feat(mcp): accept a language for course generation

[abdulrafey1]

Part of: Sparkth UI, emails, API errors, and AI-generated content are English-only

Fourth of the phase-2 generation-language stack. Does not close the issue.

What

Lets an MCP caller name the language a generated course should be written in, and makes the
course-generation prompt state that language explicitly.

Changes

• feat(mcp): optional language BCP 47 tag on CourseGenerationPromptRequest, resolved at the
  tool boundary
• feat(mcp): the course-generation prompt names the output language and covers titles, lesson
  text, assessment questions, answer options and feedback
• fix(mcp): remove "in the user's language" from the prompt — undefined on a path with no user
• test(mcp): explicit, omitted and unsupported tags, each with a negative control, plus the
  published tool schema

How to Test

1. uv run pytest tests/mcp/ -v — 19 tests pass.
2. Call the tool over MCP with {"course_params": {"course_name": "...", "course_description": "...", "language": "es"}}
   and confirm the returned prompt names Spanish.
3. Omit language, and separately pass a nonsense tag such as "klingon" — both return a prompt
   naming DEFAULT_LANGUAGE, and neither errors.
4. Inspect the tool's published schema and confirm language is optional and its description
   states the fallback.

Notes

No migration, no env var, no dependency. Public MCP tool signature change — additive and
backward compatible: existing callers that omit language get the platform default.

The MCP server has an audit middleware but no identity layer, so there is no user to resolve a
stored preference from — the calling agent supplies the tag. This is why the two generation paths
are not symmetric.

The tag is validated by falling back, not by erroring: an omitted, misspelled or withdrawn tag
resolves to DEFAULT_LANGUAGE, so a bad value never fails a whole agent-driven generation run.
The published field description promises exactly that, and a test pins the promise.

The directive deliberately covers the course only, not the calling agent's own clarifying
questions. language is the course language and says nothing about what language the agent's
human speaks — an English speaker may commission a Spanish course, and the agent should keep
asking its questions in the language of its own conversation.

frontend/lib/api/generated.ts is unchanged: the MCP tool schema is not part of the REST OpenAPI
document, and make test.frontend.api reports no drift.

Manual smoke test. The tests mock nothing about the model's actual output, so before relying
on this: generate a full course through the MCP path in es and in fr, have a speaker review
register and terminology, and confirm assessment questions, answer options and feedback are all
translated.

This description was written with the assistance of an LLM (Claude).