[octo-patch]

Reason: the CLI can persist model profiles, but has no MiniMax preset and sends text requests and API-key validation to DashScope-specific paths. Add opt-in minimax and minimax-cn profiles so a fresh MiniMax login can actually run text requests with MiniMax-M3.1-Flash.

• Validate keys only against the selected official HTTPS origin's /v1/models, and retain the selected profile regardless of Aliyun-style key prefixes.
• Route Chat Completions and Responses through their /v1 endpoints for those origins; keep existing transport, streaming and response handling.
• Preserve saved and explicit models, including MiniMax-M3, MiniMax-M2.7 and custom IDs. MiniMax presets claim only text.chat and do not materialize media defaults.
• Map explicit reasoning requests to MiniMax's thinking / reasoning fields and reject unsupported thinking budgets. Add English/Chinese command help and login examples.

Validation: 53 focused tests passed (new MiniMax tests plus existing API-key and config-priority tests). Tests exercise actual Client HTTP serialization, bearer headers, both regions, both APIs, streaming/non-streaming, fresh defaults, saved/custom/legacy models, and unchanged non-MiniMax routing. Scoped formatting, lint and type checks passed for all nine changed files; git diff --check passed.

Validation limits: no dependency installation or paid inference was performed. The pinned Vite+ 0.3.3 was unavailable in the local cache; focused tests used cached Vitest 2.1.9 with a test-import alias, and scoped static checks used cached Vite+ 1.1.0 and Node types 24.8.1. Runtime dependencies were linked from existing caches. The full pinned build/e2e suite was not run. GitHub CI for commit 607c99e7b5a4e9760352ee268a8034e0c7c67a6f currently reports action_required, with no check runs yet.

API references: OpenAI-compatible text, models endpoint, Responses, China Responses.