Fix(minimax-m3-free): Add MiniMax M3 Free model to opencode/zen

[bitsnaps]

Summary

Fixes for the rejected PR's findings:

• Missing OpenCode model descriptor — src/integrations/models/opencode.ts:611-628
  Added opencode-minimax-m3-free with conservative free-tier metadata (chat-only, no vision, no reasoning,
  131_072/32_768 tokens), matching the existing *-free opencode descriptors. Resolves the dangling modelDescriptorId
  reference from the catalog and unblocks getModelsForGateway('opencode') enrichment.
• Zen model count in opencode.ts:23 — bumped 41 models → 42 models.
• Zen model count in test opencode.test.ts:287 — bumped expect(...).toBe(41) → 42. The cross-reference test at
  line 327 (which was the actual failure mode in the rejected PR) will now also pass.
• User-facing docs — README.md:165 and docs/advanced-setup.md:197 bumped to 42 models.
• Generated artifact — src/integrations/generated/integrationArtifacts.generated.ts:326 hand-synced to 42 models.

[jatmn]

I found an issue that needs to be addressed before this is ready.

Findings

• [P1] Route MiniMax M3 Free through the correct OpenCode metadata and endpoint
  src/integrations/gateways/opencode.ts:77
  The newly added minimax-m3-free catalog entry falls through to the default /chat/completions OpenAI-compatible path, and its descriptor in src/integrations/models/opencode.ts marks the model as non-reasoning, non-vision, and 131k context. OpenCode's current model catalog for minimax-m3-free lists it as an Anthropic-backed model with reasoning enabled, text/image/video input, 200k context, and a 32k output limit. As written, selecting this model will use the wrong transport shape and OpenClaude will hide or under-budget the capabilities users need for the model. Please wire the catalog entry the same way the other OpenCode Anthropic-message models are routed and update the descriptor metadata to match the provider catalog.