fix(runtime/provider/tui): advisory gate, request reuse, and gateway fail-fast

[fennoai]

Requested by @phantom5099

Summary

• downgrade OpenAI-compatible/Anthropic/Gemini local estimate gate policy to advisory
• reuse prepared provider requests between estimate and generate to avoid duplicate request builds
• make TUI gateway stream fail-fast on runtime payload version mismatch to prevent repeated error events

[phantom5099]

@xgopilot 给“估算失败”加降级通道。
现在最该补的是：EstimateInputTokens() 失败时，不要直接结束 run，而是发一条诊断事件后跳过 proactive budget gate，继续进入 Generate()。
只有“配置非法、模型非法、driver 非法”这类确定不可恢复错误才继续硬失败。这样才能真正消掉“估算阶段单点故障”

[fennoai]

codeagent is processing your request, please wait a moment...