@modusensus/dsh-mneme 0.5.2 → 0.6.0

package/test/normalize-decisions.test.js

@@ -0,0 +1,120 @@
+// Field-name normalization guard (v0.5.3): thinking-type models
+// (deepseek-v4-flash etc.) sometimes ignore the prompt's exact decision schema
+// and emit alias keys —实测方案 A 输出 "consolidation"/"target_ids"、方案 B
+// 输出 "action"/"targetIds"，都不是插件要求的 "action"/"ids"。normalizeDecisions
+// 在 validateDecisions 之前把这些变体重写回规范字段名，让"语义正确但 schema
+// 不听话"的输出仍被应用，而不是整单被拒。
+import test from "node:test";
+import assert from "node:assert/strict";
+import { normalizeDecisions, createDreamScheduler } from "../src/dream.js";
+import { createStore } from "../src/store.js";
+import { createService } from "../src/service.js";
+
+test("normalizeDecisions: canonical decisions pass through untouched", () => {
+  const raw = [
+    { action: "merge", ids: ["m1", "m2"], keepSource: "m1", title: "t", content: "c", importance: 4, reason: "主题相近" },
+    { action: "conflict", winner: "m3", loser: "m4", reason: "矛盾" },
+    { action: "update", ids: ["m5"], content: "修正" },
+    { action: "archive", ids: ["m6"], reason: "过时" },
+    { action: "create", type: "pattern", title: "新模式", content: "发现", importance: 3 }
+  ];
+  assert.deepEqual(normalizeDecisions(raw), raw);
+});
+
+test("normalizeDecisions: alias keys rewritten to canonical names", () => {
+  const raw = [
+    { consolidation: "merge", target_ids: ["m1", "m2"], keep_source: "m1", new_title: "t", new_content: "c", priority: 4, rationale: "主题相近" },
+    { action: "conflict", winner_id: "m3", loser_id: "m4", why: "矛盾" },
+    { action: "update", targetIds: ["m5"], merged_content: "修正" },
+    { action: "archive", memory_ids: ["m6"] }
+  ];
+  assert.deepEqual(normalizeDecisions(raw), [
+    { action: "merge", ids: ["m1", "m2"], keepSource: "m1", title: "t", content: "c", importance: 4, reason: "主题相近" },
+    { action: "conflict", winner: "m3", loser: "m4", reason: "矛盾" },
+    { action: "update", ids: ["m5"], content: "修正" },
+    { action: "archive", ids: ["m6"] }
+  ]);
+});
+
+test("normalizeDecisions: action value synonyms normalized", () => {
+  const raw = [
+    { action: "archived", ids: ["m1"], reason: "重复" },
+    { action: "consolidation", target_ids: ["m2", "m3"], keep_source: "m2" },
+    { action: "combine", targetIds: ["m4", "m5"], keeper: "m5" },
+    { action: "Remove", memory_ids: ["m6"] }
+  ];
+  const normalized = normalizeDecisions(raw);
+  assert.equal(normalized[0].action, "archive");
+  assert.equal(normalized[1].action, "merge");
+  assert.equal(normalized[1].ids[0], "m2");
+  assert.equal(normalized[2].action, "merge");
+  assert.equal(normalized[2].keepSource, "m5");
+  assert.equal(normalized[3].action, "archive");
+});
+
+test("normalizeDecisions: wrapper object around the array is unwrapped", () => {
+  const raw = { consolidation: [{ action: "merge", ids: ["m1"], keepSource: "m1", title: "t", content: "c" }] };
+  assert.deepEqual(normalizeDecisions(raw), [
+    { action: "merge", ids: ["m1"], keepSource: "m1", title: "t", content: "c" }
+  ]);
+  const raw2 = { decisions: [{ action: "archive", target_ids: ["m9"] }] };
+  assert.deepEqual(normalizeDecisions(raw2), [{ action: "archive", ids: ["m9"] }]);
+});
+
+test("normalizeDecisions: single decision object (non-array) normalized too", () => {
+  const raw = { action: "update", targetIds: ["m1"], new_content: "修正" };
+  assert.deepEqual(normalizeDecisions(raw), [{ action: "update", ids: ["m1"], content: "修正" }]);
+});
+
+test("normalizeDecisions: single-string ids become a one-element array", () => {
+  const raw = [{ action: "archive", target_id: "m7" }];
+  // target_id (singular) is not an alias key — the raw key survives untouched
+  // and validateDecisions rejects it (safe side). But targetIds (plural, string)
+  // IS mapped and wrapped.
+  const raw2 = [{ action: "archive", targetIds: "m7" }];
+  assert.deepEqual(normalizeDecisions(raw2), [{ action: "archive", ids: ["m7"] }]);
+  assert.deepEqual(normalizeDecisions(raw), [{ action: "archive", target_id: "m7" }]);
+});
+
+test("normalizeDecisions: create keeps its type field (never mistaken for action)", () => {
+  const raw = { action: "create", type: "preference", title: "语言", content: "中文", importance: 3 };
+  assert.deepEqual(normalizeDecisions(raw), [
+    { action: "create", type: "preference", title: "语言", content: "中文", importance: 3 }
+  ]);
+});
+
+test("normalizeDecisions: null / non-parseable returns as-is (caller's no-json branch)", () => {
+  assert.equal(normalizeDecisions(null), null);
+  assert.equal(normalizeDecisions(undefined), undefined);
+  assert.equal(normalizeDecisions("not json"), "not json");
+});
+
+test("dream runDream applies alias-key consolidation output end-to-end", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const dream = createDreamScheduler({ onRun: () => Promise.resolve({ ok: true, skipped: true }) });
+  const { memory: a } = service.saveWithDedupe({ type: "project", title: "插件", content: "旧", importance: 3 });
+  const { memory: b } = service.saveWithDedupe({ type: "project", title: "插件2", content: "新细节", importance: 4 });
+  const ctx = {
+    logger: { warn: () => {} },
+    agentDefaultModel: { currentSelection: () => ({ provider: "mock", model: "mock-model" }) },
+    llm: {
+      async *stream(options) {
+        const userText = options.messages.find((m) => m.role === "user")?.content?.[0]?.text ?? "";
+        if (userText.startsWith("id=")) {
+          // deepseek-v4-flash 实测风格：consolidation 当 action、target_ids 当 ids
+          yield { type: "text-delta", index: 0, text: JSON.stringify([
+            { consolidation: "merge", target_ids: [a.id, b.id], keep_source: b.id, new_title: "插件总览", new_content: "合并内容", priority: 4 }
+          ]) };
+        } else {
+          yield { type: "text-delta", index: 0, text: "记忆库总览：默认摘要。" };
+        }
+        yield { type: "finish", reason: { kind: "stop" } };
+      }
+    }
+  };
+  const result = await dream.runDream(ctx, service, {});
+  assert.equal(result.ok, true, "alias-key output must not be rejected wholesale");
+  assert.ok(result.applied > 0, "alias-key decisions still land changes");
+  store.close();
+});

package/test/reasoning-effort.test.js

@@ -36,6 +36,11 @@ test("issue#9: reasoningEffort config defaults to none and rejects unknown value
   assert.equal(cfg.sleepReasoningEffort, "none");
   assert.equal(Config({ dreamReasoningEffort: "high" }).dreamReasoningEffort, "high");
   assert.equal(Config({ sleepReasoningEffort: "medium" }).sleepReasoningEffort, "medium");
+  // v0.5.3: 'off' explicitly disables thinking — required for thinking-type
+  // models (deepseek-v4-flash) whose reasoning otherwise drains the whole
+  // token budget and returns an empty consolidation body.
+  assert.equal(Config({ dreamReasoningEffort: "off" }).dreamReasoningEffort, "off");
+  assert.equal(Config({ sleepReasoningEffort: "off" }).sleepReasoningEffort, "off");
   assert.throws(() => Config({ dreamReasoningEffort: "bogus" }), "invalid effort rejected");
   assert.throws(() => Config({ sleepReasoningEffort: "ultra" }), "invalid effort rejected");
 });
@@ -107,6 +112,28 @@ test("issue#9: dream forwards dreamReasoningEffort on both LLM calls", async ()
   store.close();
 });
 
+test("issue#9: dream forwards dreamReasoningEffort:off (thinking disabled) on both LLM calls", async () => {
+  const store = createStore(":memory:");
+  const service = createService({ store, mirror: null, config: {} });
+  const dream = createDreamScheduler({ onRun: () => Promise.resolve({ ok: true, skipped: true }) });
+  const { memory: a } = service.saveWithDedupe({ type: "project", title: "插件", content: "旧", importance: 3 });
+  const { memory: b } = service.saveWithDedupe({ type: "project", title: "插件2", content: "新细节", importance: 4 });
+  const captured = [];
+  const ctx = dreamCtx({
+    captured,
+    onConsolidation: () => JSON.stringify([
+      { action: "merge", ids: [a.id, b.id], keepSource: b.id, title: "插件总览", content: "合并内容", importance: 4 }
+    ])
+  });
+  const result = await dream.runDream(ctx, service, { dreamReasoningEffort: "off" });
+  assert.equal(result.ok, true);
+  assert.equal(captured.length, 2);
+  for (const options of captured) {
+    assert.equal(options.reasoningEffort, "off", `reasoningEffort:"off" forwarded on ${options.purpose}`);
+  }
+  store.close();
+});
+
 // ---------------------------------------------------------------- sleep passthrough
 
 function sleepSetup() {

package/test/reranker.test.js

@@ -1,240 +1,240 @@
-import test from "node:test";
-import assert from "node:assert/strict";
-import { spawnSync } from "node:child_process";
-import { fileURLToPath } from "node:url";
-import { LocalReranker } from "../src/reranker.js";
-
-/** Injected scorer: records (query, passage) calls, returns a fixed score. */
-function makeScorer(scoreOf) {
-  const calls = [];
-  const fn = async (query, passage) => {
-    calls.push({ query, passage });
-    return typeof scoreOf === "function" ? scoreOf(passage) : scoreOf;
-  };
-  return { fn, calls };
-}
-
-/** Fake feature-extraction engine: each input embeds to [1, k*0.25, 0, 0]. */
-function makeFakeExtractor(dim = 4, calls = []) {
-  let k = 0;
-  const fn = async (texts) => {
-    calls.push({ count: texts.length });
-    const rows = texts.map(() => {
-      const n = k++;
-      return Array.from({ length: dim }, (__, j) => (j === 0 ? 1 : j === 1 ? n * 0.25 : 0));
-    });
-    const flat = Float32Array.from(rows.flat());
-    return { data: flat, dims: [texts.length, dim] };
-  };
-  fn.dispose = () => {
-    fn.disposed = true;
-  };
-  return fn;
-}
-
-/** Fake text-classification engine exposing tokenizer + model for the tc path. */
-function makeFakeTc(logitsPerRow) {
-  const tokenizer = (texts, opts) => ({ texts, text_pair: opts.text_pair });
-  const model = async (inputs) => {
-    const n = inputs.texts.length;
-    const flat = new Float32Array(logitsPerRow.slice(0, n * 2));
-    return { logits: { data: flat, dims: [n, 2] } };
-  };
-  return { tokenizer, model, dispose: () => {} };
-}
-
-function candidates(ids) {
-  return ids.map((id) => ({ id, title: `T-${id}`, content: `C-${id}` }));
-}
-
-test("init with injected scorePair skips model loading", async () => {
-  let factoryCalled = false;
-  const { fn, calls } = makeScorer(0.7);
-  const r = new LocalReranker({
-    scorePair: fn,
-    engineFactory: async () => {
-      factoryCalled = true;
-      throw new Error("should never load");
-    }
-  });
-  await r.init();
-  assert.equal(factoryCalled, false);
-  const out = await r.rerank("q", candidates(["a"]));
-  assert.deepEqual(out, [{ id: "a", score: 0.7 }]);
-  assert.equal(calls.length, 1);
-  assert.equal(calls[0].query, "q");
-  assert.equal(calls[0].passage, "T-a\nC-a");
-});
-
-test("rerank returns results sorted by descending score", async () => {
-  const { fn } = makeScorer((p) => (p.includes("A") ? 0.9 : p.includes("B") ? 0.5 : 0.2));
-  const r = new LocalReranker({ scorePair: fn });
-  await r.init();
-  const out = await r.rerank("q", candidates(["A", "B", "C"]));
-  assert.deepEqual(out.map((x) => x.id), ["A", "B", "C"]);
-  assert.deepEqual(out.map((x) => x.score), [0.9, 0.5, 0.2]);
-});
-
-test("rerank returns [] for empty candidates", async () => {
-  const { fn } = makeScorer(0.9);
-  const r = new LocalReranker({ scorePair: fn });
-  await r.init();
-  assert.deepEqual(await r.rerank("q", []), []);
-});
-
-test("rerank drops candidates below scoreThreshold", async () => {
-  const { fn } = makeScorer((p) => (p.includes("A") ? 0.8 : p.includes("B") ? 0.4 : 0.05));
-  const r = new LocalReranker({ scorePair: fn, scoreThreshold: 0.3 });
-  await r.init();
-  const out = await r.rerank("q", candidates(["A", "B", "C"]));
-  assert.deepEqual(out.map((x) => x.id), ["A", "B"]);
-});
-
-test("rerank clamps scores into 0..1", async () => {
-  const { fn } = makeScorer((p) => (p.includes("A") ? 1.5 : p.includes("B") ? -0.2 : NaN));
-  const r = new LocalReranker({ scorePair: fn, scoreThreshold: 0 });
-  await r.init();
-  const out = await r.rerank("q", candidates(["A", "B", "C"]));
-  assert.deepEqual(out.map((x) => x.score), [1, 0, 0]);
-});
-
-test("rerank truncates to maxCandidates", async () => {
-  const { fn, calls } = makeScorer(0.6);
-  const r = new LocalReranker({ scorePair: fn, maxCandidates: 2 });
-  await r.init();
-  const out = await r.rerank("q", candidates(["a", "b", "c", "d"]));
-  assert.equal(calls.length, 2);
-  assert.equal(out.length, 2);
-});
-
-test("passage falls back to title when content is missing", async () => {
-  const { fn, calls } = makeScorer(0.8);
-  const r = new LocalReranker({ scorePair: fn });
-  await r.init();
-  await r.rerank("q", [{ id: "x", title: "only-title", content: "" }]);
-  assert.equal(calls[0].passage, "only-title");
-});
-
-test("rerank propagates injected scorer failures", async () => {
-  const bad = async () => {
-    throw new Error("scorer boom");
-  };
-  const r = new LocalReranker({ scorePair: bad });
-  await r.init();
-  await assert.rejects(() => r.rerank("q", candidates(["a"])), /scorer boom/);
-});
-
-test("rerank throws on non-array candidates", async () => {
-  const r = new LocalReranker({ scorePair: async () => 0.5 });
-  await r.init();
-  await assert.rejects(() => r.rerank("q", "not-an-array"), /array/);
-});
-
-test("init throws when no strategy can load", async () => {
-  const r = new LocalReranker({
-    engineFactory: async () => {
-      throw new Error("Unsupported pipeline");
-    }
-  });
-  await assert.rejects(() => r.init(), /LocalReranker failed to load/);
-});
-
-test("init cascades rerank -> tc -> feature-extraction and batches", async () => {
-  const calls = [];
-  const loader = async (task) => {
-    if (task === "feature-extraction") return makeFakeExtractor(4, calls);
-    throw new Error(`Unsupported pipeline: ${task}`);
-  };
-  const r = new LocalReranker({ engineFactory: loader, batchSize: 2 });
-  await r.init();
-  const out = await r.rerank("q", candidates(["a", "b", "c"]));
-  // Query k=0 [1,0,0,0]; a,b,c embed at k=1..3 -> cosine 0.970, 0.894, 0.8,
-  // all above the default threshold.
-  assert.equal(out.length, 3);
-  assert.deepEqual(out.map((x) => x.id), ["a", "b", "c"]);
-  assert.ok(out[0].score >= out[1].score && out[1].score >= out[2].score);
-  // Query embed (cached) + two batches of size 2 and 1.
-  assert.deepEqual(calls.map((c) => c.count), [1, 2, 1]);
-  // Query vector is cached: a second rerank only makes batch calls.
-  await r.rerank("q", candidates(["a", "b", "c"]));
-  assert.deepEqual(calls.map((c) => c.count), [1, 2, 1, 2, 1]);
-});
-
-test("text-classification strategy scores via logit delta", async () => {
-  const loader = async (task) => {
-    if (task === "text-classification") return makeFakeTc([0, 1, 0, 0, -2, 2]);
-    throw new Error(`Unsupported pipeline: ${task}`);
-  };
-  const r = new LocalReranker({ engineFactory: loader });
-  await r.init();
-  const out = await r.rerank("q", candidates(["A", "B", "C"]));
-  // sigmoid(1-0)=0.731, sigmoid(0)=0.5, sigmoid(2-(-2))=0.982.
-  assert.deepEqual(out.map((x) => x.id), ["C", "A", "B"]);
-  assert.ok(Math.abs(out[0].score - 0.982) < 1e-3);
-  assert.ok(Math.abs(out[1].score - 0.731) < 1e-3);
-  assert.ok(Math.abs(out[2].score - 0.5) < 1e-3);
-});
-
-test("modelHash follows the embedder convention", () => {
-  const a = new LocalReranker({ model: "Xenova/bge-reranker-base" });
-  const b = new LocalReranker({ model: "Xenova/bge-reranker-base" });
-  assert.equal(a.modelHash, b.modelHash);
-  assert.match(a.modelHash, /^Xenova\/bge-reranker-base#[0-9a-f]+$/);
-  assert.notEqual(a.modelHash, new LocalReranker({ model: "other/model" }).modelHash);
-});
-
-test("dispose releases the loaded pipeline", async () => {
-  const extractor = makeFakeExtractor(4);
-  const r = new LocalReranker({
-    engineFactory: async (task) => {
-      if (task === "feature-extraction") return extractor;
-      throw new Error("Unsupported pipeline");
-    }
-  });
-  await r.init();
-  assert.ok(r.pipeline);
-  r.dispose();
-  assert.equal(r.pipeline, null);
-  assert.equal(extractor.disposed, true);
-});
-
-test("default pipeline loader mirrors cache_dir onto env.cacheDir (issue #13)", () => {
-  // The fix lives in the module's *default* loader — the dynamic-import of
-  // @huggingface/transformers that the in-process tests bypass by injecting
-  // engineFactory. So it is exercised in a child node process under the
-  // --experimental-test-module-mocks flag: the transformers module is mocked
-  // with an empty env, LocalReranker uses the real default loader, and we
-  // assert env.cacheDir picks up the constructor's cache_dir. A regression
-  // (loader no longer mirroring) fails the child and surfaces here as a
-  // non-zero exit.
-  const srcUrl = new URL("../src/reranker.js", import.meta.url).href;
-  const cacheDir = "/tmp/dsh-mneme-cache-mirror-test";
-  const script = `
-    import { test } from "node:test";
-    const cacheDir = ${JSON.stringify(cacheDir)};
-    test("cache_dir is mirrored onto env.cacheDir", async (t) => {
-      t.mock.module("@huggingface/transformers", {
-        namedExports: {
-          env: {},
-          pipeline: async () => ({ dispose: () => {} })
-        }
-      });
-      const { LocalReranker } = await import(${JSON.stringify(srcUrl)});
-      const r = new LocalReranker({ cacheDir, device: "cpu" });
-      await r.init();
-      const { env } = await import("@huggingface/transformers");
-      if (env.cacheDir !== cacheDir) {
-        throw new Error("env.cacheDir not mirrored from cache_dir: " + env.cacheDir);
-      }
-      console.log("CACHE_MIRROR_OK");
-    });
-  `;
-  const res = spawnSync(process.execPath, [
-    "--experimental-test-module-mocks",
-    "--input-type=module",
-    "-e",
-    script
-  ], { encoding: "utf8", cwd: fileURLToPath(new URL("..", import.meta.url)) });
-  assert.equal(res.status, 0, `cache-dir mirror child failed:\n${res.stdout}\n${res.stderr}`);
-  assert.match(res.stdout, /CACHE_MIRROR_OK/);
-});
+import test from "node:test";
+import assert from "node:assert/strict";
+import { spawnSync } from "node:child_process";
+import { fileURLToPath } from "node:url";
+import { LocalReranker } from "../src/reranker.js";
+
+/** Injected scorer: records (query, passage) calls, returns a fixed score. */
+function makeScorer(scoreOf) {
+  const calls = [];
+  const fn = async (query, passage) => {
+    calls.push({ query, passage });
+    return typeof scoreOf === "function" ? scoreOf(passage) : scoreOf;
+  };
+  return { fn, calls };
+}
+
+/** Fake feature-extraction engine: each input embeds to [1, k*0.25, 0, 0]. */
+function makeFakeExtractor(dim = 4, calls = []) {
+  let k = 0;
+  const fn = async (texts) => {
+    calls.push({ count: texts.length });
+    const rows = texts.map(() => {
+      const n = k++;
+      return Array.from({ length: dim }, (__, j) => (j === 0 ? 1 : j === 1 ? n * 0.25 : 0));
+    });
+    const flat = Float32Array.from(rows.flat());
+    return { data: flat, dims: [texts.length, dim] };
+  };
+  fn.dispose = () => {
+    fn.disposed = true;
+  };
+  return fn;
+}
+
+/** Fake text-classification engine exposing tokenizer + model for the tc path. */
+function makeFakeTc(logitsPerRow) {
+  const tokenizer = (texts, opts) => ({ texts, text_pair: opts.text_pair });
+  const model = async (inputs) => {
+    const n = inputs.texts.length;
+    const flat = new Float32Array(logitsPerRow.slice(0, n * 2));
+    return { logits: { data: flat, dims: [n, 2] } };
+  };
+  return { tokenizer, model, dispose: () => {} };
+}
+
+function candidates(ids) {
+  return ids.map((id) => ({ id, title: `T-${id}`, content: `C-${id}` }));
+}
+
+test("init with injected scorePair skips model loading", async () => {
+  let factoryCalled = false;
+  const { fn, calls } = makeScorer(0.7);
+  const r = new LocalReranker({
+    scorePair: fn,
+    engineFactory: async () => {
+      factoryCalled = true;
+      throw new Error("should never load");
+    }
+  });
+  await r.init();
+  assert.equal(factoryCalled, false);
+  const out = await r.rerank("q", candidates(["a"]));
+  assert.deepEqual(out, [{ id: "a", score: 0.7 }]);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].query, "q");
+  assert.equal(calls[0].passage, "T-a\nC-a");
+});
+
+test("rerank returns results sorted by descending score", async () => {
+  const { fn } = makeScorer((p) => (p.includes("A") ? 0.9 : p.includes("B") ? 0.5 : 0.2));
+  const r = new LocalReranker({ scorePair: fn });
+  await r.init();
+  const out = await r.rerank("q", candidates(["A", "B", "C"]));
+  assert.deepEqual(out.map((x) => x.id), ["A", "B", "C"]);
+  assert.deepEqual(out.map((x) => x.score), [0.9, 0.5, 0.2]);
+});
+
+test("rerank returns [] for empty candidates", async () => {
+  const { fn } = makeScorer(0.9);
+  const r = new LocalReranker({ scorePair: fn });
+  await r.init();
+  assert.deepEqual(await r.rerank("q", []), []);
+});
+
+test("rerank drops candidates below scoreThreshold", async () => {
+  const { fn } = makeScorer((p) => (p.includes("A") ? 0.8 : p.includes("B") ? 0.4 : 0.05));
+  const r = new LocalReranker({ scorePair: fn, scoreThreshold: 0.3 });
+  await r.init();
+  const out = await r.rerank("q", candidates(["A", "B", "C"]));
+  assert.deepEqual(out.map((x) => x.id), ["A", "B"]);
+});
+
+test("rerank clamps scores into 0..1", async () => {
+  const { fn } = makeScorer((p) => (p.includes("A") ? 1.5 : p.includes("B") ? -0.2 : NaN));
+  const r = new LocalReranker({ scorePair: fn, scoreThreshold: 0 });
+  await r.init();
+  const out = await r.rerank("q", candidates(["A", "B", "C"]));
+  assert.deepEqual(out.map((x) => x.score), [1, 0, 0]);
+});
+
+test("rerank truncates to maxCandidates", async () => {
+  const { fn, calls } = makeScorer(0.6);
+  const r = new LocalReranker({ scorePair: fn, maxCandidates: 2 });
+  await r.init();
+  const out = await r.rerank("q", candidates(["a", "b", "c", "d"]));
+  assert.equal(calls.length, 2);
+  assert.equal(out.length, 2);
+});
+
+test("passage falls back to title when content is missing", async () => {
+  const { fn, calls } = makeScorer(0.8);
+  const r = new LocalReranker({ scorePair: fn });
+  await r.init();
+  await r.rerank("q", [{ id: "x", title: "only-title", content: "" }]);
+  assert.equal(calls[0].passage, "only-title");
+});
+
+test("rerank propagates injected scorer failures", async () => {
+  const bad = async () => {
+    throw new Error("scorer boom");
+  };
+  const r = new LocalReranker({ scorePair: bad });
+  await r.init();
+  await assert.rejects(() => r.rerank("q", candidates(["a"])), /scorer boom/);
+});
+
+test("rerank throws on non-array candidates", async () => {
+  const r = new LocalReranker({ scorePair: async () => 0.5 });
+  await r.init();
+  await assert.rejects(() => r.rerank("q", "not-an-array"), /array/);
+});
+
+test("init throws when no strategy can load", async () => {
+  const r = new LocalReranker({
+    engineFactory: async () => {
+      throw new Error("Unsupported pipeline");
+    }
+  });
+  await assert.rejects(() => r.init(), /LocalReranker failed to load/);
+});
+
+test("init cascades rerank -> tc -> feature-extraction and batches", async () => {
+  const calls = [];
+  const loader = async (task) => {
+    if (task === "feature-extraction") return makeFakeExtractor(4, calls);
+    throw new Error(`Unsupported pipeline: ${task}`);
+  };
+  const r = new LocalReranker({ engineFactory: loader, batchSize: 2 });
+  await r.init();
+  const out = await r.rerank("q", candidates(["a", "b", "c"]));
+  // Query k=0 [1,0,0,0]; a,b,c embed at k=1..3 -> cosine 0.970, 0.894, 0.8,
+  // all above the default threshold.
+  assert.equal(out.length, 3);
+  assert.deepEqual(out.map((x) => x.id), ["a", "b", "c"]);
+  assert.ok(out[0].score >= out[1].score && out[1].score >= out[2].score);
+  // Query embed (cached) + two batches of size 2 and 1.
+  assert.deepEqual(calls.map((c) => c.count), [1, 2, 1]);
+  // Query vector is cached: a second rerank only makes batch calls.
+  await r.rerank("q", candidates(["a", "b", "c"]));
+  assert.deepEqual(calls.map((c) => c.count), [1, 2, 1, 2, 1]);
+});
+
+test("text-classification strategy scores via logit delta", async () => {
+  const loader = async (task) => {
+    if (task === "text-classification") return makeFakeTc([0, 1, 0, 0, -2, 2]);
+    throw new Error(`Unsupported pipeline: ${task}`);
+  };
+  const r = new LocalReranker({ engineFactory: loader });
+  await r.init();
+  const out = await r.rerank("q", candidates(["A", "B", "C"]));
+  // sigmoid(1-0)=0.731, sigmoid(0)=0.5, sigmoid(2-(-2))=0.982.
+  assert.deepEqual(out.map((x) => x.id), ["C", "A", "B"]);
+  assert.ok(Math.abs(out[0].score - 0.982) < 1e-3);
+  assert.ok(Math.abs(out[1].score - 0.731) < 1e-3);
+  assert.ok(Math.abs(out[2].score - 0.5) < 1e-3);
+});
+
+test("modelHash follows the embedder convention", () => {
+  const a = new LocalReranker({ model: "Xenova/bge-reranker-base" });
+  const b = new LocalReranker({ model: "Xenova/bge-reranker-base" });
+  assert.equal(a.modelHash, b.modelHash);
+  assert.match(a.modelHash, /^Xenova\/bge-reranker-base#[0-9a-f]+$/);
+  assert.notEqual(a.modelHash, new LocalReranker({ model: "other/model" }).modelHash);
+});
+
+test("dispose releases the loaded pipeline", async () => {
+  const extractor = makeFakeExtractor(4);
+  const r = new LocalReranker({
+    engineFactory: async (task) => {
+      if (task === "feature-extraction") return extractor;
+      throw new Error("Unsupported pipeline");
+    }
+  });
+  await r.init();
+  assert.ok(r.pipeline);
+  r.dispose();
+  assert.equal(r.pipeline, null);
+  assert.equal(extractor.disposed, true);
+});
+
+test("default pipeline loader mirrors cache_dir onto env.cacheDir (issue #13)", () => {
+  // The fix lives in the module's *default* loader — the dynamic-import of
+  // @huggingface/transformers that the in-process tests bypass by injecting
+  // engineFactory. So it is exercised in a child node process under the
+  // --experimental-test-module-mocks flag: the transformers module is mocked
+  // with an empty env, LocalReranker uses the real default loader, and we
+  // assert env.cacheDir picks up the constructor's cache_dir. A regression
+  // (loader no longer mirroring) fails the child and surfaces here as a
+  // non-zero exit.
+  const srcUrl = new URL("../src/reranker.js", import.meta.url).href;
+  const cacheDir = "/tmp/dsh-mneme-cache-mirror-test";
+  const script = `
+    import { test } from "node:test";
+    const cacheDir = ${JSON.stringify(cacheDir)};
+    test("cache_dir is mirrored onto env.cacheDir", async (t) => {
+      t.mock.module("@huggingface/transformers", {
+        namedExports: {
+          env: {},
+          pipeline: async () => ({ dispose: () => {} })
+        }
+      });
+      const { LocalReranker } = await import(${JSON.stringify(srcUrl)});
+      const r = new LocalReranker({ cacheDir, device: "cpu" });
+      await r.init();
+      const { env } = await import("@huggingface/transformers");
+      if (env.cacheDir !== cacheDir) {
+        throw new Error("env.cacheDir not mirrored from cache_dir: " + env.cacheDir);
+      }
+      console.log("CACHE_MIRROR_OK");
+    });
+  `;
+  const res = spawnSync(process.execPath, [
+    "--experimental-test-module-mocks",
+    "--input-type=module",
+    "-e",
+    script
+  ], { encoding: "utf8", cwd: fileURLToPath(new URL("..", import.meta.url)) });
+  assert.equal(res.status, 0, `cache-dir mirror child failed:\n${res.stdout}\n${res.stderr}`);
+  assert.match(res.stdout, /CACHE_MIRROR_OK/);
+});