feat: optimize GPT-5.6 defaults and reasoning

[aheritier]

Summary

• add GPT-5.6 Sol, Terra, and Luna support and refresh the embedded models.dev catalog
• support the GPT-5.6 reasoning ladder (none, low, medium, high, xhigh, max) while preserving older-model behavior
• update OpenAI, ChatGPT, Vercel, builtin-agent defaults, examples, schema guidance, and provider documentation
• propagate trusted OpenAI-vendor identity across native and JS/WASM provider factories, including adversarial and clone-path coverage

Validation

• task build
• task test
• golangci-lint run (0 issues)
• GOOS=js GOARCH=wasm go test -exec=$(go env GOROOT)/lib/wasm/go_js_wasm_exec ./pkg/model/provider/...
• git diff --check

The repository custom linter still reports the pre-existing Lint/DeferMutexUnlock warning in untouched pkg/tools/mcp/session_client.go.

Direct OpenAI and ChatGPT examples and defaults use the unsuffixed gpt-5.6 alias. Vercel remains pinned to openai/gpt-5.6-sol because its catalog exposes explicit variants rather than the unsuffixed alias.

References

• OpenAI: Using GPT-5.6 — documents the GPT-5.6 family, model selection, and that the gpt-5.6 alias routes to gpt-5.6-sol.
• OpenAI: Reasoning models — reasoning effort and Responses API guidance.
• OpenAI model reference: GPT-5.6
• OpenAI model reference: GPT-5.6 Sol
• OpenAI model reference: GPT-5.6 Terra
• OpenAI model reference: GPT-5.6 Luna
• OpenAI Codex models — ChatGPT/Codex model availability.
• models.dev: GPT-5.6 Sol
• models.dev: GPT-5.6 Terra
• models.dev: GPT-5.6 Luna

[aheritier]

👋 This PR has merge conflicts with the base branch. Please rebase or merge the latest base branch and resolve them. I've moved it to draft and added status/needs-rebase; it'll be picked back up automatically once the conflicts are cleared.

[Sayt-0]

I'm not sure that gpt 5.6 has a max reasoning effort according to the docs : https://developers.openai.com/api/docs/guides/reasoning

[aheritier]

Good catch—the generic reasoning guide is incomplete here. It says the accepted values are model-dependent, but its generic list currently stops at xhigh.

The GPT-5.6-specific guide explicitly documents max:

• “GPT-5.6 supports max reasoning effort”
• Supported efforts are none, low, medium, high, xhigh, and max

https://developers.openai.com/api/docs/guides/latest-model

The implementation gates max to GPT-5.6+ only, so older OpenAI models won’t expose it.