@remnic/core 1.1.12 → 1.1.14

package/src/recall-mmr.test.ts

@@ -0,0 +1,808 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import {
+  applyMmrToCandidates,
+  reorderRecallResultsWithMmr,
+  summarizeMmrDiversity,
+  normalizeTokens,
+  type MmrCandidate,
+  type MmrRecallResult,
+} from "./recall-mmr.js";
+
+function makeCandidate(
+  id: string,
+  content: string,
+  score: number,
+  embedding?: number[],
+): MmrCandidate {
+  return { id, content, score, embedding: embedding ?? null };
+}
+
+test("applyMmrToCandidates returns [] for empty input", () => {
+  const out = applyMmrToCandidates({ candidates: [] });
+  assert.deepEqual(out, []);
+});
+
+test("applyMmrToCandidates is a no-op for a single candidate", () => {
+  const input = [makeCandidate("a", "hello", 1)];
+  const out = applyMmrToCandidates({ candidates: input });
+  assert.equal(out.length, 1);
+  assert.equal(out[0]?.id, "a");
+});
+
+test("applyMmrToCandidates does not mutate input", () => {
+  const input: MmrCandidate[] = [
+    makeCandidate("a", "one", 0.9, [1, 0, 0, 0]),
+    makeCandidate("b", "two", 0.8, [0, 1, 0, 0]),
+  ];
+  const before = input.map((c) => c.id).join(",");
+  applyMmrToCandidates({ candidates: input, budget: 1 });
+  assert.equal(
+    input.map((c) => c.id).join(","),
+    before,
+    "input order should be unchanged",
+  );
+});
+
+test("applyMmrToCandidates collapses near-duplicate clusters (cosine)", () => {
+  // Five near-duplicate candidates all pointing toward [1,0,0,0] with tiny
+  // perturbations, plus four diverse candidates along orthogonal axes.
+  const epsilon = 0.001;
+  const nearDupe = (idx: number): number[] => [1, epsilon * idx, 0, 0];
+  const candidates: MmrCandidate[] = [
+    makeCandidate("dupe-1", "alpha one", 0.99, nearDupe(1)),
+    makeCandidate("dupe-2", "alpha two", 0.98, nearDupe(2)),
+    makeCandidate("dupe-3", "alpha three", 0.97, nearDupe(3)),
+    makeCandidate("dupe-4", "alpha four", 0.96, nearDupe(4)),
+    makeCandidate("dupe-5", "alpha five", 0.95, nearDupe(5)),
+    makeCandidate("div-y", "beta", 0.9, [0, 1, 0, 0]),
+    makeCandidate("div-z", "gamma", 0.85, [0, 0, 1, 0]),
+    makeCandidate("div-w", "delta", 0.8, [0, 0, 0, 1]),
+    makeCandidate("div-a", "epsilon", 0.75, [1, 1, 0, 0]),
+  ];
+
+  // Budget = 4 — with 4 distinct axes available (plus a 5th dupe cluster),
+  // MMR should pick 1 dupe + 3 diverse before ever revisiting the cluster.
+  const selected = applyMmrToCandidates({
+    candidates,
+    lambda: 0.5, // balance relevance + diversity
+    topN: 40,
+    budget: 4,
+  });
+
+  assert.equal(selected.length, 4);
+
+  const dupeCount = selected.filter((c) => c.id.startsWith("dupe-")).length;
+  assert.ok(
+    dupeCount <= 1,
+    `expected at most 1 representative from the duplicate cluster, got ${dupeCount}: ${selected
+      .map((c) => c.id)
+      .join(",")}`,
+  );
+
+  // The highest-relevance duplicate should still be included.
+  const ids = selected.map((c) => c.id);
+  assert.ok(ids.includes("dupe-1"), `expected dupe-1 in selection: ${ids}`);
+  // All three fully diverse axes should also surface.
+  assert.ok(ids.includes("div-y"));
+  assert.ok(ids.includes("div-z"));
+  assert.ok(ids.includes("div-w"));
+});
+
+test("applyMmrToCandidates falls back to Jaccard when embeddings are missing", () => {
+  // Three candidates with heavy lexical overlap + two diverse ones.
+  const candidates: MmrCandidate[] = [
+    makeCandidate(
+      "s1",
+      "The cat sat on the red mat in the sunny kitchen",
+      0.95,
+    ),
+    makeCandidate(
+      "s2",
+      "The cat sat on the red mat in the sunny kitchen today",
+      0.94,
+    ),
+    makeCandidate(
+      "s3",
+      "A cat is sitting on the red mat in the kitchen",
+      0.93,
+    ),
+    makeCandidate(
+      "d1",
+      "Quantum tunneling explains alpha decay in heavy nuclei",
+      0.5,
+    ),
+    makeCandidate(
+      "d2",
+      "Bicycle gears shift when pedaling uphill on pavement",
+      0.4,
+    ),
+  ];
+
+  const selected = applyMmrToCandidates({
+    candidates,
+    lambda: 0.4, // push harder toward diversity so clustering shows up
+    topN: 40,
+    budget: 3,
+  });
+
+  assert.equal(selected.length, 3);
+  const ids = selected.map((c) => c.id);
+  // The top-scored lexical duplicate should be included...
+  assert.ok(ids.includes("s1"));
+  // ...but MMR should prefer the diverse candidates over the other near-dupes.
+  const sxCount = ids.filter((id) => id.startsWith("s")).length;
+  assert.ok(
+    sxCount < 3,
+    `MMR should not fill the slate with lexical duplicates: ${ids.join(",")}`,
+  );
+  assert.ok(
+    ids.includes("d1") || ids.includes("d2"),
+    `at least one diverse candidate should appear: ${ids.join(",")}`,
+  );
+});
+
+test("applyMmrToCandidates is sensitive to lambda", () => {
+  const candidates: MmrCandidate[] = [
+    makeCandidate("a", "alpha one", 1.0, [1, 0.001, 0, 0]),
+    makeCandidate("b", "alpha two", 0.99, [1, 0.002, 0, 0]),
+    makeCandidate("c", "beta", 0.5, [0, 1, 0, 0]),
+    makeCandidate("d", "gamma", 0.4, [0, 0, 1, 0]),
+  ];
+
+  // Lambda = 1.0 → pure relevance, no diversity adjustment.
+  const pureRelevance = applyMmrToCandidates({
+    candidates,
+    lambda: 1.0,
+    budget: 2,
+  });
+  assert.deepEqual(pureRelevance.map((c) => c.id), ["a", "b"]);
+
+  // Lambda = 0.2 → diversity dominates, the second pick should avoid the
+  // near-duplicate of `a`.
+  const diverse = applyMmrToCandidates({
+    candidates,
+    lambda: 0.2,
+    budget: 2,
+  });
+  assert.equal(diverse[0]?.id, "a", "highest-relevance should still lead");
+  assert.notEqual(
+    diverse[1]?.id,
+    "b",
+    "diversity-heavy MMR should avoid the near-duplicate",
+  );
+});
+
+test("applyMmrToCandidates preserves all candidates when budget >= length", () => {
+  const candidates: MmrCandidate[] = [
+    makeCandidate("a", "one", 1, [1, 0, 0]),
+    makeCandidate("b", "two", 0.9, [0, 1, 0]),
+    makeCandidate("c", "three", 0.8, [0, 0, 1]),
+  ];
+  const out = applyMmrToCandidates({
+    candidates,
+    budget: 10,
+  });
+  assert.equal(out.length, 3);
+  const ids = out.map((c) => c.id).sort();
+  assert.deepEqual(ids, ["a", "b", "c"]);
+});
+
+test("applyMmrToCandidates respects topN clamping", () => {
+  const candidates: MmrCandidate[] = Array.from({ length: 50 }, (_, i) =>
+    makeCandidate(`c-${i}`, `content ${i}`, 1 - i * 0.01, [1, i * 0.001, 0]),
+  );
+  const out = applyMmrToCandidates({
+    candidates,
+    topN: 10,
+    budget: 5,
+  });
+  assert.equal(out.length, 5);
+  // All selected should come from the top 10 (indices 0-9).
+  for (const c of out) {
+    const idx = Number.parseInt(c.id.split("-")[1]!, 10);
+    assert.ok(idx < 10, `selected candidate ${c.id} outside topN window`);
+  }
+});
+
+test("applyMmrToCandidates handles budget=0 as empty selection", () => {
+  const candidates = [makeCandidate("a", "one", 1), makeCandidate("b", "two", 0.5)];
+  const out = applyMmrToCandidates({ candidates, budget: 0 });
+  // budget=0 is clamped to 0, nothing selected
+  assert.equal(out.length, 0);
+});
+
+test("summarizeMmrDiversity reports higher before-sim for dup-heavy input", () => {
+  const dupHeavy: MmrCandidate[] = [
+    makeCandidate("a", "the cat sat on the mat", 1, [1, 0.001, 0]),
+    makeCandidate("b", "a cat sat on the mat", 0.99, [1, 0.002, 0]),
+    makeCandidate("c", "cats sit on mats", 0.98, [1, 0.003, 0]),
+  ];
+  const diversified = applyMmrToCandidates({
+    candidates: dupHeavy.concat([
+      makeCandidate("d", "rocket thrust vectoring", 0.5, [0, 1, 0]),
+      makeCandidate("e", "violin bow rosin", 0.4, [0, 0, 1]),
+    ]),
+    lambda: 0.4,
+    budget: 3,
+  });
+
+  const report = summarizeMmrDiversity(dupHeavy, diversified, 40);
+  assert.ok(
+    report.avgPairwiseSimAfter <= report.avgPairwiseSimBefore + 1e-9,
+    `avg pairwise similarity should not increase after MMR: before=${report.avgPairwiseSimBefore} after=${report.avgPairwiseSimAfter}`,
+  );
+});
+
+test("normalizeTokens lowercases and strips punctuation", () => {
+  assert.deepEqual(
+    [...normalizeTokens("Hello, World! 42 TIMES.")],
+    ["hello", "world", "42", "times"],
+  );
+  assert.deepEqual([...normalizeTokens("")], []);
+  assert.deepEqual([...normalizeTokens("   ")], []);
+});
+
+test("applyMmrToCandidates end-to-end: duplicate sentences reduced vs disabled", () => {
+  // Small seeded fixture: 4 near-duplicate "fact" sentences + 3 diverse.
+  const fixture: MmrCandidate[] = [
+    makeCandidate(
+      "f1",
+      "The user prefers dark mode for their terminal",
+      0.95,
+    ),
+    makeCandidate(
+      "f2",
+      "User likes dark mode in the terminal interface",
+      0.94,
+    ),
+    makeCandidate(
+      "f3",
+      "Prefers dark mode setting for terminal usage",
+      0.93,
+    ),
+    makeCandidate(
+      "f4",
+      "The terminal should use dark mode",
+      0.92,
+    ),
+    makeCandidate("d1", "Coffee is brewed at 195 degrees fahrenheit", 0.8),
+    makeCandidate("d2", "The sunset reflected orange on the ocean waves", 0.75),
+    makeCandidate("d3", "Linear algebra underpins modern machine learning", 0.7),
+  ];
+
+  // Without MMR: top-4 would be f1..f4, all near-duplicates.
+  const noMmrTop4 = fixture.slice(0, 4);
+  const mmrTop4 = applyMmrToCandidates({
+    candidates: fixture,
+    lambda: 0.4,
+    budget: 4,
+  });
+
+  const noMmrIds = noMmrTop4.map((c) => c.id);
+  const mmrIds = mmrTop4.map((c) => c.id);
+
+  const countDupes = (ids: string[]) =>
+    ids.filter((id) => id.startsWith("f")).length;
+
+  assert.ok(
+    countDupes(mmrIds) < countDupes(noMmrIds),
+    `MMR should strictly reduce duplicate cluster count: noMmr=${noMmrIds} mmr=${mmrIds}`,
+  );
+  assert.ok(
+    mmrIds.some((id) => id.startsWith("d")),
+    `MMR should surface at least one diverse candidate: ${mmrIds}`,
+  );
+});
+
+// ---------------------------------------------------------------------------
+// Regression: path-first keying for reorderRecallResultsWithMmr
+// (ChatGPT Codex P2 review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "reorderRecallResultsWithMmr preserves distinct results that share a docid",
+  () => {
+    // Two results with IDENTICAL docids but DIFFERENT paths. A docid-keyed
+    // implementation would collapse these into one and silently drop a valid
+    // recall candidate.
+    const results: MmrRecallResult[] = [
+      {
+        docid: "fact-007",
+        path: "memories/facts/source-a/fact-007.md",
+        snippet: "Some fact from source A",
+        score: 0.95,
+      },
+      {
+        docid: "fact-007",
+        path: "memories/facts/source-b/fact-007.md",
+        snippet: "A very different fact from source B",
+        score: 0.92,
+      },
+      {
+        docid: "fact-008",
+        path: "memories/facts/source-a/fact-008.md",
+        snippet: "Yet another unrelated fact",
+        score: 0.9,
+      },
+    ];
+
+    const { reordered } = reorderRecallResultsWithMmr(results);
+
+    assert.equal(
+      reordered.length,
+      results.length,
+      "no candidate should be dropped when docids collide across paths",
+    );
+    const paths = reordered.map((r) => r.path).sort();
+    assert.deepEqual(paths, [
+      "memories/facts/source-a/fact-007.md",
+      "memories/facts/source-a/fact-008.md",
+      "memories/facts/source-b/fact-007.md",
+    ]);
+  },
+);
+
+test(
+  "reorderRecallResultsWithMmr keeps results with empty path AND empty docid distinct",
+  () => {
+    // Pathological input: every result has empty path and empty docid. The
+    // stable index-suffix fallback must still give each candidate its own
+    // key so none of them collapse.
+    const results: MmrRecallResult[] = [
+      { docid: "", path: "", snippet: "alpha content", score: 0.9 },
+      { docid: "", path: "", snippet: "beta content", score: 0.8 },
+      { docid: "", path: "", snippet: "gamma content", score: 0.7 },
+    ];
+
+    const { reordered } = reorderRecallResultsWithMmr(results);
+    assert.equal(reordered.length, 3);
+    const snippets = reordered.map((r) => r.snippet).sort();
+    assert.deepEqual(snippets, ["alpha content", "beta content", "gamma content"]);
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: head-of-list diversity metric actually reflects reordering
+// (Cursor Bugbot Medium review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "summarizeMmrDiversity with small sample size reflects head-of-list reordering",
+  () => {
+    // Input: 5 near-duplicate high-score "a*" candidates up front, then 3
+    // diverse candidates. With `budget = candidates.length`, MMR reorders but
+    // drops nothing — the two slices pre- and post-MMR contain the same
+    // SET. If the diversity metric used a large sample size it would report
+    // identical before/after averages. With a small head-of-list sample size
+    // the head *does* differ, so avgPairwiseSimAfter < avgPairwiseSimBefore.
+    const candidates: MmrCandidate[] = [
+      { id: "a1", content: "alpha", score: 0.99, embedding: [1, 0.001, 0] },
+      { id: "a2", content: "alpha", score: 0.98, embedding: [1, 0.002, 0] },
+      { id: "a3", content: "alpha", score: 0.97, embedding: [1, 0.003, 0] },
+      { id: "a4", content: "alpha", score: 0.96, embedding: [1, 0.004, 0] },
+      { id: "a5", content: "alpha", score: 0.95, embedding: [1, 0.005, 0] },
+      { id: "b1", content: "beta", score: 0.9, embedding: [0, 1, 0] },
+      { id: "b2", content: "gamma", score: 0.85, embedding: [0, 0, 1] },
+      { id: "b3", content: "delta", score: 0.8, embedding: [1, 1, 0] },
+    ];
+
+    const reordered = applyMmrToCandidates({
+      candidates,
+      lambda: 0.3, // push hard toward diversity
+      budget: candidates.length, // same set, just reordered — same bug regime
+    });
+    assert.equal(reordered.length, candidates.length);
+
+    // Default small sample size (10 is larger than our 8-element pool, so
+    // everything is compared — this recreates the bug the comment flagged).
+    const trivialReport = summarizeMmrDiversity(
+      candidates,
+      reordered,
+      candidates.length,
+    );
+    // With full-pool sampling the before/after average pairwise similarity
+    // is order-independent — the bug condition.
+    assert.ok(
+      Math.abs(trivialReport.avgPairwiseSimBefore - trivialReport.avgPairwiseSimAfter) < 1e-9,
+      "sanity: full-pool sampling is order-independent and hides diversity gains",
+    );
+
+    // With a small head-of-list sample size (4), MMR's reordering *must*
+    // show up as a strictly lower average pairwise similarity.
+    const headReport = summarizeMmrDiversity(candidates, reordered, 4);
+    assert.ok(
+      headReport.avgPairwiseSimAfter + 1e-9 < headReport.avgPairwiseSimBefore,
+      `head-of-list sample should show MMR diversity gain: before=${headReport.avgPairwiseSimBefore} after=${headReport.avgPairwiseSimAfter}`,
+    );
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: MMR must run before truncation so diverse candidates just below
+// the cutoff can be promoted into the final injected set.
+// (ChatGPT Codex P2 review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "reorderRecallResultsWithMmr promotes sub-cutoff diverse candidates when run pre-slice",
+  () => {
+    // Simulate a near-duplicate cluster of 5 high-score candidates followed
+    // by a truly diverse candidate at position 5 (just below a final limit
+    // of 5). If MMR ran only on the already-truncated top-5, the diverse
+    // candidate would never make it into the injected set. Running MMR on
+    // the full 6-candidate pool and then slicing to 5 must pull the diverse
+    // candidate into the final slice.
+    const results: MmrRecallResult[] = [
+      { docid: "a1", path: "p/a1", snippet: "alpha fact one", score: 0.99 },
+      { docid: "a2", path: "p/a2", snippet: "alpha fact two", score: 0.98 },
+      { docid: "a3", path: "p/a3", snippet: "alpha fact three", score: 0.97 },
+      { docid: "a4", path: "p/a4", snippet: "alpha fact four", score: 0.96 },
+      { docid: "a5", path: "p/a5", snippet: "alpha fact five", score: 0.95 },
+      {
+        docid: "d1",
+        path: "p/d1",
+        snippet: "orthogonal concept rocket fuel chemistry",
+        score: 0.94,
+      },
+    ];
+
+    // Simulate the buggy "slice then MMR" behavior.
+    const sliceFirst = results.slice(0, 5);
+    const sliceFirstMmr = reorderRecallResultsWithMmr(sliceFirst, {
+      lambda: 0.3,
+    }).reordered;
+    const sliceFirstIds = sliceFirstMmr.map((r) => r.docid);
+    assert.ok(
+      !sliceFirstIds.includes("d1"),
+      "sanity: slicing before MMR cannot recover the sub-cutoff diverse candidate",
+    );
+
+    // Now the correct "MMR then slice" behavior.
+    const mmrFirst = reorderRecallResultsWithMmr(results, {
+      lambda: 0.3,
+    }).reordered.slice(0, 5);
+    const mmrFirstIds = mmrFirst.map((r) => r.docid);
+    assert.equal(mmrFirst.length, 5);
+    assert.ok(
+      mmrFirstIds.includes("d1"),
+      `MMR-then-slice should promote the diverse candidate into the top-5: ${mmrFirstIds.join(",")}`,
+    );
+  },
+);
+
+// NOTE: the orchestrator's `diversifyAndLimitRecallResults` helper is a
+// private method, but its zero-limit behavior is covered by the pure
+// equivalent below: when the caller wants zero results, MMR plus a slice of
+// zero should yield an empty array. This mirrors the legacy
+// `.slice(0, recallResultLimit)` semantics when `recallResultLimit === 0`
+// (which happens when `memoriesSectionEnabled` is false). Cursor Bugbot flagged
+// this regression on PR #391.
+
+test(
+  "MMR + slice(0) idiom yields empty array when limit is zero",
+  () => {
+    const results: MmrRecallResult[] = [
+      { docid: "a", path: "p/a", snippet: "one", score: 0.9 },
+      { docid: "b", path: "p/b", snippet: "two", score: 0.8 },
+    ];
+    const { reordered } = reorderRecallResultsWithMmr(results);
+    const limited = reordered.slice(0, 0);
+    assert.equal(
+      limited.length,
+      0,
+      "slice(0, 0) must still yield an empty array after MMR",
+    );
+  },
+);
+
+test(
+  "reorderRecallResultsWithMmr with topN=0 is a full no-op preserving input order",
+  () => {
+    // AGENTS.md §4: never coerce zero limits to non-zero. A topN of 0 means
+    // "apply MMR over an empty window" — no candidate gets dropped and the
+    // original order is preserved.
+    const results: MmrRecallResult[] = [
+      { docid: "a", path: "p/a", snippet: "one", score: 0.9 },
+      { docid: "b", path: "p/b", snippet: "two", score: 0.8 },
+      { docid: "c", path: "p/c", snippet: "three", score: 0.7 },
+    ];
+    const { reordered } = reorderRecallResultsWithMmr(results, { topN: 0 });
+    assert.equal(reordered.length, results.length);
+    assert.deepEqual(
+      reordered.map((r) => r.docid),
+      ["a", "b", "c"],
+      "topN=0 should be a no-op and preserve input order",
+    );
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: Unicode-aware Jaccard tokenization for multilingual snippets
+// (ChatGPT Codex P2 review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "normalizeTokens preserves non-ASCII scripts for multilingual Jaccard",
+  () => {
+    // Cyrillic: two near-identical snippets must share tokens, not collapse
+    // to an empty set like the old [^a-z0-9]+ tokenizer did.
+    const ruA = normalizeTokens("Привет, мир! Это тест.");
+    const ruB = normalizeTokens("Привет мир это тест");
+    assert.ok(ruA.size > 0, "Cyrillic snippet should produce non-empty tokens");
+    let ruOverlap = 0;
+    for (const t of ruA) if (ruB.has(t)) ruOverlap += 1;
+    assert.ok(
+      ruOverlap >= 3,
+      `Cyrillic near-duplicates should share most tokens, overlap=${ruOverlap}`,
+    );
+
+    // Greek: same near-duplicate invariant.
+    const grA = normalizeTokens("Καλημέρα κόσμε");
+    const grB = normalizeTokens("καλημέρα κόσμε!");
+    assert.ok(grA.size > 0);
+    let grOverlap = 0;
+    for (const t of grA) if (grB.has(t)) grOverlap += 1;
+    assert.ok(
+      grOverlap >= 2,
+      `Greek near-duplicates should share tokens, overlap=${grOverlap}`,
+    );
+
+    // CJK (Chinese): no whitespace word breaks, so tokens are per-character.
+    // Two snippets that share most characters must produce overlapping token
+    // sets, not the empty set the old tokenizer returned.
+    const zhA = normalizeTokens("用户喜欢深色模式");
+    const zhB = normalizeTokens("用户偏好深色模式");
+    assert.ok(zhA.size > 0, "CJK snippet should produce non-empty tokens");
+    let zhOverlap = 0;
+    for (const t of zhA) if (zhB.has(t)) zhOverlap += 1;
+    assert.ok(
+      zhOverlap >= 5,
+      `Chinese near-duplicates should share most chars, overlap=${zhOverlap}`,
+    );
+  },
+);
+
+test(
+  "reorderRecallResultsWithMmr detects non-ASCII near-duplicates and diversifies",
+  () => {
+    // Without Unicode-aware tokenization, the Jaccard fallback returned 0 for
+    // all non-Latin snippets and MMR degenerated to plain relevance ordering.
+    // After the fix the duplicate Cyrillic cluster must give up at least one
+    // head-of-list slot to the diverse candidate.
+    const results: MmrRecallResult[] = [
+      { docid: "ru1", path: "p/ru1", snippet: "Пользователь предпочитает тёмный режим в терминале", score: 0.99 },
+      { docid: "ru2", path: "p/ru2", snippet: "Пользователь предпочитает тёмный режим в терминалах", score: 0.98 },
+      { docid: "ru3", path: "p/ru3", snippet: "Пользователь любит тёмный режим в терминале", score: 0.97 },
+      { docid: "ru4", path: "p/ru4", snippet: "Пользователь предпочитает тёмную тему в терминале", score: 0.96 },
+      { docid: "dv1", path: "p/dv1", snippet: "Ракетное топливо на основе жидкого кислорода", score: 0.9 },
+    ];
+    const { reordered, diversity } = reorderRecallResultsWithMmr(results, {
+      lambda: 0.3,
+      diversitySampleSize: 4,
+    });
+    const headIds = reordered.slice(0, 4).map((r) => r.docid);
+    assert.ok(
+      headIds.includes("dv1"),
+      `diverse Cyrillic candidate should be promoted into head: ${headIds.join(",")}`,
+    );
+    assert.ok(
+      diversity.headReorderCount > 0,
+      `headReorderCount should be > 0 after MMR promoted a diverse candidate, got ${diversity.headReorderCount}`,
+    );
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: min-max normalization used to defeat MMR for tight score
+// clusters. After the fix, scores that already live in [0, 1] pass through
+// untouched so the (1 - lambda) * maxSim penalty can actually outweigh a
+// 0.01 relevance gap.
+// (Cursor Bugbot Medium review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "applyMmrToCandidates diversifies tight reranker score clusters",
+  () => {
+    // Three near-duplicates tightly clustered in [0.93, 0.95], plus one
+    // genuinely diverse candidate slightly below at 0.92. The old min-max
+    // normalization would scale the diverse candidate's relevance to 0.0 and
+    // it could never beat any duplicate; after the fix it must be promoted.
+    const candidates: MmrCandidate[] = [
+      { id: "dup1", content: "alpha content", score: 0.95, embedding: [1, 0.001, 0] },
+      { id: "dup2", content: "alpha content near", score: 0.94, embedding: [1, 0.002, 0] },
+      { id: "dup3", content: "alpha content close", score: 0.93, embedding: [1, 0.003, 0] },
+      { id: "div", content: "completely orthogonal", score: 0.92, embedding: [0, 1, 0] },
+    ];
+    const reordered = applyMmrToCandidates({
+      candidates,
+      lambda: 0.7,
+      budget: 2,
+    });
+    const ids = reordered.map((c) => c.id);
+    assert.equal(reordered.length, 2);
+    assert.ok(
+      ids.includes("div"),
+      `tight-cluster MMR should promote diverse candidate into top-2: ${ids.join(",")}`,
+    );
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: diversity metric must emit an actionable signal even in
+// reorder-only mode. `kept/considered` alone is unactionable when MMR never
+// drops candidates; the new `headReorderCount` field makes the metric useful.
+// (Cursor Bugbot Low-severity review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "summarizeMmrDiversity emits headReorderCount > 0 when MMR reorders the head",
+  () => {
+    // Near-dup cluster up front plus diverse candidates behind — MMR with
+    // low lambda must swap at least one head-of-list slot, producing a
+    // positive headReorderCount signal.
+    const candidates: MmrCandidate[] = [
+      { id: "a1", content: "dup alpha one", score: 0.99, embedding: [1, 0, 0] },
+      { id: "a2", content: "dup alpha two", score: 0.98, embedding: [1, 0.001, 0] },
+      { id: "a3", content: "dup alpha three", score: 0.97, embedding: [1, 0.002, 0] },
+      { id: "b1", content: "diverse beta", score: 0.8, embedding: [0, 1, 0] },
+      { id: "b2", content: "diverse gamma", score: 0.75, embedding: [0, 0, 1] },
+    ];
+    const reordered = applyMmrToCandidates({
+      candidates,
+      lambda: 0.3,
+      budget: candidates.length,
+    });
+    const report = summarizeMmrDiversity(candidates, reordered, 3);
+    assert.equal(report.considered, candidates.length);
+    assert.equal(report.kept, reordered.length);
+    assert.ok(
+      report.headReorderCount > 0,
+      `headReorderCount should be > 0 after MMR swapped head slots, got ${report.headReorderCount}`,
+    );
+  },
+);
+
+test(
+  "summarizeMmrDiversity headReorderCount is 0 for an identity reorder",
+  () => {
+    // If MMR leaves the head untouched (e.g. all candidates are already
+    // maximally diverse), headReorderCount should be exactly 0 so callers
+    // can distinguish "no head change" from "head was reshuffled".
+    const candidates: MmrCandidate[] = [
+      { id: "x", content: "one", score: 0.9, embedding: [1, 0, 0] },
+      { id: "y", content: "two", score: 0.8, embedding: [0, 1, 0] },
+      { id: "z", content: "three", score: 0.7, embedding: [0, 0, 1] },
+    ];
+    const reordered = applyMmrToCandidates({
+      candidates,
+      lambda: 0.7,
+      budget: candidates.length,
+    });
+    // Sanity: MMR should not reshuffle a perfectly diverse pool.
+    assert.deepEqual(
+      reordered.map((c) => c.id),
+      ["x", "y", "z"],
+    );
+    const report = summarizeMmrDiversity(candidates, reordered);
+    assert.equal(report.headReorderCount, 0);
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: single-element input must report considered=1 kept=1, not the
+// 0/0 sentinel meant for the empty case.
+// (Cursor Bugbot Low-severity review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "reorderRecallResultsWithMmr reports considered=1 kept=1 for single-element input",
+  () => {
+    const results: MmrRecallResult[] = [
+      { docid: "only", path: "p/only", snippet: "sole result", score: 0.9 },
+    ];
+    const { reordered, diversity } = reorderRecallResultsWithMmr(results);
+    assert.equal(reordered.length, 1);
+    assert.equal(diversity.considered, 1);
+    assert.equal(diversity.kept, 1);
+    assert.equal(diversity.headReorderCount, 0);
+  },
+);
+
+// ---------------------------------------------------------------------------
+// Regression: diversity metric must detect head swaps even when two results
+// share the same base key (path or docid). Previously the reordered-side
+// candidate ids were rebuilt with the *post-MMR* position index, so the
+// position suffix cancelled out and two same-path results swapping into the
+// same head position looked like a no-op.
+// (Cursor Bugbot Low-severity review comment on PR #391)
+// ---------------------------------------------------------------------------
+
+test(
+  "reorderRecallResultsWithMmr detects head swaps for results sharing a base key",
+  () => {
+    // Both results share the same `path` and empty `docid`, so the base
+    // portion of makeRecallKey is identical. Only the input-index suffix
+    // distinguishes them. If the diversity report rebuilt the reordered
+    // candidate ids with the post-MMR position index, a swap at position 0
+    // would be invisible because both reorderings produce the same
+    // `"shared-path::0"` id. After the fix the original-index-suffixed ids
+    // are preserved, so the head swap is counted.
+    const results: MmrRecallResult[] = [
+      {
+        docid: "",
+        path: "shared/path.md",
+        // Strongly duplicate-cluster-forming snippet at top.
+        snippet: "dup alpha content alpha",
+        score: 0.99,
+      },
+      {
+        docid: "",
+        path: "shared/path.md",
+        snippet: "dup alpha content alpha near",
+        score: 0.98,
+      },
+      {
+        docid: "",
+        path: "shared/path.md",
+        // Orthogonal diverse candidate.
+        snippet: "rocket propellant chemistry",
+        score: 0.97,
+      },
+    ];
+    const { reordered, diversity } = reorderRecallResultsWithMmr(results, {
+      lambda: 0.3,
+      diversitySampleSize: 2,
+    });
+    // MMR should have surfaced the diverse "rocket" candidate into the head
+    // two — verify by checking the first two snippets include the diverse
+    // one (order of the cluster pair is implementation-dependent).
+    const headSnippets = reordered.slice(0, 2).map((r) => r.snippet);
+    assert.ok(
+      headSnippets.some((s) => s && s.includes("rocket")),
+      `diverse candidate should be in head-of-list: ${headSnippets.join(" | ")}`,
+    );
+    // The real regression: headReorderCount must be > 0 even though all
+    // three inputs share the same base key.
+    assert.ok(
+      diversity.headReorderCount > 0,
+      `headReorderCount must detect same-base-key head swaps, got ${diversity.headReorderCount}`,
+    );
+  },
+);
+
+test(
+  "reorderRecallResultsWithMmr diversity report defaults to head-of-list sample",
+  () => {
+    // Integration check for the orchestration helper: even with
+    // `budget = results.length`, the default diversity report should surface
+    // the fact that MMR promoted diverse candidates to the head.
+    const results: MmrRecallResult[] = [
+      { docid: "a1", path: "p/a1", snippet: "the quick brown fox jumps over the lazy dog", score: 0.99 },
+      { docid: "a2", path: "p/a2", snippet: "the quick brown fox jumps over a lazy dog", score: 0.98 },
+      { docid: "a3", path: "p/a3", snippet: "a quick brown fox jumps over the lazy dog", score: 0.97 },
+      { docid: "a4", path: "p/a4", snippet: "quick brown fox jumping over lazy dogs", score: 0.96 },
+      { docid: "a5", path: "p/a5", snippet: "brown foxes jumping over lazy dogs", score: 0.95 },
+      { docid: "d1", path: "p/d1", snippet: "rocket thrust vectoring", score: 0.6 },
+      { docid: "d2", path: "p/d2", snippet: "violin bow rosin reservoir", score: 0.5 },
+      { docid: "d3", path: "p/d3", snippet: "coffee brewing at high altitude", score: 0.4 },
+    ];
+
+    const { reordered, diversity } = reorderRecallResultsWithMmr(results, {
+      lambda: 0.3,
+      diversitySampleSize: 4,
+    });
+    assert.equal(reordered.length, results.length);
+    assert.ok(
+      diversity.avgPairwiseSimAfter + 1e-9 < diversity.avgPairwiseSimBefore,
+      `head-of-list MMR should lower avg pairwise similarity: before=${diversity.avgPairwiseSimBefore} after=${diversity.avgPairwiseSimAfter}`,
+    );
+    // Sanity: MMR should have pulled at least one diverse candidate into the head.
+    const headDocids = reordered.slice(0, 4).map((r) => r.docid);
+    assert.ok(
+      headDocids.some((d) => d && d.startsWith("d")),
+      `expected a diverse candidate in the head after MMR: ${headDocids.join(",")}`,
+    );
+  },
+);