[octo-patch]

Reason: SQLBot's built-in MiniMax picker still offers M3 and M2.7. Offer MiniMax-M3.1-Flash first and MiniMax-M3 second for new model selection.

Changes

• Refresh only the MiniMax text model options in the supplier registry.
• Update the existing supplier guard to assert the exact model list and order.
• Keep custom model entry and saved model IDs unchanged. The existing Chat Completions endpoint, parameters, and transport remain unchanged.

Validation

• Existing supplier suite: 22 tests passed; the updated catalog assertion failed against the original options and passed after the change.
• Ruff check and format check passed for the changed Python test; git diff --check passed.
• Offline HTTP transport validation used SQLBot's real BaseChatOpenAI and the production OpenAILLM/LLMConfig classes extracted without database imports, with real LangChain/OpenAI SDKs and httpx.MockTransport. Sixteen model/endpoint/stream combinations passed, covering Flash, M3, saved M2.7 and custom IDs, both configured endpoint variants, image and tool history, streamed reasoning, and explicit Flash reasoning effort. This checks serialization only, not regional service availability.
• Cached validation dependencies: langchain-openai 0.3.28, langchain-core 0.3.86, openai 1.109.1, httpx 0.28.1. Cached Python is 3.12.13 rather than the project's required 3.11; cached pytest is 8.4.2 rather than the project's <8 range.

Limitations

The frontend build, full backend environment, database startup, browser flow, and live inference were not run. Frontend dependencies are absent, and heavy setup/builds are paused by the workspace resource gate (approximately 17 GiB free disk, below its 110 GiB threshold). No paid inference or real provider credentials were used.