测试数据 / Test data¶
生成器仅依赖Python标准库；固定配方和种子见 plan.json。 The generator uses only the Python standard library; plan.json fixes all recipes and seeds.
python tests/generate.py --out-dir build/data
python tests/generate.py --plan tests/plan.json --out-dir build/data --case 01 --case 02 --std ./std
在单题目录中运行。--case 可以重复；未知编号会报错。--std 指定已编译程序并生成对应输出。
Run inside the problem directory. Repeat --case to select inputs; unknown IDs are rejected. --std runs the compiled program to produce outputs.
小规模独立模型、输入校验、正式数据复现和五轮测量分别见仓库工具 tools/check_r34.py 和 tools/benchmark_r34.py .公开性能记录见 benchmark-R34.json。正式生成文件不入库，样例文件入库。
Independent small models and validators are in repository tool tools/check_r34.py ; reproduction and five-run measurements are in tools/benchmark_r34.py . See benchmark-R34.json. Formal files are regenerated locally; sample files are versioned.
本地交互 / Local interaction¶
g++ -std=c++17 -O2 std.cpp -o std
python judge.py --input tests/samples/02.in --program ./std --transcript trace.json
g++ -std=c++17 -O2 simulate.cpp -o simulate
隐藏文件先给n，再给排列。本地交互器读取隐藏文件，提交程序只收到n和逐次回复。--std 通过真实交互验证，并把恢复的排列保存为 .out .样例输出也是该排列，不能当作交互协议的完整输出记录。trace.json 保存实际询问和回复。
Hidden fixtures contain n followed by the permutation. The judge reads them; the submission receives only n and replies. --std validates real interaction and writes the recovered permutation to .out . Sample outputs have this same fixture format, not a complete protocol transcript. trace.json records actual queries and replies.
simulate.cpp 直接读取隐藏文件，只是算法验证工具，输出恢复的排列及下一行的询问次数。它与提交程序共享 solver.hpp .不要将模拟器提交为交互解答。
simulate.cpp reads a hidden fixture directly and prints the recovered permutation, then the query count on the next line. It is a testing harness sharing solver.hpp , not an interactive submission.
交互器会拒绝非法查询、超过800次询问、错误排列、多余输出、异常退出和超时。默认30秒是本地诊断超时；正式时限见题面及测量。 The judge rejects invalid queries, more than 800 queries, wrong permutations, extra output, unsuccessful exit, and timeout. Its default 30 seconds is a diagnostic timeout; use the statement and measurements for the evaluation limit.