feat(deepseek): use responses api for v4 flash

[MrMushrooooom]

Summary

• mark direct DeepSeek V4 Flash as a Responses API model
• apply the same shape to OpenCode Zen, Zen Free, and Go catalog entries
• cite DeepSeek's official Responses API documentation

Context

DeepSeek documents V4 Flash as the only currently supported Responses API model: https://api-docs.deepseek.com/guides/responses_api

Requires anomalyco/opencode#40011 to consume the per-model shape.

Testing

• bun validate
• git diff --check
• bun test: 131 pass; existing unrelated failures remain for missing open-weight links and ungenerated SDK snapshot/effect artifacts

[github-actions]

Action items

• [medium] [possible mistake] providers/opencode/models/deepseek-v4-flash.toml:16 - Check: provider.shape must describe the inference surface of that catalog provider, not only the upstream model. Why: These OpenCode Zen/Go entries set shape = "responses" and cite only DeepSeek’s native Responses guide. That does not establish Zen/Go routing. providers/opencode/provider.toml documents /zen/v1/responses for OpenAI and /zen/v1/chat/completions for compatible models; these DeepSeek rows still use the provider default @ai-sdk/openai-compatible (no npm override), unlike OpenCode GPT rows that select the OpenAI surface with npm = "@ai-sdk/openai". Applying Responses here can mis-route clients once shape is consumed. Action: Verify OpenCode Zen, Zen Free, and Go actually expose deepseek-v4-flash on a Responses endpoint; cite that OpenCode surface (and align npm/endpoint notes if needed). If they still use chat-completions, drop shape = "responses" from the three OpenCode catalog files and keep it only on direct DeepSeek if that native claim holds.