fix: 主线 B3/B4 结构化工具调用与完成前验证

- 打通 Worker → WorkerManager → Provider 的 tools 传递链路，ProviderManager/adapter
  返回结构化 tool_calls 给 WorkerRuntime
- OpenAI-compatible/Anthropic adapter 发送工具 schema，并解析 provider 返回的
  tool_calls/tool_use；OpenAI 工具名使用 fs.write ↔ fs__write 双向映射
- 修复独立复审发现的 OpenAI 协议隐患：assistant tool_use blocks 必须转换为
  assistant.tool_calls，后续 role=tool 消息的 tool_call_id 必须匹配前一轮
  tool_calls[].id；不再把 tool_use JSON 字符串化为普通文本
- ExecutorRole 优先消费原生 tool_calls，回灌 canonical tool_result block；移除
  fs.write(...)/shell.run(...) 函数调用正则解析，只保留严格 JSON tool_call
  fallback 与 filename code block 兼容
- DONE 前执行 verification-before-completion：任务要求 build/compile/run/test/编译/
  运行/测试时必须实际 shell.run 验证，失败不 checkpoint、不返回 completed
- fs.write 覆盖已有文件也强制 read-before-write，补齐 Claude Code 文件状态纪律
- 新增 packages/workers/test/executor-role.test.ts 行为测试：原生 tool_calls 执行、
  verification 失败不得 completed

真实验收:
- TSC=0
- bun test packages/workers/test/executor-role.test.ts: 2 pass / 0 fail
- OpenAI converter 探针确认 assistant.tool_calls 与 role=tool 的 tool_call_id 匹配
- 真实 GLM Worker C++ 编译运行任务通过，worker verification 记录实际命令:
  c++ hello.cpp -o /tmp/aircoding-verify && /tmp/aircoding-verify

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

packages/workers/src/WorkerRuntime.ts

@@ -35,7 +35,7 @@ export interface LLMRequest {
 export interface LLMResponse {
   content: string
   usage?: { input_tokens: number; output_tokens: number }
-  tool_calls?: Array<{ name: string; arguments: Record<string, unknown> }>
+  tool_calls?: Array<{ id?: string; name: string; arguments: Record<string, unknown> }>
 }
 
 export class WorkerRuntime {