@modusensus/dsh-mneme 0.8.14 → 0.8.15

package/test/baseline-persist.test.js

@@ -1,248 +1,248 @@
-import test from "node:test";
-import assert from "node:assert/strict";
-import { mkdtempSync, rmSync, readFileSync } from "node:fs";
-import { tmpdir } from "node:os";
-import { join } from "node:path";
-import { fileURLToPath } from "node:url";
-import { DatabaseSync } from "node:sqlite";
-import { createStore } from "../src/store.js";
-import { createService } from "../src/service.js";
-import { createDreamScheduler } from "../src/dream.js";
-
-// 回归（issue #89 基线半边）：阈值基线 baseline{count,chars} 只活在调度器闭包里，
-// 进程重启归零。库里记忆远多于阈值时，重启后第一条写入（更新类写入同样计入
-// notifyWrite）就让 overBase 成立，绕过阈值开跑一整轮巩固——0.8.11 实测四次重启
-// 分别只新增 6/7/0/8 条记忆（阈值 10 条 / 5000 字符），每轮约 16 万 tokens。
-// #291 只持久化了 lastRunAt（时间闸门那半边），本文件锁另一半：
-//   1. lastDreamBaseline 只认「成功轮落下的完整基线」，按 run_type 隔离；
-//   2. 播种后重启不再绕过阈值——同库同写入的无种子对照组仍触发，锁的正是这个差异；
-//   3. 播种不妨碍阈值正常累积（计数长到 +10 仍然触发）；
-//   4. runDream 成功轮把基线落库并回给调度器（进程内刷新与跨重启播种同一份数字）；
-//   5. 老库两列由幂等迁移补上，存量行认不出基线 → 退回零基线（升级前行为）。
-
-/** 与阈值判定同口径的活跃库规模（测试侧独立实现，避免与被测代码共用一份过滤条件）。 */
-function activeSize(store) {
-  const rows = store.all().filter((m) => !m.archived && m.type !== "summary" && m.type !== "document");
-  return {
-    count: rows.length,
-    chars: rows.reduce((sum, m) => sum + m.title.length + m.content.length, 0)
-  };
-}
-
-function seedMemories(service, n, tag = "a") {
-  // 标题必须全局唯一：saveWithDedupe 的去重键是 type+title+scope，同名会被合并成
-  // 同一条（条数不增长），「累计到阈值」的断言就会假失败。
-  for (let i = 0; i < n; i++) {
-    service.saveWithDedupe({ type: "history", title: `${tag}-${i}`, content: "x".repeat(300), importance: 3 });
-  }
-}
-
-test("issue#89: store.lastDreamBaseline reads the last baseline-advancing run only", () => {
-  const store = createStore(":memory:");
-  assert.equal(store.lastDreamBaseline(), null, "empty trail → null (never ran)");
-  // 非 ok 轮不推进基线（既有语义）→ 不落这两列，播种端必须无视它
-  store.saveDreamRun({ status: "failed", snapshot_hash: "h0", input_count: 1, receipt: "r0", created_at: "2026-09-30T10:00:00.000Z", run_type: "auto" });
-  assert.equal(store.lastDreamBaseline("auto"), null, "failed run carries no baseline");
-  store.saveDreamRun({ status: "ok", snapshot_hash: "h1", input_count: 1, receipt: "r1", created_at: "2026-09-30T10:01:00.000Z", run_type: "auto", store_count: 4, store_chars: 400 });
-  assert.deepEqual(store.lastDreamBaseline("auto"), { count: 4, chars: 400 });
-  // sleep/organize 的轮次不参与 auto 的阈值判定
-  store.saveDreamRun({ status: "ok", snapshot_hash: "h2", input_count: 1, receipt: "r2", created_at: "2026-09-30T10:02:00.000Z", run_type: "sleep", store_count: 9, store_chars: 900 });
-  assert.deepEqual(store.lastDreamBaseline("auto"), { count: 4, chars: 400 }, "sleep rows are filtered out of the auto baseline");
-  assert.deepEqual(store.lastDreamBaseline("sleep"), { count: 9, chars: 900 });
-  // 半个基线（只有条数没有字符数）写入侧就该拒绝：否则阈值会按 0 字符起算
-  store.saveDreamRun({ status: "ok", snapshot_hash: "h3", input_count: 1, receipt: "r3", created_at: "2026-09-30T10:03:00.000Z", run_type: "auto", store_count: 7 });
-  assert.deepEqual(store.lastDreamBaseline("auto"), { count: 4, chars: 400 }, "half a baseline is not accepted");
-  store.close();
-});
-
-test("issue#89: a seeded baseline survives a restart — the first write no longer bypasses the threshold", async () => {
-  const store = createStore(":memory:");
-  const service = createService({ store, mirror: null, config: {} });
-  seedMemories(service, 12); // 条数与字符数都远超阈值（12 ≥ 10 条）
-  // 上一进程的成功轮把「当时的库规模」落进审计表，然后进程重启：
-  const size = activeSize(store);
-  store.saveDreamRun({
-    status: "ok", snapshot_hash: "h", input_count: 1, receipt: "r",
-    created_at: new Date(Date.now() - 10).toISOString(), run_type: "auto",
-    store_count: size.count, store_chars: size.chars
-  });
-  const opts = { onRun: async () => ({ ok: false }), thresholdCount: 10, thresholdChars: 5000, delayMs: 0, logger: { warn: () => {} } };
-  const seeded = createDreamScheduler({ ...opts, baselineSeed: store.lastDreamBaseline("auto") });
-  assert.equal(seeded.maybeSchedule(service), false, "seeded baseline blocks the first write after restart");
-  // 对照组：同一个库、同一次写入，没有种子就是被修掉的那个缺陷。若哪天播种被忽略，
-  // 上面那条断言会与这条同真同假——所以两条必须成对存在。
-  const unseeded = createDreamScheduler(opts);
-  assert.equal(unseeded.maybeSchedule(service), true, "without the seed the threshold is bypassed (the defect this test locks)");
-  await unseeded.dispose(); // 清掉 0ms 待跑计时器，别让对照组真开跑
-  store.close();
-});
-
-test("issue#89: a seeded baseline still lets the threshold accumulate", async () => {
-  const store = createStore(":memory:");
-  const service = createService({ store, mirror: null, config: {} });
-  seedMemories(service, 12);
-  const dream = createDreamScheduler({
-    onRun: async () => ({ ok: false }), thresholdCount: 10, thresholdChars: 5000, delayMs: 0,
-    baselineSeed: activeSize(store), logger: { warn: () => {} }
-  });
-  seedMemories(service, 5, "b");
-  assert.equal(dream.maybeSchedule(service), false, "half the threshold is not enough");
-  seedMemories(service, 5, "c"); // 累计 +10 条
-  assert.equal(dream.maybeSchedule(service), true, "the threshold still fires once the store grew by it");
-  await dream.dispose();
-  store.close();
-});
-
-test("issue#89: the run-reported baseline is what the scheduler keeps (single source of truth)", async () => {
-  const store = createStore(":memory:");
-  const service = createService({ store, mirror: null, config: {} });
-  seedMemories(service, 12);
-  const reported = { count: 50, chars: 50000 };
-  const dream = createDreamScheduler({
-    onRun: async () => ({ ok: true, baseline: reported }),
-    thresholdCount: 10, thresholdChars: 5000, delayMs: 0, logger: { warn: () => {} }
-  });
-  assert.equal(dream.maybeSchedule(service), true, "no seed yet → the first run is allowed");
-  await new Promise((r) => setTimeout(r, 10)); // 让 run 收尾并刷新基线
-  // 运行后再长 15 条：若调度器改用「现场计算」的基线（12），这里会触发；用 run
-  // 自报的 50 则不触发。这条差异就是「审计行与内存基线同一份数字」的锁。
-  seedMemories(service, 15, "grow");
-  assert.equal(dream.maybeSchedule(service), false, "the persisted-reported baseline governs, not the live store size");
-  await dream.dispose();
-  store.close();
-});
-
-test("issue#89: an ok run persists its post-run store size; a failed run persists none", async () => {
-  const store = createStore(":memory:");
-  const service = createService({ store, mirror: null, config: {} });
-  const a = service.saveWithDedupe({ type: "project", title: "旧1", content: "第一段内容" });
-  const b = service.saveWithDedupe({ type: "project", title: "旧2", content: "第二段内容" });
-  const dream = createDreamScheduler({ thresholdCount: 1, thresholdChars: 0, delayMs: 0, logger: { warn: () => {} } });
-
-  // 失败轮（无路由）：不能留下任何基线——否则重启播种会拿到一个「没推进过」的假基线
-  const failed = await dream.runDream({ llm: { stream: async function* () {} }, logger: { warn: () => {} } }, service, {});
-  assert.equal(failed.ok, false, "no route → failed run");
-  assert.equal(failed.baseline, undefined, "failed run reports no baseline");
-  assert.equal(store.lastDreamBaseline("auto"), null, "failed run persists no baseline");
-
-  // 成功轮：落库的数字 = 本轮结束后的活跃库规模（合并已归档一方，summary 不计入）
-  let calls = 0;
-  const ctx = {
-    llm: {
-      stream: async function* () {
-        calls++;
-        const text = calls === 1
-          ? JSON.stringify([{ action: "merge", ids: [a.memory.id, b.memory.id], title: "合并标题", content: "合并后的内容", importance: 4, keepSource: a.memory.id }])
-          : "记忆库总览摘要文本";
-        yield { type: "text-delta", text };
-        yield { type: "finish", reason: { kind: "ok" } };
-      }
-    },
-    logger: { warn: () => {} }
-  };
-  const ok = await dream.runDream(ctx, service, { dreamProvider: "deepseek", dreamModel: "deepseek-chat" });
-  assert.equal(ok.ok, true, "run with decisions + summary is ok");
-  const expected = activeSize(store);
-  assert.deepEqual(ok.baseline, expected, "the run reports the post-run store size");
-  assert.deepEqual(store.lastDreamBaseline("auto"), expected, "the audit row carries the same numbers");
-  store.close();
-});
-
-test("issue#89: legacy dream_runs gains the baseline columns on open (idempotent ALTER)", () => {
-  const dir = mkdtempSync(join(tmpdir(), "dsh-mneme-baseline-migrate-"));
-  const dbPath = join(dir, "legacy.db");
-  try {
-    const legacy = new DatabaseSync(dbPath);
-    legacy.exec(`CREATE TABLE dream_runs (
-      id TEXT PRIMARY KEY, created_at TEXT NOT NULL, status TEXT NOT NULL, error TEXT,
-      provider TEXT, model TEXT, snapshot_hash TEXT NOT NULL, input_count INTEGER NOT NULL,
-      input TEXT, decisions TEXT, outcome TEXT, applied INTEGER NOT NULL DEFAULT 0,
-      summary_stored INTEGER NOT NULL DEFAULT 0, receipt TEXT NOT NULL,
-      policy_epoch INTEGER NOT NULL DEFAULT 0, run_type TEXT NOT NULL DEFAULT 'auto', skipped TEXT
-    );`);
-    legacy.prepare(
-      `INSERT INTO dream_runs (id, created_at, status, snapshot_hash, input_count, applied, summary_stored, receipt)
-       VALUES ('legacy', '2026-09-30T10:00:00.000Z', 'ok', 'h', 1, 0, 0, 'r')`
-    ).run();
-    legacy.close();
-
-    const store = createStore(dbPath);
-    const cols = store.db.prepare("PRAGMA table_info(dream_runs)").all().map((c) => c.name);
-    assert.ok(cols.includes("store_count") && cols.includes("store_chars"), "baseline columns added to legacy db");
-    // 存量行两列皆 NULL：反推不出「那轮结束时库有多大」，认不出就不认——退回零基线
-    assert.equal(store.lastDreamBaseline("auto"), null, "legacy rows carry no baseline");
-    store.close();
-
-    const again = createStore(dbPath); // 二次打开：ALTER 幂等，不抛 duplicate column
-    assert.equal(again.lastDreamBaseline("auto"), null);
-    again.close();
-  } finally {
-    rmSync(dir, { recursive: true, force: true });
-  }
-});
-
-test("issue#89: an unusable seed is treated as no baseline at all", () => {
-  const store = createStore(":memory:");
-  const service = createService({ store, mirror: null, config: {} });
-  seedMemories(service, 12);
-  const opts = { onRun: async () => ({ ok: false }), thresholdCount: 10, thresholdChars: 5000, delayMs: 0, logger: { warn: () => {} } };
-  // 库里的持久化值可能是老库的 NULL、或手工改坏的库：半个基线（只有条数）若被当成
-  // 「12 条 0 字符」，阈值会按 0 字符起算——比不认还糟。非法种子一律退回零基线。
-  for (const bad of [{ count: 12 }, { chars: 2000 }, { count: -1, chars: 2000 }, { count: 1.5, chars: 10 }, "nope"]) {
-    const sched = createDreamScheduler({ ...opts, baselineSeed: bad });
-    assert.equal(sched.maybeSchedule(service), true, `unusable seed ${JSON.stringify(bad)} falls back to the zero baseline`);
-    sched.dispose();
-  }
-  store.close();
-});
-
-test("issue#89: a failing baseline read never breaks the run nor fakes a baseline", async () => {
-  const store = createStore(":memory:");
-  const service = createService({ store, mirror: null, config: {} });
-  const a = service.saveWithDedupe({ type: "project", title: "旧1", content: "第一段内容" });
-  // finish 阶段读库失败（关库 / IO 故障）：必须不反噬巩固本身（审计行照写），也不
-  // 写一个猜出来的基线（落 NULL → 下次重启退回零基线，而不是拿假基线抬高阈值）。
-  // 故障点在「总览出文本之后」才武装：runDream 前面还要读两次库（候选集与总览输入），
-  // 过早抛错会把整轮打成 failed，验的就不是收尾取证这条路了。
-  const realAll = service.all.bind(service);
-  let armed = false;
-  let calls = 0;
-  service.all = () => {
-    if (armed) throw new Error("store closed");
-    return realAll();
-  };
-  const ctx = {
-    llm: {
-      stream: async function* () {
-        if (calls++ === 0) {
-          yield { type: "text-delta", text: JSON.stringify([{ action: "keep", ids: [a.memory.id] }]) };
-        } else {
-          armed = true;
-          yield { type: "text-delta", text: "记忆库总览摘要文本" };
-        }
-        yield { type: "finish", reason: { kind: "ok" } };
-      }
-    },
-    logger: { warn: () => {} }
-  };
-  const dream = createDreamScheduler({ thresholdCount: 1, thresholdChars: 0, delayMs: 0, logger: { warn: () => {} } });
-  const ok = await dream.runDream(ctx, service, { dreamProvider: "deepseek", dreamModel: "deepseek-chat" });
-  assert.equal(ok.ok, true, "the run still completes when the baseline read fails");
-  assert.equal(ok.baseline, undefined, "no made-up baseline is reported");
-  assert.ok(store.listDreamRuns({ limit: 1 })[0], "the audit row is still written");
-  assert.equal(store.listDreamRuns({ limit: 1 })[0].store_count, undefined, "the audit row carries no baseline");
-  store.close();
-});
-
-// --- CodeRabbit 在 #291 上的同类要求：index.js 的 seed 是单行无分支接线，仓库没有
-// 插件级挂载 harness（apply 需要完整宿主 ctx），改为源码级锁——删掉该注入行即红。
-// import 用 URL 相对本文件解析，不依赖测试进程 CWD。
-const indexSrc = readFileSync(fileURLToPath(new URL("../src/index.js", import.meta.url)), "utf8");
-
-test("issue#89: index.js wires the dream baseline seed from the audit trail", () => {
-  assert.match(
-    indexSrc,
-    /baselineSeed:\s*store\.lastDreamBaseline\("auto"\)/,
-    "dream scheduler must seed its threshold baseline from dream_runs (run_type='auto') — remove this line and the restart-bypass returns"
-  );
-});
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, rmSync, readFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { DatabaseSync } from "node:sqlite";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+import { createDreamScheduler } from "../src/dream.js";
+
+// 回归（issue #89 基线半边）：阈值基线 baseline{count,chars} 只活在调度器闭包里，
+// 进程重启归零。库里记忆远多于阈值时，重启后第一条写入（更新类写入同样计入
+// notifyWrite）就让 overBase 成立，绕过阈值开跑一整轮巩固——0.8.11 实测四次重启
+// 分别只新增 6/7/0/8 条记忆（阈值 10 条 / 5000 字符），每轮约 16 万 tokens。
+// #291 只持久化了 lastRunAt（时间闸门那半边），本文件锁另一半：
+//   1. lastDreamBaseline 只认「成功轮落下的完整基线」，按 run_type 隔离；
+//   2. 播种后重启不再绕过阈值——同库同写入的无种子对照组仍触发，锁的正是这个差异；
+//   3. 播种不妨碍阈值正常累积（计数长到 +10 仍然触发）；
+//   4. runDream 成功轮把基线落库并回给调度器（进程内刷新与跨重启播种同一份数字）；
+//   5. 老库两列由幂等迁移补上，存量行认不出基线 → 退回零基线（升级前行为）。
+
+/** 与阈值判定同口径的活跃库规模（测试侧独立实现，避免与被测代码共用一份过滤条件）。 */
+function activeSize(store) {
+  const rows = store.all().filter((m) => !m.archived && m.type !== "summary" && m.type !== "document");
+  return {
+    count: rows.length,
+    chars: rows.reduce((sum, m) => sum + m.title.length + m.content.length, 0)
+  };
+}
+
+function seedMemories(service, n, tag = "a") {
+  // 标题必须全局唯一：saveWithDedupe 的去重键是 type+title+scope，同名会被合并成
+  // 同一条（条数不增长），「累计到阈值」的断言就会假失败。
+  for (let i = 0; i < n; i++) {
+    service.saveWithDedupe({ type: "history", title: `${tag}-${i}`, content: "x".repeat(300), importance: 3 });
+  }
+}
+
+test("issue#89: store.lastDreamBaseline reads the last baseline-advancing run only", () => {
+  const store = createStore(":memory:");
+  assert.equal(store.lastDreamBaseline(), null, "empty trail → null (never ran)");
+  // 非 ok 轮不推进基线（既有语义）→ 不落这两列，播种端必须无视它
+  store.saveDreamRun({ status: "failed", snapshot_hash: "h0", input_count: 1, receipt: "r0", created_at: "2026-09-30T10:00:00.000Z", run_type: "auto" });
+  assert.equal(store.lastDreamBaseline("auto"), null, "failed run carries no baseline");
+  store.saveDreamRun({ status: "ok", snapshot_hash: "h1", input_count: 1, receipt: "r1", created_at: "2026-09-30T10:01:00.000Z", run_type: "auto", store_count: 4, store_chars: 400 });
+  assert.deepEqual(store.lastDreamBaseline("auto"), { count: 4, chars: 400 });
+  // sleep/organize 的轮次不参与 auto 的阈值判定
+  store.saveDreamRun({ status: "ok", snapshot_hash: "h2", input_count: 1, receipt: "r2", created_at: "2026-09-30T10:02:00.000Z", run_type: "sleep", store_count: 9, store_chars: 900 });
+  assert.deepEqual(store.lastDreamBaseline("auto"), { count: 4, chars: 400 }, "sleep rows are filtered out of the auto baseline");
+  assert.deepEqual(store.lastDreamBaseline("sleep"), { count: 9, chars: 900 });
+  // 半个基线（只有条数没有字符数）写入侧就该拒绝：否则阈值会按 0 字符起算
+  store.saveDreamRun({ status: "ok", snapshot_hash: "h3", input_count: 1, receipt: "r3", created_at: "2026-09-30T10:03:00.000Z", run_type: "auto", store_count: 7 });
+  assert.deepEqual(store.lastDreamBaseline("auto"), { count: 4, chars: 400 }, "half a baseline is not accepted");
+  store.close();
+});
+
+test("issue#89: a seeded baseline survives a restart — the first write no longer bypasses the threshold", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  seedMemories(service, 12); // 条数与字符数都远超阈值（12 ≥ 10 条）
+  // 上一进程的成功轮把「当时的库规模」落进审计表，然后进程重启：
+  const size = activeSize(store);
+  store.saveDreamRun({
+    status: "ok", snapshot_hash: "h", input_count: 1, receipt: "r",
+    created_at: new Date(Date.now() - 10).toISOString(), run_type: "auto",
+    store_count: size.count, store_chars: size.chars
+  });
+  const opts = { onRun: async () => ({ ok: false }), thresholdCount: 10, thresholdChars: 5000, delayMs: 0, logger: { warn: () => {} } };
+  const seeded = createDreamScheduler({ ...opts, baselineSeed: store.lastDreamBaseline("auto") });
+  assert.equal(seeded.maybeSchedule(service), false, "seeded baseline blocks the first write after restart");
+  // 对照组：同一个库、同一次写入，没有种子就是被修掉的那个缺陷。若哪天播种被忽略，
+  // 上面那条断言会与这条同真同假——所以两条必须成对存在。
+  const unseeded = createDreamScheduler(opts);
+  assert.equal(unseeded.maybeSchedule(service), true, "without the seed the threshold is bypassed (the defect this test locks)");
+  await unseeded.dispose(); // 清掉 0ms 待跑计时器，别让对照组真开跑
+  store.close();
+});
+
+test("issue#89: a seeded baseline still lets the threshold accumulate", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  seedMemories(service, 12);
+  const dream = createDreamScheduler({
+    onRun: async () => ({ ok: false }), thresholdCount: 10, thresholdChars: 5000, delayMs: 0,
+    baselineSeed: activeSize(store), logger: { warn: () => {} }
+  });
+  seedMemories(service, 5, "b");
+  assert.equal(dream.maybeSchedule(service), false, "half the threshold is not enough");
+  seedMemories(service, 5, "c"); // 累计 +10 条
+  assert.equal(dream.maybeSchedule(service), true, "the threshold still fires once the store grew by it");
+  await dream.dispose();
+  store.close();
+});
+
+test("issue#89: the run-reported baseline is what the scheduler keeps (single source of truth)", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  seedMemories(service, 12);
+  const reported = { count: 50, chars: 50000 };
+  const dream = createDreamScheduler({
+    onRun: async () => ({ ok: true, baseline: reported }),
+    thresholdCount: 10, thresholdChars: 5000, delayMs: 0, logger: { warn: () => {} }
+  });
+  assert.equal(dream.maybeSchedule(service), true, "no seed yet → the first run is allowed");
+  await new Promise((r) => setTimeout(r, 10)); // 让 run 收尾并刷新基线
+  // 运行后再长 15 条：若调度器改用「现场计算」的基线（12），这里会触发；用 run
+  // 自报的 50 则不触发。这条差异就是「审计行与内存基线同一份数字」的锁。
+  seedMemories(service, 15, "grow");
+  assert.equal(dream.maybeSchedule(service), false, "the persisted-reported baseline governs, not the live store size");
+  await dream.dispose();
+  store.close();
+});
+
+test("issue#89: an ok run persists its post-run store size; a failed run persists none", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const a = service.saveWithDedupe({ type: "project", title: "旧1", content: "第一段内容" });
+  const b = service.saveWithDedupe({ type: "project", title: "旧2", content: "第二段内容" });
+  const dream = createDreamScheduler({ thresholdCount: 1, thresholdChars: 0, delayMs: 0, logger: { warn: () => {} } });
+
+  // 失败轮（无路由）：不能留下任何基线——否则重启播种会拿到一个「没推进过」的假基线
+  const failed = await dream.runDream({ llm: { stream: async function* () {} }, logger: { warn: () => {} } }, service, {});
+  assert.equal(failed.ok, false, "no route → failed run");
+  assert.equal(failed.baseline, undefined, "failed run reports no baseline");
+  assert.equal(store.lastDreamBaseline("auto"), null, "failed run persists no baseline");
+
+  // 成功轮：落库的数字 = 本轮结束后的活跃库规模（合并已归档一方，summary 不计入）
+  let calls = 0;
+  const ctx = {
+    llm: {
+      stream: async function* () {
+        calls++;
+        const text = calls === 1
+          ? JSON.stringify([{ action: "merge", ids: [a.memory.id, b.memory.id], title: "合并标题", content: "合并后的内容", importance: 4, keepSource: a.memory.id }])
+          : "记忆库总览摘要文本";
+        yield { type: "text-delta", text };
+        yield { type: "finish", reason: { kind: "ok" } };
+      }
+    },
+    logger: { warn: () => {} }
+  };
+  const ok = await dream.runDream(ctx, service, { dreamProvider: "deepseek", dreamModel: "deepseek-chat" });
+  assert.equal(ok.ok, true, "run with decisions + summary is ok");
+  const expected = activeSize(store);
+  assert.deepEqual(ok.baseline, expected, "the run reports the post-run store size");
+  assert.deepEqual(store.lastDreamBaseline("auto"), expected, "the audit row carries the same numbers");
+  store.close();
+});
+
+test("issue#89: legacy dream_runs gains the baseline columns on open (idempotent ALTER)", () => {
+  const dir = mkdtempSync(join(tmpdir(), "dsh-mneme-baseline-migrate-"));
+  const dbPath = join(dir, "legacy.db");
+  try {
+    const legacy = new DatabaseSync(dbPath);
+    legacy.exec(`CREATE TABLE dream_runs (
+      id TEXT PRIMARY KEY, created_at TEXT NOT NULL, status TEXT NOT NULL, error TEXT,
+      provider TEXT, model TEXT, snapshot_hash TEXT NOT NULL, input_count INTEGER NOT NULL,
+      input TEXT, decisions TEXT, outcome TEXT, applied INTEGER NOT NULL DEFAULT 0,
+      summary_stored INTEGER NOT NULL DEFAULT 0, receipt TEXT NOT NULL,
+      policy_epoch INTEGER NOT NULL DEFAULT 0, run_type TEXT NOT NULL DEFAULT 'auto', skipped TEXT
+    );`);
+    legacy.prepare(
+      `INSERT INTO dream_runs (id, created_at, status, snapshot_hash, input_count, applied, summary_stored, receipt)
+       VALUES ('legacy', '2026-09-30T10:00:00.000Z', 'ok', 'h', 1, 0, 0, 'r')`
+    ).run();
+    legacy.close();
+
+    const store = createStore(dbPath);
+    const cols = store.db.prepare("PRAGMA table_info(dream_runs)").all().map((c) => c.name);
+    assert.ok(cols.includes("store_count") && cols.includes("store_chars"), "baseline columns added to legacy db");
+    // 存量行两列皆 NULL：反推不出「那轮结束时库有多大」，认不出就不认——退回零基线
+    assert.equal(store.lastDreamBaseline("auto"), null, "legacy rows carry no baseline");
+    store.close();
+
+    const again = createStore(dbPath); // 二次打开：ALTER 幂等，不抛 duplicate column
+    assert.equal(again.lastDreamBaseline("auto"), null);
+    again.close();
+  } finally {
+    rmSync(dir, { recursive: true, force: true });
+  }
+});
+
+test("issue#89: an unusable seed is treated as no baseline at all", () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  seedMemories(service, 12);
+  const opts = { onRun: async () => ({ ok: false }), thresholdCount: 10, thresholdChars: 5000, delayMs: 0, logger: { warn: () => {} } };
+  // 库里的持久化值可能是老库的 NULL、或手工改坏的库：半个基线（只有条数）若被当成
+  // 「12 条 0 字符」，阈值会按 0 字符起算——比不认还糟。非法种子一律退回零基线。
+  for (const bad of [{ count: 12 }, { chars: 2000 }, { count: -1, chars: 2000 }, { count: 1.5, chars: 10 }, "nope"]) {
+    const sched = createDreamScheduler({ ...opts, baselineSeed: bad });
+    assert.equal(sched.maybeSchedule(service), true, `unusable seed ${JSON.stringify(bad)} falls back to the zero baseline`);
+    sched.dispose();
+  }
+  store.close();
+});
+
+test("issue#89: a failing baseline read never breaks the run nor fakes a baseline", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const a = service.saveWithDedupe({ type: "project", title: "旧1", content: "第一段内容" });
+  // finish 阶段读库失败（关库 / IO 故障）：必须不反噬巩固本身（审计行照写），也不
+  // 写一个猜出来的基线（落 NULL → 下次重启退回零基线，而不是拿假基线抬高阈值）。
+  // 故障点在「总览出文本之后」才武装：runDream 前面还要读两次库（候选集与总览输入），
+  // 过早抛错会把整轮打成 failed，验的就不是收尾取证这条路了。
+  const realAll = service.all.bind(service);
+  let armed = false;
+  let calls = 0;
+  service.all = () => {
+    if (armed) throw new Error("store closed");
+    return realAll();
+  };
+  const ctx = {
+    llm: {
+      stream: async function* () {
+        if (calls++ === 0) {
+          yield { type: "text-delta", text: JSON.stringify([{ action: "keep", ids: [a.memory.id] }]) };
+        } else {
+          armed = true;
+          yield { type: "text-delta", text: "记忆库总览摘要文本" };
+        }
+        yield { type: "finish", reason: { kind: "ok" } };
+      }
+    },
+    logger: { warn: () => {} }
+  };
+  const dream = createDreamScheduler({ thresholdCount: 1, thresholdChars: 0, delayMs: 0, logger: { warn: () => {} } });
+  const ok = await dream.runDream(ctx, service, { dreamProvider: "deepseek", dreamModel: "deepseek-chat" });
+  assert.equal(ok.ok, true, "the run still completes when the baseline read fails");
+  assert.equal(ok.baseline, undefined, "no made-up baseline is reported");
+  assert.ok(store.listDreamRuns({ limit: 1 })[0], "the audit row is still written");
+  assert.equal(store.listDreamRuns({ limit: 1 })[0].store_count, undefined, "the audit row carries no baseline");
+  store.close();
+});
+
+// --- CodeRabbit 在 #291 上的同类要求：index.js 的 seed 是单行无分支接线，仓库没有
+// 插件级挂载 harness（apply 需要完整宿主 ctx），改为源码级锁——删掉该注入行即红。
+// import 用 URL 相对本文件解析，不依赖测试进程 CWD。
+const indexSrc = readFileSync(fileURLToPath(new URL("../src/index.js", import.meta.url)), "utf8");
+
+test("issue#89: index.js wires the dream baseline seed from the audit trail", () => {
+  assert.match(
+    indexSrc,
+    /baselineSeed:\s*store\.lastDreamBaseline\("auto"\)/,
+    "dream scheduler must seed its threshold baseline from dream_runs (run_type='auto') — remove this line and the restart-bypass returns"
+  );
+});