InfoRSI/tests/vocab.test.ts

import assert from "node:assert/strict";
import { test } from "node:test";
import { Value } from "typebox/value";
import {
  canonical,
  hash,
  JudgedResult,
  RunDefinition,
  SUBMISSION_KINDS,
  SUBMISSION_SCHEMAS,
  type Checklist,
  type Submission,
  type SubmissionKind,
} from "../src/vocab/index.ts";

// 预期字符串手写；预期哈希用命令行 `printf '%s' '<预期字符串>' | sha256sum` 算出。

test("规范序列化：键的顺序不同，结果相同", () => {
  const expected = '{"a":[true,null,"x"],"b":1,"c":{"y":"中","z":2.5}}';
  const one = { b: 1, a: [true, null, "x"], c: { z: 2.5, y: "中" } };
  const two = { c: { y: "中", z: 2.5 }, a: [true, null, "x"], b: 1 };
  assert.equal(canonical(one), expected);
  assert.equal(canonical(two), expected);
  assert.equal(hash(one), "668eb8c3a097ac33abf5f690ab18c5f753b911ff1c9bbfd6163deb5939b68e5d");
  assert.equal(hash(two), "668eb8c3a097ac33abf5f690ab18c5f753b911ff1c9bbfd6163deb5939b68e5d");
});

test("规范序列化：嵌套数组保持顺序，数组里的对象也排键", () => {
  const value = [{ b: 0, a: [{ d: 1, c: -1 }] }, [1, [2, []]], {}, "s"];
  const expected = '[{"a":[{"c":-1,"d":1}],"b":0},[1,[2,[]]],{},"s"]';
  assert.equal(canonical(value), expected);
  assert.equal(hash(value), "521ca7249ce6c90890b5645dd2f8d46b42f6e000df5ad86aff42415035c7e8cb");
  assert.equal(hash({}), "44136fa355b3678a1146ad16f7e8649e94fb4fc21fe77e8310c060f61caaff8a");
});

test("规范序列化：字符串转义（引号、换行、制表、控制字符、反斜杠、中文）与 -0", () => {
  const value = { "中": [-0], z: -0, s: 'a"b\n\t\u0001中\\' };
  const expected = String.raw`{"s":"a\"b\n\t\u0001中\\","z":0,"中":[0]}`;
  assert.equal(canonical(value), expected);
  assert.equal(hash(value), "d4605ee7ceddf1373f60f837ae86ce24de7319705bbcbce4706309fc4415c79a");
});

test("规范序列化：NaN 和 Infinity 报错，嵌套在里面也报错", () => {
  for (const bad of [NaN, Infinity, -Infinity, { a: [1, NaN] }, [{ x: Infinity }]]) {
    assert.throws(() => canonical(bad), /有限数/);
    assert.throws(() => hash(bad), /有限数/);
  }
});

test("规范序列化：带自定义原型的对象和非 JSON 值报错", () => {
  class Point {
    x = 1;
  }
  const bad: unknown[] = [
    new Point(),
    Object.create(null),
    new Date(0),
    new Map(),
    { a: new Point() },
    { a: undefined },
    () => 1,
    1n,
  ];
  for (const value of bad) assert.throws(() => canonical(value), /纯 JSON/);
});

// 每种提交一份合法样例，同时用来核对字段写的是不是方法.md 第五部分的意思。
const range = { q1: 2.6, q5: 2.62, q9: 2.63 };
const factors = [
  { name: "多起点局部优化", method: "从 8 个随机起点各跑一次 SLSQP，取最好的", keySettings: "起点数 8", continues: null },
  {
    name: "停滞期扰动",
    method: "连续 20 步没有改进时扰动半径",
    keySettings: "幅度 0.004",
    offBehavior: "不扰动",
    continues: "r1-c2.p1",
  },
];
const metrics = {
  defined: [
    {
      id: "first-feasible-seconds",
      meaning: "第一次得到合法摆法的时间",
      unit: "秒",
      direction: "minimize" as const,
      howMeasured: "从进程启动计时，到第一个通过校验的摆法写出为止，每次运行记一次",
      missing: "recordMissing" as const,
    },
  ],
  reused: ["wall-seconds"],
};
const body = "## 为什么可能有效\n……";

const samples: { [K in SubmissionKind]: Submission<K> } = {
  entry: { entries: [{ name: "起点分布", body: body }, { name: "收尾修复", body: body }] },
  discussion: {
    body,
    entries: [
      { name: "a", body },
      { name: "b", body },
      { name: "c", body },
    ],
  },
  claim: { choices: ["researcher-2-e1", "researcher-1-e2"] },
  proposal: { factors, safeguards: ["final-shrink"], metrics, authorEstimate: range, body },
  firstPrediction: { range, body: "理由" },
  critique: { target: "p2", body: "（推测）扰动可能和起点数混杂" },
  response: {
    replies: [{ critique: "cr-1", reply: "改了 p2 的幅度" }],
    changes: { factors: [factors[0]!] },
    clauses: ["在 solve.py 的 main 里加多起点循环"],
    body: "答复",
  },
  unification: {
    factors,
    safeguards: [],
    metrics: { defined: [], reused: [] },
    body,
    replies: [],
    clauses: ["c1 的内容", "c2 的内容"],
    sources: [{ proposal: "r2-p3", coverage: "体现在 p1" }],
  },
  secondPrediction: { range: { q1: 2.6, q5: 2.6, q9: 2.6 }, body: "理由" },
  implementation: {
    factorLocations: [{ factor: "p1", locations: ["solve.py:10-40"] }],
    safeguardLocations: [{ safeguard: "final-shrink", locations: ["solve.py:88", "util.py"] }],
    unassigned: [{ location: "solve.py:5", note: "补了一个 import" }],
    concerns: "无",
    body: "说明",
  },
  repair: {
    factorLocations: [],
    safeguardLocations: [],
    unassigned: [],
    concerns: "无",
    body: "按核验意见补上 c2",
  },
  verification: {
    clauses: [{ clause: "c1", found: true, where: "solve.py:10-40" }],
    factors: [{ factor: "p1", found: false, where: "看了 solve.py 全文" }],
    safeguards: [{ safeguard: "final-shrink", found: true, where: "solve.py:88" }],
    unassigned: [{ location: "solve.py:5", outOfScope: false, reason: "只是 import" }],
    body: "说明",
  },
  explanation: {
    explanations: [
      { body: "p1 在出力", cells: [{ factor: "p1", range }, { factor: "p2", range }] },
      { body: "我没有具体的机制判断", cells: [{ factor: "p1", range: { q1: 2.3, q5: 2.55, q9: 2.64 } }] },
    ],
  },
  ablationImplementation: { alsoOff: [], cannotRun: null, concerns: "无" },
  ablationVerification: { factors: [{ factor: "p1", off: true, where: "solve.py:10-40" }] },
  wrapUp: { body: "被支持的：……" },
};

test("每种提交都有阶段、角色和字段定义", () => {
  assert.deepEqual(Object.keys(SUBMISSION_SCHEMAS).sort(), Object.keys(SUBMISSION_KINDS).sort());
});

for (const kind of Object.keys(SUBMISSION_KINDS) as SubmissionKind[]) {
  test(`合法样例通过字段定义检查：${kind}`, () => {
    const schema = SUBMISSION_SCHEMAS[kind];
    const errors = [...Value.Errors(schema, samples[kind])].map(
      (e) => `${e.instancePath} ${e.message}`,
    );
    assert.deepEqual(errors, []);
  });
}

test("关掉消融时，不带 cells 的解释通过字段定义检查", () => {
  const submission = { explanations: [{ body: "机制：p1 在出力" }] };
  assert.equal(Value.Check(SUBMISSION_SCHEMAS.explanation, submission), true);
});

// 复现：TypeBox 的 Record 对带格式的键只生成 patternProperties，不合格式的键原先完全不检查。
test("环境变量和指标取值里，不合格式的键被拒收", () => {
  const run = {
    command: ["python3", "solve.py"],
    workdir: "/work",
    image: { digest: `sha256:${"a".repeat(64)}`, runtime: "runc" },
    wallSeconds: 120,
    cores: 1,
    inheritEnv: false,
    env: { OMP_NUM_THREADS: "1" },
  };
  assert.equal(Value.Check(RunDefinition, run), true);
  assert.equal(Value.Check(RunDefinition, { ...run, env: { "1BAD-KEY": "x" } }), false);
  assert.equal(Value.Check(RunDefinition, { ...run, env: { "x-y": 5 } }), false);
  const judged = {
    verdict: "valid",
    verdictReason: "合法",
    score: { status: "measured", value: 2.6 },
    metrics: { "wall-seconds": { status: "measured", value: 3 } },
    answerPath: { status: "read", value: "main" },
    safeguardTriggers: { status: "unreadable", reason: "没写" },
    extra: {},
  };
  assert.equal(Value.Check(JudgedResult, judged), true);
  const badMetrics = { "../x": { status: "measured", value: 1 } };
  assert.equal(Value.Check(JudgedResult, { ...judged, metrics: badMetrics }), false);
});

test("对号清单：16 种提交种类每种都有一项，不多不少（类型断言，npm run check 时检查）", () => {
  type Missing = Exclude<SubmissionKind, Checklist["kind"]>;
  type Extra = Exclude<Checklist["kind"], SubmissionKind>;
  const none: [Missing, Extra] extends [never, never] ? true : false = true;
  assert.equal(none, true);
});