eval: 完整研究的运行、统计与审阅导出（任务书 17 第三块的检验代码）

- eval/full/run.ts：配置照 eval/round-zero/research.json 只换编号，真实跑到停下； launches/machine/timeline/outcome 照任务书 16 惯例；--kill-after-first-ablation-start 在第一个消融运行确认启动后、结果落账前 SIGKILL 控制进程，恢复续跑到底， kill-report.json 写三条检查（同一容器不重跑、只扣一次、结果照常写入）。 - eval/lib/full.ts：账目核对 fullAccounts——逐轮走向（候选分数、采纳、选格、每格核验重做次数与扣次、结算、押的因子）加 12 项程序账目检查（阶段开关成对、扣次=确认启动数、余额为 0、解释冻结早于首格启动、每轮最多一条结算、解封在收尾关闭之后且覆盖全部选中候选、跑不起来/不跑的格零扣次、第 2 轮起基线链重算、停在"导出还没写"等）与 CPU/墙钟比（<0.9 列出并附同时段负载，只列事实）。 - eval/full/stats.ts、export.ts：统计输出与按轮按格的审阅材料导出。 - 不沿用 eval/forward/acceptance.ts 的"启动区间互不重叠"（这次 3 份同时跑）。 - tests/full-eval.test.ts 4 条：真实账本 + rules 真实计算走完一轮含消融的研究，撑几成与配合关系标签按方法.md 6.6 手算对照。npm run check 通过，npm test 476 条通过。
2026-10-04 14:09:09 -04:00
parent 144b88480b
commit 0ae495f8e8
6 changed files with 1606 additions and 1 deletions
@@ -0,0 +1,273 @@
+import { mkdirSync, writeFileSync } from "node:fs";
+import { join, resolve } from "node:path";
+import {
+    Board,
+    cellId,
+    explanationId,
+    protocolMetrics,
+    splitCell,
+    type TaskMaterial,
+} from "../../src/materials/index.ts";
+import { ablationOutcome } from "../../src/submissions/index.ts";
+import { checkoutVersion, compareVersions } from "../../src/workspace/index.ts";
+import { jsonFile } from "../lib/forward.ts";
+import { finalSubmissions, ofKind, readLedger } from "../lib/ledger.ts";
+
+// 任务书 17 第 3.2 节研究质量评估的审阅材料：解释、选格决定、消融格（协议、位置表、代码与 diff、实现与核验往来）、
+// 结算、收尾、下一轮提案。只导出事实和原文，不替审阅者下结论。
+// node eval/full/export.ts <研究输出目录> <全新绝对审阅目录>
+const [dir, output, ...extra] = process.argv.slice(2);
+if (!dir || !output || extra.length || output !== resolve(output)) {
+    throw new Error(
+        "用法：node eval/full/export.ts <研究输出目录> <全新绝对审阅目录>",
+    );
+}
+mkdirSync(output); // 不覆盖以前的审阅材料。
+const study = join(dir, "study");
+const records = readLedger(study);
+const board = new Board(records);
+const repo = join(study, "repo.git");
+const finals = finalSubmissions(records);
+const roundOf = (stageId: string) => board.stage(stageId).round;
+const roundIds = [
+    ...new Set(ofKind(records, "decision").map((d) => d.content.round)),
+]
+    .filter((r) => r >= 1)
+    .sort((a, b) => a - b);
+
+const checkout = (version: string, path: string) => {
+    mkdirSync(path);
+    checkoutVersion(repo, version, path);
+};
+const json = (v: unknown) => `${JSON.stringify(v, null, 2)}\n`;
+
+const index = [];
+for (const round of roundIds) {
+    const roundDir = join(output, `round-${round}`);
+    mkdirSync(roundDir);
+
+    // 本轮提案（下一轮提案在 round-<r+1> 里，审阅"收尾→接力"时把两份对着读）。
+    const proposals = finals.filter(
+        (s) =>
+            s.content.kind === "proposal" &&
+            roundOf(s.content.stageId) === round,
+    );
+    jsonFile(join(roundDir, "proposals.json"), proposals);
+    writeFileSync(
+        join(roundDir, "proposals.txt"),
+        proposals
+            .map(
+                (s) =>
+                    `${"=".repeat(20)} ${s.content.item}（${s.content.author}，记录 ${s.seq}）\n${json(s.content.content)}`,
+            )
+            .join("\n"),
+    );
+
+    // 解释（任务书 17 第 3.2 节第 1 样）：作者、正文、每格三个数。
+    const explanations = finals.filter(
+        (s) =>
+            s.content.kind === "explanation" &&
+            roundOf(s.content.stageId) === round,
+    );
+    jsonFile(join(roundDir, "explanations.json"), explanations);
+    writeFileSync(
+        join(roundDir, "explanations.txt"),
+        explanations
+            .flatMap((s) =>
+                (
+                    s.content.content as {
+                        explanations: Array<{
+                            body: string;
+                            cells?: Array<{ factor: string; range: unknown }>;
+                        }>;
+                    }
+                ).explanations.map(
+                    (e, i) =>
+                        `${"=".repeat(20)} ${explanationId(round, s.content.author, i + 1)}（${s.content.author} 的第 ${i + 1} 条，记录 ${s.seq}）\n${e.body}\n三个数：${(e.cells ?? []).map((c) => `${c.factor} ${JSON.stringify(c.range)}`).join("；")}`,
+                ),
+            )
+            .join("\n\n"),
+    );
+
+    // 选格决定全文（第 2 样）：因子、合并范围、各组合的信息价值/成本/净值、选中与没选的原因都在这条决定里。
+    const cellsDecision =
+        ofKind(records, "decision").find(
+            (d) => d.content.decision === "cells" && d.content.round === round,
+        ) ?? null;
+    const chosenCells = board.decision("cells", round)?.output.chosen ?? null;
+    jsonFile(join(roundDir, "cells-decision.json"), cellsDecision);
+
+    // 结算全文（第 4 样）：结算记录和同一事务里押的因子决定。
+    const settlement =
+        ofKind(records, "settlement").find((s) => s.content.round === round) ??
+        null;
+    const bet =
+        ofKind(records, "decision").find(
+            (d) => d.content.decision === "bet" && d.content.round === round,
+        ) ?? null;
+    jsonFile(join(roundDir, "settlement.json"), { settlement, bet });
+
+    // 收尾全文（第 5 样）。
+    const wrapUps = finals.filter(
+        (s) =>
+            s.content.kind === "wrapUp" && roundOf(s.content.stageId) === round,
+    );
+    jsonFile(join(roundDir, "wrapups.json"), wrapUps);
+    writeFileSync(
+        join(roundDir, "wrapups.txt"),
+        wrapUps
+            .map(
+                (s) =>
+                    `${"=".repeat(20)} ${s.content.author}（记录 ${s.seq}）\n${(s.content.content as { body: string }).body}`,
+            )
+            .join("\n\n"),
+    );
+
+    // 消融格（第 3 样）：冻结协议、位置表、完整代码与格代码及 diff、alsoOff、concerns、每次核验逐项结论与退回、重做记录。
+    const cells = [];
+    if (chosenCells !== null) {
+        const adopt = board.decision("adopt", round)!;
+        const candidate = adopt.output.adopted!;
+        const full = board.code(candidate);
+        const baseline = board.code(board.baselineOf(round));
+        for (const key of chosenCells) {
+            const cell = cellId(candidate, key);
+            const cellDir = join(roundDir, "cells", cell);
+            mkdirSync(cellDir, { recursive: true });
+            jsonFile(join(cellDir, "protocol.json"), {
+                record: ofKind(records, "protocol")
+                    .filter((p) => p.content.candidate === candidate)
+                    .at(-1),
+                resolvedMetrics: protocolMetrics(
+                    board,
+                    board.protocol(candidate),
+                ),
+                factorLocations: full.locations,
+                baselineVersion: baseline.version,
+                fullVersion: full.version,
+            });
+            checkout(full.version, join(cellDir, "full"));
+            checkout(baseline.version, join(cellDir, "baseline"));
+            // 每次实现和代码版本对号（同一事务里先写提交、紧接着写代码；交"跑不起来"时没有代码记录）。
+            const impls = finals.filter(
+                (s) =>
+                    s.content.kind === "ablationImplementation" &&
+                    s.content.target === cell,
+            );
+            const codes = ofKind(records, "code").filter(
+                (c) => c.content.object === cell,
+            );
+            const attempts = impls.map((s) => ({
+                submission: s,
+                code: codes.find((c) => c.seq === s.seq + 1) ?? null,
+            }));
+            jsonFile(join(cellDir, "implementations.json"), attempts);
+            const verifications = finals
+                .filter(
+                    (s) =>
+                        s.content.kind === "ablationVerification" &&
+                        s.content.target === cell,
+                )
+                .map((s) => {
+                    const impl =
+                        impls.filter((x) => x.seq < s.seq).at(-1) ?? null;
+                    return {
+                        submission: s,
+                        judgedImplSeq: impl?.seq ?? null,
+                        outcome:
+                            impl === null
+                                ? null
+                                : ablationOutcome(
+                                      s.content.content as Parameters<
+                                          typeof ablationOutcome
+                                      >[0],
+                                      key,
+                                      (
+                                          impl.content.content as {
+                                              alsoOff: string[];
+                                          }
+                                      ).alsoOff,
+                                  ),
+                    };
+                });
+            jsonFile(join(cellDir, "verifications.json"), verifications);
+            jsonFile(
+                join(cellDir, "rejections.json"),
+                ofKind(records, "rejection").filter(
+                    (r) => r.content.target === cell,
+                ),
+            );
+            const lastCode = codes.at(-1) ?? null;
+            if (lastCode !== null) {
+                checkout(lastCode.content.version, join(cellDir, "cell"));
+                jsonFile(
+                    join(cellDir, "diff.json"),
+                    compareVersions(
+                        repo,
+                        full.version,
+                        lastCode.content.version,
+                    ),
+                );
+            }
+            for (const attempt of attempts) {
+                if (attempt.code === null) continue;
+                const path = join(cellDir, `code-${attempt.code.seq}`);
+                checkout(attempt.code.content.version, path);
+                jsonFile(
+                    join(cellDir, `diff-${attempt.code.seq}.json`),
+                    compareVersions(
+                        repo,
+                        full.version,
+                        attempt.code.content.version,
+                    ),
+                );
+            }
+            jsonFile(
+                join(cellDir, "calls.json"),
+                ofKind(records, "callStarted")
+                    .filter((c) => {
+                        const material = c.content.materials as TaskMaterial;
+                        return (
+                            material.object === cell &&
+                            [
+                                "ablationImplementation",
+                                "ablationVerification",
+                            ].includes(material.kind)
+                        );
+                    })
+                    .map((c) => ({
+                        record: c,
+                        sessions: join(
+                            study,
+                            "calls",
+                            c.content.call,
+                            "sessions",
+                        ),
+                    })),
+            );
+            cells.push({
+                cell,
+                factor: splitCell(cell).key,
+                dir: cellDir,
+                ran: lastCode !== null,
+            });
+        }
+    }
+
+    writeFileSync(
+        join(roundDir, "review.txt"),
+        `任务书17第3.2节，只读本目录，中文直接回复，不写文件。\n` +
+            `第1样（解释）：读 explanations.txt、cells-decision.json、settlement.json。看不到消融结果地回答：机制是否具体、写了什么情况下不成立没有、三个数和正文是否一致、范围宽窄有没有依据、同轮几人的解释有没有真正不同、"没有具体机制判断"用得诚不诚实；再对 settlement.json 看每条解释在每格被确认还是否定。\n` +
+            `第2样（选格）：读 cells-decision.json，列出各格合并范围、各组合信息价值/成本/净值、选中和没选的原因。\n` +
+            `第3样（消融实现与核验，每格一个全新上下文）：读 cells/<格>/ 下的 protocol.json、full/、cell/、diff.json、implementations.json、verifications.json、rejections.json。判断：是不是只关了指定的因子；保障原样保留没有、关掉的因子的指标写 null 没有；答案路径标签和关掉以后的分支对不对得上；核验员每次逐项结论判得对不对；重做的只改了被指出的地方没有。\n` +
+            `第5样（收尾与接力）：读 wrapups.txt、settlement.json，再对着 round-${round + 1}/proposals.txt（没有就是最后一轮）看下一轮提案用上了哪些发现。\n`,
+    );
+    index.push({
+        round,
+        dir: roundDir,
+        hasAblation: chosenCells !== null,
+        cells,
+    });
+}
+jsonFile(join(output, "index.json"), { study: dir, rounds: index });
+console.log(`审阅材料：${output}`);