Add MiniMax M3.1 Flash preset and thinking controls

[octo-patch]

Reason: MiniMax's current direct API text lineup is MiniMax-M3.1-Flash and MiniMax-M3. The preset still offered M2.7, and OpenCode's inherited M3 none variant would send an invalid disabled-thinking request for Flash.

This makes Flash the first installation choice and recommends the two current models, without migrating saved or custom model IDs. Both presets carry the documented 1M context and 524,288 output limits. Flash uses adaptive thinking and low/medium/high/xhigh/max effort (default max), with incompatible inherited variants disabled. M3 keeps adaptive/disabled thinking and the protocol's default, without effort. The OpenAI-compatible SDK maps reasoningEffort to reasoning_effort; Anthropic maps effort to output_config.effort.

The settings display uses Flash's $0.15/$0.60/$0.015 input/output/cache-read rates and M3's Standard <=512k baseline of $0.30/$1.20/$0.06 per million tokens, with its higher-input and Priority pricing explained beside the model. Unknown cache-write pricing remains unset. No media assets or non-MiniMax providers are changed.

Validation: 67 tests passed across MiniMax preset, provider-engine adapter, and shared provider preferences suites; app TypeScript check passed; maintainability audit and git diff --check passed. An independent capture probe using the pinned OpenCode SDK versions (@ai-sdk/anthropic 3.0.111 and @ai-sdk/openai-compatible 2.0.41) verified 36 model/variant requests across all four endpoint choices, including the exact effort/thinking wire fields. No live inference was performed. The frozen dependency install populated dependencies but returned the existing unapproved lefthook build-policy error; the installed checker was invoked directly, and the installer-added placeholder was removed.

CI: the Ubuntu and macOS test jobs and design-studio build terminate during dependency installation with ERR_PNPM_IGNORED_BUILDS for lefthook@2.1.6, before tests run. This matches the local frozen-install limitation; the repository's dependency build-approval policy is unchanged. Business-components and i18n-audit passed. Evidence: Ubuntu test job, macOS test job, design-studio job.

Limits: OpenCode's flat cost schema cannot calculate request-size or Priority billing; the UI states those tiers. OpenCode may apply its own runtime output cap below the model's documented maximum. The repository mentions a fraimz skill that is absent from this checkout; its available maintainability skill was applied. Existing image/video/voice PRs and the auth-curation PR are separate work.

Sources:

• https://platform.minimax.io/docs/guides/models-intro.md
• https://platform.minimax.io/docs/pricing/overview.md
• https://platform.minimax.io/docs/api-reference/text-chat-openai.md
• https://platform.minimax.io/docs/api-reference/text-chat-anthropic.md