Strip non-schema message fields for Mistral chat requests

[jvmenen]

Summary

• Mistral's chat completions API uses strict Pydantic validation (extra="forbid") and rejects any field on a message object that is not part of its defined schema.
• Space Agent's prompt pipeline attaches internal bookkeeping fields such as tokenCount, kind, and source to message objects for prompt budgeting and UI display, which causes Mistral to reject the request.
• This PR adds a headless _core/mistral module with shared endpoint detection and a body-rewrite helper, wired through extension hooks on prepareOnscreenAgentApiRequest and prepareAdminAgentApiRequest. Non-Mistral requests are left untouched.

Design notes

• Provider detection runs against the configured upstream API endpoint (not the proxied fetch URL), since frontend fetches may be rerouted through /api/proxy.
• The body rewrite preserves only the message fields accepted by Mistral's schema: role, content, name, tool_calls, tool_call_id. If Mistral expands the accepted schema later, extend ALLOWED_MESSAGE_FIELDS in request.js rather than removing the strip.
• Provider-specific HTTP policy is kept inside this module instead of branching inside _core/onscreen_agent/llm.js or _core/admin/views/agent/api.js, so the admin and onscreen surfaces stay in sync via one shared helper.
• AGENTS.md documents ownership and contracts for the new module.

Test plan

• Send a chat request from the onscreen agent against a Mistral endpoint and confirm the request succeeds (previously rejected with a 422 from Pydantic validation).
• Send a chat request from the admin agent view against a Mistral endpoint and confirm the same.
• Send a request against a non-Mistral endpoint (e.g. OpenAI / OpenRouter) and confirm the request body is unchanged.
• Verify in the desktop build that the bundled app/**/* assets include the new _core/mistral module after repackaging.