feat(chat): generate replies and content in the user's language

[abdulrafey1]

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

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

What

Makes the chat plugin write replies and generated course content in the user's stored preferred
language, instead of leaving the language to the model's guess.

Changes

• feat(plugins): get_learning_design_system_prompt takes a resolved BCP 47 tag and names the
  output language explicitly in the system prompt
• feat(plugins): resolve the tag once per request from the signed-in user's stored preference
• fix(plugins): replace the ambiguous "Write in the user's language" instruction
• test(plugins): prompt rendering per supported tag, a regression guard on the removed sentence,
  and route-level tests that the stored preference reaches the provider and re-resolves per request
• docs(core): correct the config comment and the user guide, both of which said generated content
  ignores the preference

How to Test

1. uv run pytest sparkth/plugins/chat/tests/test_prompt.py sparkth/plugins/chat/tests/test_completion_language.py -v
2. PATCH /api/v1/user/me with {"language": "es"}, then start a chat and ask for a course
   outline in English — the reply and the outline come back in Spanish.
3. Change the preference to fr mid-conversation and send another message: the next turn is in
   French, and earlier messages are left as they were.
4. Clear the preference with {"language": null} and confirm replies fall back to
   DEFAULT_LANGUAGE.

Notes

No migration, no env var, no dependency.

The system prompt is rendered fresh on every request, so language re-resolves per turn. That is
intended: changing the setting applies from the next message onward and no conversation-level
pinning exists.

The template hands the model the refusal sentence and tells it to send that sentence verbatim, so
the new language directive carves it out explicitly. Without the carve-out the prompt contradicts
itself and the model's refusal would drift away from the deterministic refusal streamed on the
out-of-scope path.

Manual smoke test — not automatable, and the one thing the suite cannot cover. Every test here
mocks the LLM, so they prove the language reaches the prompt, not that the output is good:

1. Generate a full course in es and in fr. Have a speaker check register and terminology, and
   confirm it does not read as translated English.
2. Generate a course from an English source document into the target language — the highest-value
   case, and the one the directive's "regardless of the language of any uploaded source documents"
   clause exists for.
3. Confirm assessment questions and their answer options and feedback are all translated.
   Partial translation of quiz options is a classic LLM failure.
4. With a non-English preference set, ask explicitly "translate this into Spanish" (PR 1 unblocks
   that phrasing) and confirm the explicit request and the standing language directive interact
   sensibly rather than fighting.

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