@sanqianx/project-knowledge 5.0.24 → 5.0.26

package/_site/_test/parallel-retrieval-followup-citation-test.js

@@ -0,0 +1,219 @@
+'use strict';
+// Offline regression for the RET-007 attribution contract.
+//
+// What is locked here: a promoted knowledge entry must carry the FULL commit
+// SHA and the project it belongs to, a later session must retrieve that entry
+// by a semantic query that contains NO identifier, and an unrelated hex query
+// must not reach the entry through the exact-commit channel.
+//
+// What is NOT claimed here: the analyzer is a DECLARED STUB (no model is
+// called) and the vector endpoint is a DECLARED STUB that returns no rows, so
+// this file says nothing about model behaviour or about vector recall. The
+// real-model chain and the real index live in
+// qa-autopilot/tools/product-probes/parallel-retrieval-followup-ret007.js,
+// which is deliberately outside the offline CI.
+//
+// Everything runs in isolation: own data directory, own knowledge root, own
+// temporary repository, own HTTP ports, no ambient module address, no real
+// HOME.
+const assert = require('node:assert/strict');
+const fs = require('node:fs');
+const path = require('node:path');
+const http = require('node:http');
+const crypto = require('node:crypto');
+
+const shell = require('./helpers/parallel-retrieval-followup-shell');
+// Isolate before any product module is loaded.
+shell.isolateProcess();
+
+const { createEngineServer } = require('../../_modules/knowledge-engine/src/server');
+const AtomicFile = require('../../_modules/knowledge-engine/src/runtime/atomic-file');
+
+const TOKEN = 'RET007OFFLINE';
+const results = [];
+
+function record(caseId, status, detail) {
+  results.push({ caseId, status, detail });
+  console.log(`CASE ${caseId} ${status} ${JSON.stringify(detail)}`);
+}
+
+async function stubVectorServer() {
+  const requests = [];
+  const server = http.createServer((req, res) => {
+    let body = '';
+    req.on('data', chunk => { body += chunk; });
+    req.on('end', () => {
+      requests.push({ url: req.url, method: req.method });
+      res.writeHead(200, { 'Content-Type': 'application/json' });
+      res.end(JSON.stringify({ ok: true, results: [], errors: [], projects: [] }));
+    });
+  });
+  await new Promise(resolve => server.listen(0, '127.0.0.1', resolve));
+  return {
+    requests,
+    url: `http://127.0.0.1:${server.address().port}`,
+    close: () => { server.closeAllConnections(); return new Promise(resolve => server.close(resolve)); },
+  };
+}
+
+function forbiddenBridge() {
+  const guard = name => async () => { throw new Error(`event bridge must not be contacted from the test lane (${name})`); };
+  return {
+    registerConsumer: guard('registerConsumer'),
+    listJournals: guard('listJournals'),
+    readEvents: guard('readEvents'),
+    ackConsumerCursor: guard('ackConsumerCursor'),
+    close() {},
+  };
+}
+
+// Declared stub analyzer: stages one Markdown file plus a manifest that
+// carries the four required evidence references, exactly like the real
+// analyzer's contract. It never contacts a model.
+function stubAnalyzer(events = []) {
+  return {
+    sourceReadMode: 'isolated-frozen-evidence',
+    async runClaim(input) {
+      events.push({ kind: 'start', projectId: input.projectId, sha: input.claim.commitSha });
+      const relative = `changes/${input.claim.commitSha.slice(0, 12)}.md`;
+      const content = `# 离线归属条目\n\n${TOKEN} 事实：重试上限调整为 7 次，提交 ${input.claim.commitSha}。\n`;
+      AtomicFile.writeFileAtomic(path.join(input.stagingPath, 'files', ...relative.split('/')), content);
+      AtomicFile.writeJsonAtomic(input.manifestPath, {
+        schema: 'knowledge-staging-manifest/v1',
+        projectId: input.projectId,
+        runId: input.claim.runId,
+        commitSha: input.claim.commitSha,
+        operations: [{
+          path: relative,
+          operation: 'create',
+          sha256: `sha256:${crypto.createHash('sha256').update(content).digest('hex')}`,
+          reason: 'Frozen Git evidence',
+          evidenceReferences: [
+            `commit:${input.claim.commitSha}`,
+            `patch:${input.claim.patchHash}`,
+            `conversation:${input.claim.conversationSnapshotHash}`,
+            `retrieval:${input.claim.retrievalManifestHash}`,
+          ],
+        }],
+      });
+      events.push({ kind: 'finish', projectId: input.projectId, sha: input.claim.commitSha });
+      return { ok: true };
+    },
+  };
+}
+
+async function startEngine({ dataDir, vectorUrl, analyzer }) {
+  fs.mkdirSync(dataDir, { recursive: true });
+  const runtime = createEngineServer({ dataDir, vectorUrl, bridge: forbiddenBridge(), analyzer });
+  await runtime.ready;
+  await new Promise(resolve => runtime.server.listen(0, '127.0.0.1', resolve));
+  const base = `http://127.0.0.1:${runtime.server.address().port}`;
+  return {
+    ...runtime,
+    base,
+    async call(route, body) {
+      const response = await fetch(`${base}${route}`, {
+        method: body === undefined ? 'GET' : 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: body === undefined ? undefined : JSON.stringify(body),
+      });
+      return { status: response.status, body: await response.json() };
+    },
+  };
+}
+
+(async () => {
+  const root = shell.makeTempDir('parallel-retrieval-followup-citation');
+  const vector = await stubVectorServer();
+  const analyzerEvents = [];
+  let engine = null;
+  try {
+    const repo = shell.makeRepo(root, 'repo-citation', [
+      { message: '初始提交', files: { 'README.md': '# citation\n' } },
+      { message: '事实提交', files: { 'note.txt': `${TOKEN} 重试上限调整为 7 次。\n` } },
+    ]);
+    const knowledgePath = shell.makeKnowledgeDir(root, 'knowledge-citation');
+    engine = await startEngine({
+      dataDir: path.join(root, 'engine-citation'),
+      vectorUrl: vector.url,
+      analyzer: stubAnalyzer(analyzerEvents),
+    });
+    const registration = await engine.call('/api/projects', {
+      projectId: 'project-citation',
+      name: 'citation',
+      repoPath: repo.repoPath,
+      knowledgePath,
+      brainId: 'brain-default',
+      domainId: 'domain-tools',
+    });
+    assert.equal(registration.status, 200, `the project registration failed: ${JSON.stringify(registration.body)}`);
+
+    const run = engine.coordinator.enqueue({ projectId: 'project-citation', commitSha: repo.head });
+    const settled = await shell.eventually(() => ['succeeded', 'failed'].includes(engine.store.getRun(run.runId).state), 20000);
+    assert.equal(settled, true, 'the run must reach a terminal state');
+    const finished = engine.store.getRun(run.runId);
+    assert.equal(finished.state, 'succeeded', `the promotion failed: ${JSON.stringify(finished.error)}`);
+
+    const commitSha = repo.head;
+    const entryRelative = `changes/${commitSha.slice(0, 12)}.md`;
+    const promoted = fs.readFileSync(path.join(knowledgePath, ...entryRelative.split('/')), 'utf8');
+    assert.ok(promoted.includes(TOKEN), 'the promoted body must carry the planted fact');
+    assert.ok(promoted.includes(commitSha), 'the promoted body must carry the full commit SHA');
+
+    // A later session asks with a semantic query that contains no identifier.
+    const semanticQuery = `${TOKEN} 重试上限调整为几次`;
+    assert.ok(!/[0-9a-f]{7}/i.test(semanticQuery), 'the semantic query must not contain a hex identifier');
+
+    const search = await engine.coordinator.retrieval.search({ query: semanticQuery, projectId: 'project-citation', client: 'workbench', sessionId: 'offline-session' });
+    const rows = search.results || [];
+    const hit = rows.find(row => row.commitSha === commitSha);
+    assert.ok(hit, `the promoted entry must be retrieved with its full SHA; rows: ${JSON.stringify(rows.map(row => ({ entry: row.entry_id, sha: row.commitSha, channels: row.match_channels })))}`);
+    assert.equal(hit.projectId, 'project-citation', 'the retrieved entry must keep its project');
+    assert.ok(String(hit.chunk_text || hit.snippet || '').includes(TOKEN), 'the retrieved entry must carry the planted body');
+    assert.ok(!(hit.match_channels || []).includes('exact-commit'), 'a semantic query must not reach the entry through the exact-commit channel');
+
+    const asked = await engine.coordinator.retrieval.ask({ query: semanticQuery, projectId: 'project-citation' });
+    const citation = (asked.citations || []).find(entry => entry.commitSha === commitSha);
+    assert.ok(citation, 'the answer must cite the promoted entry with its full SHA');
+    assert.equal(citation.projectId, 'project-citation', 'the citation must keep its project');
+    assert.ok(String(citation.text || '').includes(TOKEN), 'the citation must carry the planted body');
+
+    // Control: an unrelated hex query must not reach the entry.
+    const controlQuery = 'deadbeefcafe';
+    const control = await engine.coordinator.retrieval.search({ query: controlQuery, projectId: 'project-citation', client: 'workbench', sessionId: 'offline-session' });
+    const controlHit = (control.results || []).find(row => row.commitSha === commitSha);
+    assert.equal(controlHit, undefined, 'an unrelated hex query must not return the promoted entry');
+    assert.equal((control.results || []).some(row => (row.match_channels || []).includes('exact-commit')), false, 'an unrelated hex query must not use the exact-commit channel');
+
+    record('RET-007-CITE', 'PASS', {
+      commitSha,
+      entryRelative,
+      retrievedRows: rows.length,
+      hitEntryId: hit.entry_id,
+      hitChannels: hit.match_channels,
+      citationEntryId: citation.entryId,
+      controlRows: (control.results || []).length,
+      analyzer: 'declared stub analyzer (no model call)',
+      vector: 'declared stub returning no rows',
+    });
+
+    const summary = {
+      lane: 'parallel-retrieval-followup',
+      file: 'parallel-retrieval-followup-citation-test.js',
+      cases: results,
+      model: 'declared stub analyzer (no model call)',
+      vector: 'declared stub returning no rows',
+    };
+    const target = process.env.PARALLEL_RETRIEVAL_EVIDENCE;
+    if (target) {
+      fs.mkdirSync(path.dirname(target), { recursive: true });
+      fs.writeFileSync(target, JSON.stringify(summary, null, 2), 'utf8');
+    }
+    console.log('parallel-retrieval-followup-citation-test PASS');
+  } finally {
+    if (engine) await engine.close();
+    await vector.close();
+    shell.removeTree(root);
+    shell.cleanupIsolatedProcess();
+  }
+})().catch(error => { console.error(error); process.exitCode = 1; });

package/_site/_test/parallel-retrieval-followup-corpus-test.js

@@ -0,0 +1,287 @@
+'use strict';
+// Offline regression for the RET-008 harness contract.
+//
+// This file locks the pieces of the RET-008 comparison that do NOT need a
+// model and do NOT need the vector-hub source checkout:
+//
+//   1) the fixed corpus and the human-labelled ground truth stay consistent
+//      (unique ids, every relevant document exists, no query carries a hex
+//      identifier, no query is a verbatim copy of a document sentence);
+//   2) the ranking and latency helpers compute what the report claims;
+//   3) the engine and the PINNED vector-hub package agree on the body hash,
+//      which is the precondition the engine needs before it may merge a
+//      vector-hub chunk row into a retrieved fact;
+//   4) the PINNED package still has the identity the comparison was made
+//      against, and its query cache still BEHAVES the way the comparison
+//      assumed: an identical query is served from the cache, a failed query
+//      is evicted, and a late failure of an evicted owner cannot destroy the
+//      replacement cached for the same text. The behaviour is observed
+//      through the public API with deterministic counts; the source text of
+//      the pinned build is never matched.
+//
+// What is NOT claimed here: the embedder is a DECLARED STUB (a deterministic
+// character-gram hashing vector, NOT a neural model), so this file says
+// nothing about retrieval quality or recall. The real-model measurement
+// (Xenova/bge-small-zh-v1.5) and the baseline/candidate comparison live in
+// qa-autopilot/tools/product-probes/parallel-retrieval-followup-ret008.js,
+// which is deliberately outside the offline CI.
+//
+// Everything runs in isolation: own temp root, own hub root, own ports, no
+// ambient module address, no real HOME, no network.
+const assert = require('node:assert/strict');
+const fs = require('node:fs');
+const path = require('node:path');
+const crypto = require('node:crypto');
+
+const shell = require('./helpers/parallel-retrieval-followup-shell');
+// Isolate before any product module is loaded.
+shell.isolateProcess();
+
+const corpusDefinitions = require('./helpers/parallel-retrieval-followup-corpus');
+const ranking = require('../../_modules/knowledge-engine/src/retrieval-ranking');
+
+const HUB_DIST = path.join(shell.REPO_ROOT, 'node_modules', '@sanqianx', 'vector-hub', 'dist', 'cjs');
+const DIMENSIONS = 512;
+const results = [];
+
+function record(caseId, status, detail) {
+  results.push({ caseId, status, detail });
+  console.log(`CASE ${caseId} ${status} ${JSON.stringify(detail)}`);
+}
+
+// DECLARED STUB embedder: a deterministic character-gram hashing vector.
+// It is not a model and makes no claim about semantic quality; it only gives
+// the hub real numbers to store and compare. The counters separate the
+// document embeddings (a batch) from the query embeddings (a single string),
+// because the query cache is only observable on the single-string path.
+function stubEmbedder() {
+  const state = { batchCalls: 0, singleCalls: 0, failNext: false, deferFor: null, pending: [] };
+  const vector = text => {
+    const out = new Array(DIMENSIONS).fill(0);
+    const source = String(text || '').toLowerCase();
+    for (let i = 0; i < source.length; i += 1) {
+      const gram = source.slice(i, i + 2);
+      const slot = crypto.createHash('sha256').update(gram).digest();
+      const index = slot.readUInt16BE(0) % DIMENSIONS;
+      out[index] += 1 + (slot[2] / 255);
+    }
+    const norm = Math.sqrt(out.reduce((sum, value) => sum + value * value, 0)) || 1;
+    return out.map(value => value / norm);
+  };
+  state.vector = vector;
+  const embedder = {
+    kind: 'declared-stub-hashing-embedder',
+    model: 'stub-hashing-embedder',
+    modelId: 'stub-hashing-embedder',
+    maxTokens: 500,
+    async createEmbeddings(inputs) {
+      if (Array.isArray(inputs)) {
+        state.batchCalls += 1;
+        return { status: 'success', output: inputs.map(vector) };
+      }
+      state.singleCalls += 1;
+      if (state.failNext) {
+        state.failNext = false;
+        return { status: 'failure', output: [] };
+      }
+      if (state.deferFor && state.deferFor(String(inputs))) {
+        return new Promise((resolve, reject) => { state.pending.push({ text: String(inputs), resolve, reject }); });
+      }
+      return { status: 'success', output: [vector(inputs)] };
+    },
+  };
+  return { embedder, state };
+}
+
+async function waitUntil(predicate, timeoutMs = 5000, intervalMs = 20) {
+  const deadline = Date.now() + timeoutMs;
+  while (Date.now() < deadline) {
+    if (predicate()) return true;
+    await new Promise(resolve => setTimeout(resolve, intervalMs));
+  }
+  return false;
+}
+
+(async () => {
+  const root = shell.makeTempDir('parallel-retrieval-followup-corpus');
+  try {
+    // ------------------------------------------------------- definitions
+    const ids = corpusDefinitions.CORPUS.map(doc => doc.id);
+    assert.equal(new Set(ids).size, ids.length, 'every corpus id must be unique');
+    const files = corpusDefinitions.CORPUS.map(doc => doc.file);
+    assert.equal(new Set(files).size, files.length, 'every corpus file name must be unique');
+    for (const doc of corpusDefinitions.CORPUS) {
+      assert.ok(doc.content.trim().length > 0, `the corpus document ${doc.id} must not be empty`);
+    }
+    assert.ok(corpusDefinitions.QUERIES.length >= 5, 'the query set must be usable');
+    for (const definition of corpusDefinitions.QUERIES) {
+      assert.ok(definition.relevant.length > 0, `the query ${definition.id} must name at least one relevant document`);
+      for (const relevant of definition.relevant) {
+        assert.ok(ids.includes(relevant), `the query ${definition.id} names the unknown document ${relevant}`);
+      }
+      assert.ok(!/[0-9a-f]{7}/i.test(definition.query), `the query ${definition.id} must not carry a hex identifier`);
+      for (const doc of corpusDefinitions.CORPUS) {
+        assert.ok(!doc.content.includes(definition.query), `the query ${definition.id} is copied verbatim from ${doc.id}`);
+      }
+    }
+    record('RET-008-DEFINITIONS', 'PASS', {
+      documents: corpusDefinitions.CORPUS.length,
+      queries: corpusDefinitions.QUERIES.length,
+      multiRelevantQueries: corpusDefinitions.QUERIES.filter(definition => definition.relevant.length > 1).map(definition => definition.id),
+    });
+
+    // ------------------------------------------------------------ helpers
+    const ranked = corpusDefinitions.evaluateRanking(['b', 'a', 'c'], ['a', 'c']);
+    assert.equal(ranked.firstRelevantRank, 2, 'the first relevant rank must be reported');
+    assert.equal(ranked.reciprocalRank, 0.5, 'the reciprocal rank must be 1/2');
+    assert.equal(ranked.recallAt1, 0, 'nothing relevant is in the top 1');
+    assert.equal(ranked.recallAt5, 1, 'both relevant documents are within the top 5');
+    const empty = corpusDefinitions.evaluateRanking(['x'], ['a']);
+    assert.equal(empty.hit, false, 'a query without any relevant document must not count as a hit');
+    assert.equal(empty.reciprocalRank, 0, 'a query without any relevant document has no reciprocal rank');
+    const latency = corpusDefinitions.summariseLatency([10, 20, 30, 40, 100]);
+    assert.equal(latency.count, 5, 'every finite sample must be counted');
+    assert.equal(latency.min, 10, 'the minimum must be reported');
+    assert.equal(latency.max, 100, 'the maximum must be reported');
+    assert.equal(latency.p50, 30, 'the median must be reported');
+    assert.equal(corpusDefinitions.summariseLatency([]).p50, null, 'an empty sample has no percentile');
+    record('RET-008-METRICS', 'PASS', { ranked, latency });
+
+    // -------------------------------------------- engine/hub body hash
+    const corpusDir = path.join(root, 'corpus');
+    const corpusIndex = corpusDefinitions.materialiseCorpus(corpusDir);
+    assert.equal(fs.readdirSync(corpusDir).length, corpusDefinitions.CORPUS.length, 'every corpus document must be written');
+    assert.equal(corpusIndex.idForFile('quota-sampling.md'), 'quota-sampling', 'a result file must map back to its corpus id');
+    assert.equal(corpusIndex.idForFile('missing.md'), null, 'an unknown file must not map to a corpus id');
+
+    const { VectorHub } = require(path.join(HUB_DIST, 'core', 'hub.js'));
+    const hub = new VectorHub({ rootPath: path.join(root, 'hub') });
+    const corpusEmbedder = stubEmbedder();
+    hub.setEmbeddings(corpusEmbedder.embedder);
+    const tracked = await hub.syncFolder('ret008-offline', corpusDir);
+    assert.equal(tracked, corpusDefinitions.CORPUS.length, 'every corpus document must reach the index');
+
+    const response = await hub.search(corpusDefinitions.QUERIES[0].query, { maxResults: 10, resultLevel: 'chunk' });
+    assert.ok(response.results.length > 0, 'the chunk search must return rows');
+    const expectedHashes = new Map();
+    for (const doc of corpusDefinitions.CORPUS) {
+      const content = fs.readFileSync(path.join(corpusDir, doc.file), 'utf8');
+      expectedHashes.set(doc.file, ranking.bodyHash(content));
+    }
+    let checked = 0;
+    for (const row of response.results) {
+      const expected = expectedHashes.get(row.file);
+      assert.ok(expected, `the search returned an unknown document ${row.file}`);
+      assert.equal(row.bodyHash, expected, `the engine and the pinned hub must agree on the body hash of ${row.file}`);
+      assert.equal(row.scoreType, 'cosine', 'a chunk row must carry the cosine score type');
+      assert.ok(Number.isInteger(row.startPos) && Number.isInteger(row.endPos), 'a chunk row must carry integer positions');
+      checked += 1;
+    }
+    record('RET-008-BODY-HASH', 'PASS', { tracked, rowsChecked: checked, distinctFiles: new Set(response.results.map(row => row.file)).size });
+
+    // The pinned dependency must still be the one the RET-008 comparison was
+    // made against. A dependency change has to re-run the explicit probe.
+    // Only the PACKAGE IDENTITY is asserted here: the source text of the
+    // pinned build is never matched, because a string match against the
+    // implementation is not a behaviour.
+    const pinnedPackage = path.join(shell.REPO_ROOT, 'node_modules', '@sanqianx', 'vector-hub');
+    const pinnedManifest = JSON.parse(fs.readFileSync(path.join(pinnedPackage, 'package.json'), 'utf8'));
+    assert.equal(pinnedManifest.name, '@sanqianx/vector-hub', 'the pinned package identity changed');
+    assert.equal(
+      pinnedManifest.version,
+      '1.0.8-rc.8',
+      `the pinned @sanqianx/vector-hub is ${pinnedManifest.version} instead of 1.0.8-rc.8; re-run qa-autopilot/tools/product-probes/parallel-retrieval-followup-ret008.js`,
+    );
+    record('RET-008-PINNED-PACKAGE', 'PASS', { pinnedName: pinnedManifest.name, pinnedVersion: pinnedManifest.version });
+
+    // ---- real query-cache behaviour of the pinned package -----------------
+    // The RET-008 comparison assumed that the pinned hub caches the query
+    // embedding and that a late failure of an evicted owner cannot destroy
+    // the replacement cached for the same text. Both are checked here through
+    // the public API with DETERMINISTIC COUNTS. The embedder is the DECLARED
+    // STUB, so this case says nothing about neural semantic quality; it only
+    // says that the cache still behaves the way the comparison assumed.
+    const cacheEmbedder = stubEmbedder();
+    const cacheHub = new VectorHub({ rootPath: path.join(root, 'hub-cache') });
+    cacheHub.setEmbeddings(cacheEmbedder.embedder);
+    const cacheTracked = await cacheHub.syncFolder('ret008-cache', corpusDir);
+    assert.equal(cacheTracked, corpusDefinitions.CORPUS.length, 'every corpus document must reach the cache index');
+    const search = text => cacheHub.search(text, { maxResults: 10, resultLevel: 'chunk' });
+
+    const queryA = corpusDefinitions.QUERIES[0].query;
+    const queryB = corpusDefinitions.QUERIES[1].query;
+    const beforeFirst = cacheEmbedder.state.singleCalls;
+    await search(queryA);
+    const afterFirst = cacheEmbedder.state.singleCalls;
+    assert.ok(afterFirst > beforeFirst, 'the first query must be embedded');
+    await search(queryA);
+    const afterRepeated = cacheEmbedder.state.singleCalls;
+    assert.equal(afterRepeated, afterFirst, 'the identical query must be served from the query cache');
+    const beforeSecond = cacheEmbedder.state.singleCalls;
+    await search(queryB);
+    const afterDifferent = cacheEmbedder.state.singleCalls;
+    assert.ok(afterDifferent > beforeSecond, 'a different query must be embedded again');
+    record('RET-008-QUERY-CACHE-HIT', 'PASS', {
+      firstQueryEmbeddings: afterFirst - beforeFirst,
+      repeatedQueryEmbeddings: afterRepeated - afterFirst,
+      differentQueryEmbeddings: afterDifferent - beforeSecond,
+      documentEmbeddings: cacheEmbedder.state.batchCalls,
+    });
+
+    // A failed query must not survive in the cache.
+    const queryFail = `${queryB} 失败用例`;
+    cacheEmbedder.state.failNext = true;
+    await search(queryFail).catch(() => {});
+    const afterFailed = cacheEmbedder.state.singleCalls;
+    await search(queryFail).catch(() => {});
+    assert.ok(cacheEmbedder.state.singleCalls > afterFailed, 'a failed query must not stay cached');
+    record('RET-008-QUERY-CACHE-FAILURE-EVICTION', 'PASS', {
+      embeddingsAfterFailedQuery: afterFailed,
+      embeddingsAfterRetry: cacheEmbedder.state.singleCalls,
+    });
+
+    // The ownership guard: a late failure of an evicted owner must not
+    // destroy the replacement that was cached for the same text.
+    const ownerQuery = '归属校验用的查询文本';
+    cacheEmbedder.state.deferFor = text => text === ownerQuery;
+    const firstOwner = search(ownerQuery).catch(() => null);
+    assert.ok(await waitUntil(() => cacheEmbedder.state.pending.length === 1), 'the first owner was never created');
+    // The hub keeps 100 cached queries; 100 distinct queries evict the
+    // oldest one, which is the entry of the first owner.
+    for (let index = 0; index < 100; index += 1) await search(`驱逐查询 ${index}`);
+    const replacement = search(ownerQuery).catch(() => null);
+    assert.ok(await waitUntil(() => cacheEmbedder.state.pending.length === 2), 'the replacement owner was never created');
+    // The evicted owner now fails late. It must not delete the replacement.
+    cacheEmbedder.state.pending[0].reject(new Error('late failure of the evicted owner'));
+    await firstOwner;
+    const beforeThird = cacheEmbedder.state.singleCalls;
+    const thirdSearch = search(ownerQuery);
+    await new Promise(resolve => setTimeout(resolve, 300));
+    assert.equal(cacheEmbedder.state.pending.length, 2, 'the replacement must still be cached: no new query embedding was created');
+    assert.equal(cacheEmbedder.state.singleCalls, beforeThird, 'the replacement must still be cached: the embedder was called again');
+    cacheEmbedder.state.pending[1].resolve({ status: 'success', output: [cacheEmbedder.state.vector(ownerQuery)] });
+    const thirdResult = await thirdSearch;
+    cacheEmbedder.state.deferFor = null;
+    record('RET-008-QUERY-CACHE-OWNERSHIP', 'PASS', {
+      evictingQueries: 100,
+      embedderCallsAfterLateFailure: cacheEmbedder.state.singleCalls - beforeThird,
+      replacementRows: Array.isArray(thirdResult && thirdResult.results) ? thirdResult.results.length : null,
+    });
+
+    const summary = {
+      lane: 'parallel-retrieval-followup',
+      file: 'parallel-retrieval-followup-corpus-test.js',
+      cases: results,
+      model: 'declared stub embedder (no model call)',
+    };
+    const target = process.env.PARALLEL_RETRIEVAL_EVIDENCE;
+    if (target) {
+      fs.mkdirSync(path.dirname(target), { recursive: true });
+      fs.writeFileSync(target, JSON.stringify(summary, null, 2), 'utf8');
+    }
+    console.log('parallel-retrieval-followup-corpus-test PASS');
+  } finally {
+    shell.removeTree(root);
+    shell.cleanupIsolatedProcess();
+  }
+})().catch(error => { console.error(error); process.exitCode = 1; });