feat(buzz-agent): add MiniMax as a first-class provider

[octo-patch]

Reason: The buzz-agent runtime shipped dedicated Anthropic, OpenAI-compatible, and Databricks providers but had no MiniMax provider.

Adds MiniMax as a first-class BUZZ_AGENT_PROVIDER in the agent runtime. MiniMax speaks the OpenAI-compatible Chat Completions dialect, so this reuses the existing request path rather than introducing a new body/parse pair — matching the crate's documented "adding a provider" pattern.

Changes

• crates/buzz-agent/src/config.rs
  • New Provider::MiniMax enum variant.
  • resolve_provider accepts the region-selecting spellings minimax (global) and minimax-cn (mainland China); the key is validated in from_env, mirroring the Databricks path.
  • from_env wires MINIMAX_API_KEY (bearer), a MINIMAX_MODEL default of MiniMax-M3, and a region-derived base URL — https://api.minimax.io/v1 for minimax, https://api.minimaxi.com/v1 for minimax-cn — overridable via MINIMAX_BASE_URL. OpenAiApi::Chat is pinned so the auto host heuristic never routes to /responses.
• crates/buzz-agent/src/llm.rs
  • Routes Provider::MiniMax through the OpenAI-compatible Chat Completions request path and the static bearer token source.
• crates/buzz-agent/README.md
  • Documents the provider, its environment variables, and both regional endpoints. Notes that MiniMax's Anthropic-compatible Messages endpoint remains reachable today via BUZZ_AGENT_PROVIDER=anthropic + ANTHROPIC_BASE_URL.

Tests

• resolve_provider_accepts_minimax_region_spellings — every accepted spelling resolves to Provider::MiniMax (case-insensitive).
• minimax_default_base_url_selects_region_by_spelling — global vs mainland-China base-URL selection.
• minimax_routes_over_openai_chat_completions — a MiniMax config POSTs to /v1/chat/completions with the configured model and parses the response.

Checks run

• cargo fmt -p buzz-agent -- --check — clean.
• cargo clippy -p buzz-agent --tests — clean.
• cargo test -p buzz-agent — 302 library tests plus the 3 new tests pass. One pre-existing, unrelated multi-threaded steering test (steer_folds_into_active_turn_without_cancelling) is flaky on this machine and passes on re-run; it does not touch the provider or config code paths.

[Chessing234]

Clean provider add following the documented OpenAI-compatible pattern. Reusing the existing request/parse path for MiniMax keeps the surface area small. README + enum/config wiring look right. One question: does MiniMax need any auth-header or base-URL default distinct from the generic OpenAI-compatible provider, or is BUZZ_AGENT_BASE_URL + key enough for first ship?