feat(llm): pass strict through tool definitions for Codex parity

[rekram1-node]

Summary

Codex CLI hardcodes strict: false on every function tool sent to the OpenAI Responses API (codex-rs/tools/src/responses_api.rs:131). This ensures MCP-sourced and dynamic schemas that don't satisfy OpenAI's structured-outputs constraints still register instead of being rejected.

This PR adds first-class strict support to the @opencode-ai/llm tool model and injects strict: false from opencode for OpenAI Responses-family providers to match Codex behavior.

Changes

@opencode-ai/llm (generic library — no policy decisions)

• schema/messages.ts: add strict?: boolean to ToolDefinition
• tool.ts: accept strict on Tool.make() (typed + dynamic configs), carry it through toDefinitions() onto the ToolDefinition
• protocols/openai-responses.ts: lowerTool passes tool.strict through conditionally — ...(tool.strict != null ? { strict: tool.strict } : {}) — same pattern as the AI SDK

opencode (owns the Codex-parity policy)

• session/llm/request.ts (AI SDK path): injects strict: false on every tool when model.api.npm is @ai-sdk/openai, @ai-sdk/azure, or @ai-sdk/amazon-bedrock/mantle — the same provider set already used for Responses item-ID stripping at transform.ts:466
• session/llm/native-runtime.ts (native path): passes strict: false via NativeTool.make() when model.api.npm === "@ai-sdk/openai" (the only provider the native gate supports)

Why not hardcode in the llm package

The llm package is a generic library. Codex-parity policy ("always send strict: false") belongs in opencode, not in the library. The library's job is to carry strict faithfully from tool construction to the wire, which is what this does.

Verification

• bun typecheck clean for both packages/llm and packages/opencode
• bun test test/provider/openai-responses.test.ts — 51 pass
• bun test test/provider/openai-responses-cache.recorded.test.ts — 1 pass
• bun test test/session/llm-native.test.ts — 16 pass