fix(opencode): stop sending tools when tool_call is false

[tobwen]

Issue for this PR

Closes #19966
Closes #35432

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

tool_call: false in model config was stored in capabilities.toolcall but never checked in the prompt loop. SessionTools.resolve() ran unconditionally, tools were sent in every LLM request with tool_choice: "auto". Providers that do not support tool calling (e.g. local Ollama models, morphllm without special flags) returned 400.

Three changes in packages/opencode/src/session/prompt.ts:

1. Guard SessionTools.resolve() behind model.capabilities.toolcall - skip tools entirely when false.
2. Set toolChoice: "none" when toolcall is false (the default undefined becomes "auto" at the provider).
3. When format.type === "json_schema" and toolcall is false, fail fast with a clear StructuredOutputError instead of pushing a misleading system prompt and making a doomed LLM call.

How did you verify your code works?

• Read the AI SDK source: prepareToolsAndToolChoice drops both empty tools and toolChoice before the HTTP body is serialized, so no tools or tool_choice key reaches the provider when toolcalling is disabled.
• bun typecheck passes.
• Existing prompt tests pass. Added a regression test under noLLMServer.instance that verifies the loop does not crash with tool_call: false.
• Built a single binary and ran the smoke test.

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR