✨ feat(llm): 引入基于完成度的 AI 提交指导与限流

- 新增 llm_guidance 策略层，根据源码、判题结果与通过率评估完成度分级
- llm-review.php 改为服务端缓冲上游输出，仅下发白名单分类的固定模板
- 新增 llm_review 与 llm_review_rate_limit 表，实现结果缓存与账户/提交两级限流
- ceinfo 模板渲染编译器标出的源码行号/列号，reinfo 仅向高完成度提交展示结构化 diff
- download.php 按完成度授权测试点下载，封堵白卷提交套取隐藏用例
- judge_client 在 ACM 非比赛场景下记录已通过测试点数，前端可显示进度
- 附 llm_guidance_policy_test 单测覆盖空白卷、注释卷、隐藏 SPJ、OJ_FULL_DIFF 等关键路径

core/judge_client/judge_client.cc

@@ -3748,6 +3748,7 @@ int main(int argc, char **argv)
 	
 	// read files and run
 	double pass_rate = 0.0;
+	int passed_test_count = 0;
 	float mark=0;
 	float total_mark=0,get_mark=0;
 	int finalACflg = ACflg;
@@ -3860,6 +3861,10 @@ int main(int argc, char **argv)
 			judge_solution(ACflg, usedtime, time_lmt, spj, p_id, infile,
 						   outfile, userfile, PEflg, lang, work_dir, topmemory,
 						   mem_lmt, solution_id, num_of_test,&pass_rate);
+			if (ACflg == OJ_AC)
+			{
+				++passed_test_count;
+			}
 			/*
    			if(usedtime > time_lmt * 1000) {          // 如果觉得的显示超时结果的计时过长，可以覆盖数据。
 					usedtime = time_lmt * 1000;
@@ -3968,6 +3973,10 @@ int main(int argc, char **argv)
 		else
 		{
 			if(ACflg==OJ_AC) pass_rate=1.0;
+			// In ACM/non-OI mode expose prefix progress only for ordinary
+			// practice submissions. Contest submissions keep the historical
+			// all-or-nothing value so hidden test progress is not leaked.
+			else if(cid==0 && num_of_test>0) pass_rate=(double)passed_test_count/num_of_test;
 			else pass_rate=0.0;
 			update_solution(solution_id, ACflg, usedtime, topmemory >> 10, sim,
 							sim_s_id, pass_rate);