.

The minimum-mcp job installs mcp==1.12.0 but leaves pydantic unpinned, so pip now resolves pydantic 2.14.1. mcp 1.12.0 imports eval_type_backport from pydantic._internal._typing_extra, which no longer exists in that version. Every test module that imports mcp.server.fastmcp then fails during collection.

• .github/workflows/ci.yml: the minimum-mcp job now installs "pydantic==2.11.10" together with "mcp==1.12.0", with a comment naming this as the known-good combination. The test matrix and pyproject.toml are unchanged.
• I pinned the exact tested version rather than a range, because the first pydantic version that breaks mcp 1.12.0 is not narrowed down.

Validation

• python -m pytest -q -p no:cacheprovider
• python -m ruff check src tests
• Syntax/build checks pass where applicable

Reproduced the job locally on Python 3.11 (Windows, uv venv), from current main:

• pip install -e ".[dev]" "mcp==1.12.0" → pydantic 2.14.1, ImportError: cannot import name 'eval_type_backport', 11 errors during collection.
• pip install -e ".[dev]" "mcp==1.12.0" "pydantic==2.11.10" → 339 passed.

python -m ruff check src tests scripts: all checks passed (no Python file changes). The workflow file parses as YAML.

Data and security

• Fixtures and screenshots use synthetic data only
• No passToken, account ID, database, export, log, or personal health record is included
• Unofficial Xiaomi status and upstream MIT attribution remain intact
• This change does not add a hosted credential proxy or coaching/medical behavior

Licensing

• I agree to the contribution licensing terms in CONTRIBUTING.md (AGPL-3.0-only, plus granting the maintainer the right to offer commercial licenses)

Compatibility

List affected devices, regions, datasets, commands, schemas, or MCP tools.

• Affected: only the minimum-mcp CI job. No devices, regions, datasets, commands, schemas or MCP tools change.
• The declared dependency range (pydantic>=2, mcp>=1.12,<2.0) stays as it is. Users who install exactly mcp 1.12.0 together with a recent pydantic would hit the same import error; that is outside this PR.