@gamaze/hicortex 0.23.1 → 0.24.0

package/dist/eval/ranking-fixtures.js

@@ -1,485 +0,0 @@
-"use strict";
-/**
- * Planted ranking fixtures for the search-trust calibration gate (#425).
- *
- * Two fixture cases, both shaped like the field failure that opened the
- * issue (the owner's live console test, 2026-09-13): a user searches a
- * DISTINCTIVE PROPER NOUN they know exists; the one memory carrying that
- * token has the best similarity of all candidates AND the literal FTS hit,
- * yet ranks behind high-strength but less-relevant memories that win on
- * effective strength + graph connections.
- *
- *   (a) sirnas_exact_match — the AC1 fixture: a low-strength, unlinked,
- *       never-accessed memory carrying an invented proper noun against
- *       base-≥0.9 rivals with links among themselves, access history, and
- *       older creation dates. The rivals share the query's DOMAIN
- *       vocabulary (harbours, docks, boats) but never the proper noun.
- *       Declared expectation: the proper-noun row ranks #1.
- *   (b) no_match_control — the AC2 fixture: a "mamma"-class query with no
- *       relevant match and NO FTS hits anywhere in the corpus. Whatever
- *       nearest-neighbor junk vector search returns must be byte-stable
- *       under the ranking work (no both-channel candidates exist, so the
- *       boost can never fire — the case pins that the no-regression
- *       guarantee is structural, not lucky).
- *   (c) linked_fallback — the #449 (items 1+3) connections fixture: a
- *       16-link vs 20-link identical-content hub pair (the saturation tie)
- *       plus an unlinked higher-similarity row vs a 4-link lower-similarity
- *       row (the direction pin). Planted in its OWN throwaway DB — the 40
- *       link-neighbor rows would pollute the case-1/2 corpus.
- *
- * Honesty rules (the planted-pairs #393 A discipline):
- *  - The texts are synthetic and generic (this package publishes to npm —
- *    no real infrastructure, people, or fleet detail), but SHAPED like the
- *    field failure: the proper noun is invented (it must appear nowhere
- *    else — verified by an invariant test), the rivals carry the metadata
- *    the production pipeline itself writes (base strength, access history,
- *    link counts).
- *  - Similarities are MEASURED at run time with the real bge-small-en-v1.5
- *    embedder (ranking-eval.ts) — never asserted into existence. The
- *    declared expectation is about ORDER under the shipped ranking, and
- *    the report prints the measured similarity/score of every top row.
- *  - Strength metadata is written through the production helpers
- *    (insertMemory + updateMemory + addLink), not hand-rolled SQL, so the
- *    fixture exercises the same rows retrieval() will see.
- */
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.LINKED_FALLBACK_FIXTURE = exports.REAL_SIRNAS_MEMORY_ID = exports.NO_MATCH_FIXTURE = exports.SIRNAS_FIXTURE = exports.RANKING_SOURCE_AGENT = exports.RANKING_PROJECT = void 0;
-exports.checkFixtureInvariants = checkFixtureInvariants;
-/** All planted rows share project + source_agent (metadata never refuses). */
-exports.RANKING_PROJECT = "ranking-eval";
-exports.RANKING_SOURCE_AGENT = "ranking-eval";
-// ---------------------------------------------------------------------------
-// Case 1 — sirnas_exact_match (AC1)
-// ---------------------------------------------------------------------------
-//
-// The target mirrors the field case's SHAPE with invented names: a home
-// harbour memory carrying one distinctive proper noun, born weak (base 0.35,
-// the "routine-but-curated" band), never accessed, unlinked, a few months
-// old. The rivals are what the decay model produces for early high-strength
-// technical rows: base 0.90-0.95, mutual links, access history, created
-// BEFORE the target. They share the harbour/boating domain vocabulary — so
-// a bare proper-noun query lands in their neighborhood — but none carries
-// the proper noun.
-const SIRNAS_ROWS = [
-    // --- rivals: high-strength, connected, older, domain-adjacent ---
-    {
-        key: "rival_fees",
-        content: "Harbour fees and guest dock rules: pay at the machine by the fuel quay, electricity on the guest dock is metered, " +
-            "and visiting boats must clear their berth by ten in the morning during the summer season.",
-        createdAt: "2026-03-02T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.9,
-        accessCount: 4,
-        lastAccessedDaysAgo: 6,
-        linksTo: ["rival_channel", "rival_maintenance"],
-    },
-    {
-        key: "rival_channel",
-        content: "Sailing notes for the home bay: the channel is marked with cardinal buoys, the shallow spit near the outer dock " +
-            "dries at low water, and the ferry has right of way inside the harbour basin.",
-        createdAt: "2026-03-16T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.95,
-        accessCount: 6,
-        lastAccessedDaysAgo: 2,
-        linksTo: ["rival_fees", "rival_maintenance"],
-    },
-    {
-        key: "rival_maintenance",
-        content: "Boat maintenance log: scrub the hull and check the anodes before launch, service the winch every spring, and " +
-            "book the boatyard crane early — the lifting slot list fills up in May.",
-        createdAt: "2026-03-30T09:00:00.000Z",
-        memoryType: "experience",
-        baseStrength: 0.9,
-        accessCount: 3,
-        lastAccessedDaysAgo: 11,
-        linksTo: ["rival_fees"],
-    },
-    {
-        key: "rival_passage",
-        content: "Passage plan habits: file the float plan before leaving the dock, plot the tidal gate times, and always take the " +
-            "safe water mark side of the reef when the light is obscured.",
-        createdAt: "2026-04-13T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.9,
-        accessCount: 2,
-        lastAccessedDaysAgo: 20,
-        linksTo: ["rival_channel"],
-    },
-    // --- the target: the exact proper-noun match, born humble ---
-    {
-        key: "target_harbour",
-        content: "Home harbour: Vindarsvik (near Stensundet). The guest berth is east of the ferry pier; hold slack when the " +
-            "fishing boats unload.",
-        createdAt: "2026-05-25T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.35,
-        accessCount: 0,
-        lastAccessedDaysAgo: 118,
-        linksTo: [],
-    },
-    // --- neutral filler: keeps neighborhoods non-trivial, off-domain ---
-    {
-        key: "fill_recipe",
-        content: "Batch-cooked a tomato and butter sauce from the Rome cookbook: San Marzano tomatoes, one whole onion halved, and far more butter than seems reasonable.",
-        createdAt: "2026-04-01T09:00:00.000Z",
-        memoryType: "experience",
-        baseStrength: 0.5,
-        accessCount: 0,
-        lastAccessedDaysAgo: 160,
-        linksTo: [],
-    },
-    {
-        key: "fill_books",
-        content: "Finished the third novel in the frontier trilogy. The middle book dragged through the mining-town chapters, but the finale pays off when the surveyor reads her grandmother's letters.",
-        createdAt: "2026-04-20T09:00:00.000Z",
-        memoryType: "experience",
-        baseStrength: 0.5,
-        accessCount: 0,
-        lastAccessedDaysAgo: 140,
-        linksTo: [],
-    },
-    {
-        key: "fill_lang",
-        content: "Language study notes: the dative prepositions finally clicked after drilling them as a sung list; irregular verbs still need spaced repetition.",
-        createdAt: "2026-05-10T09:00:00.000Z",
-        memoryType: "experience",
-        baseStrength: 0.5,
-        accessCount: 0,
-        lastAccessedDaysAgo: 120,
-        linksTo: [],
-    },
-    {
-        key: "fill_garden",
-        content: "The raised beds need refreshing before spring: compost the spent tomato vines, rotate the legume row, and prune the apple espalier before the buds swell.",
-        createdAt: "2026-06-02T09:00:00.000Z",
-        memoryType: "experience",
-        baseStrength: 0.5,
-        accessCount: 0,
-        lastAccessedDaysAgo: 100,
-        linksTo: [],
-    },
-];
-exports.SIRNAS_FIXTURE = {
-    key: "sirnas_exact_match",
-    description: "AC1: a distinctive proper-noun query must rank the exact-match (both-channel) memory first, " +
-        "ahead of high-strength domain-adjacent rivals that win on strength + connections.",
-    rows: SIRNAS_ROWS,
-    query: "Vindarsvik",
-    distinctiveToken: "vindarsvik",
-    expectedTop1: "target_harbour",
-    // Undirected degrees (#449 invariant extension): the rivals' mutual-link
-    // cluster — fees/channel 4 each, maintenance 3, passage 1 (each linksTo
-    // entry is one link row; both directions of a mutual pair count).
-    expectedUndirectedDegree: {
-        rival_fees: 4,
-        rival_channel: 4,
-        rival_maintenance: 3,
-        rival_passage: 1,
-        target_harbour: 0,
-        fill_recipe: 0,
-        fill_books: 0,
-        fill_lang: 0,
-        fill_garden: 0,
-    },
-};
-// ---------------------------------------------------------------------------
-// Case 2 — no_match_control (AC2)
-// ---------------------------------------------------------------------------
-//
-// A "mamma"-class query: a common household word in the user's language that
-// appears in NO corpus row — no FTS token, no semantic neighbor. Vector
-// search still returns nearest-neighbor junk (the field's 0.53-similarity
-// technical rows); the case pins the FULL returned list so any drift in
-// how unrelated junk is ordered is caught, and asserts structurally that
-// the ranking work cannot touch it (no both-channel candidates exist).
-exports.NO_MATCH_FIXTURE = {
-    key: "no_match_control",
-    description: "AC2: a no-match query (no FTS hits anywhere) returns a byte-stable list — no both-channel " +
-        "candidates exist, so the boost is structurally inert.",
-    // Same corpus as case 1 (minus nothing): the query matches none of it.
-    rows: SIRNAS_ROWS,
-    query: "mamma",
-    distinctiveToken: "",
-    // No declared top-1 — the assertion is stability, not order (the photo
-    // records the list; the gate compares runs).
-    expectedTop1: "",
-};
-/** The battery runs against a real snapshot copy (ranking-eval.ts case 4). */
-exports.REAL_SIRNAS_MEMORY_ID = "12980c9d-0eba-4963-95eb-0548bc2aa93e";
-// ---------------------------------------------------------------------------
-// Case 3 — linked_fallback (#449 items 1+3, PR E)
-// ---------------------------------------------------------------------------
-//
-// The connections-term fixture for the log-saturation reshape. Two probes in
-// ONE corpus (planted in its own throwaway DB by ranking-eval.ts — the 40
-// link-neighbor rows stay out of the case-1/2 corpus):
-//
-//   saturation pair — two hubs with IDENTICAL content strings and identical
-//   metadata (same baseStrength, accessCount, lastAccessedDaysAgo,
-//   createdAt, memoryType); hub A carries exactly 16 undirected links, hub B
-//   exactly 20, through disjoint off-domain neighbor rows. Declared
-//   expectation POST-change (Option A, K = 16): the two hubs' scores are
-//   byte-EQUAL — the term min(1, log1p(k)/log1p(16)) saturates at 1 for both.
-//   Under the pre-#449 linear term (connectionCount / maxConnections = k/20)
-//   they differ by 0.2 × 0.15 ≈ 0.03 composite — that red is the recorded
-//   before-picture; the harness does NOT make it pass.
-//
-//   direction pair — an unlinked row carrying the query's distinctive proper
-//   noun (the highest measured similarity) vs a linked row (exactly 4 links)
-//   whose MEASURED similarity is engineered ~0.20 LOWER, with equal metadata
-//   otherwise (equal strength/dates). Declared expectation (Option A, owner
-//   decision 2026-09-17): the unlinked row ranks ABOVE the linked one — a
-//   0.20 gap is worth 0.10 composite at the 0.50 similarity weight, more
-//   than the k=4 connections credit (0.0852 post-change). The gap must
-//   exceed 0.171 for the pin to be sound at k = 4; the wording is tuned
-//   until the real-embedder measurement (recorded in the photo/report)
-//   lands at ~0.20.
-//
-// Neighbors are DISTINCT filler rows, disjoint between the two hubs and from
-// the direction pair's cluster, all off-domain from the query — they exist
-// to carry the hub/dir-linked degrees and nothing else.
-const HUB_CONTENT = "Community noticeboard archive: the autumn fair needs tombola volunteers, choir practice moves to " +
-    "Thursday evenings, and the book-swap shelf now carries the walking-map folder.";
-/** Off-domain one-line filler rows (deterministic keys, uniform metadata). */
-function fillerRows(prefix, contents) {
-    return contents.map((content, i) => ({
-        key: `${prefix}_n${i}`,
-        content,
-        createdAt: "2026-04-15T09:00:00.000Z",
-        memoryType: "experience",
-        baseStrength: 0.5,
-        accessCount: 0,
-        lastAccessedDaysAgo: 150,
-        linksTo: [],
-    }));
-}
-const HUB_A_NEIGHBOR_CONTENTS = [
-    "Slow-cooker beans: soak overnight, then four hours on low with a bay leaf and a piece of smoked pork.",
-    "The second detective novel in the series drags until the lighthouse keeper's ledger turns up.",
-    "German grammar drill: subordinate clauses shove the verb to the end, so read the whole sentence before translating.",
-    "Raised-bed rotation: legumes follow brassicas, and lime the bed the cabbage family just emptied.",
-    "October weather pattern: the first night frosts usually arrive right after the harvest moon.",
-    "Bike maintenance: true the wheel by tightening spokes a quarter turn at a time, listening for the ping.",
-    "Chess opening note: against the locked center, flank with pawns only after the pieces are developed.",
-    "The documentary about the alpine tunnels spent too long on the geologists' cafeteria.",
-    "Coffee grinding: a burr grind for the pour-over; the blade grinder is fine enough only for the French press.",
-    "Hiking the ridge trail: start before the thermals build, and carry an extra liter once the pass is in sight.",
-    "Piano practice: separate hands at half tempo until the fingering is automatic, then join them slowly.",
-    "Sewing patch pockets: interface the fabric, and topstitch the same distance from every edge.",
-    "Bird notes: the chiffchaff repeats its own name all day; the willow warbler slides down the scale.",
-    "Sourdough rhythm: feed the starter after work, shape in the morning, bake when the poke test springs back halfway.",
-    "Replacing the washing-machine hose: shut the valve, expect a cup of water, and check the jubilee clip.",
-    "Audiobook note: the narrator's accent makes every ship's name sound like a Hebridean island.",
-];
-const HUB_B_NEIGHBOR_CONTENTS = [
-    "Winter tyres: swap before the first ice, and store the summer set flat in the dark basement.",
-    "The boardgame group settled on tiles-and-traders for the long evenings; the card game stays in the drawer.",
-    "Compost thermometer: the pile is cooking properly when it reads hotter than a warm bath for a week.",
-    "Frost-proof terracotta: empty the pots, stack them dry, and the succulents overwinter on the south sill.",
-    "The photography club theme this month is doors; shoot in morning shade and bracket the exposures.",
-    "Robot vacuum note: keep the charging dock's floor clear, or it nudges the rug and gives up mid-run.",
-    "Knife sharpening: the ceramic rod for touch-ups, the whetstone when the tomato skin resists.",
-    "The lecture on Roman concrete held the room until the slide about seawater pozzolana.",
-    "Necklace repair: jump rings open sideways, never apart, or the circle never closes again.",
-    "Ferry timetable change: the winter crossing departs an hour earlier and stops running on Sunday evenings.",
-    "Ski waxing: glide wax for the thaw, grip wax for the freeze, and never both in the same zone.",
-    "The archive boxes are labelled by decade now; the negatives live in the fire cupboard.",
-    "Beekeeping note: heft the hive — if one side lifts easily, the winter stores are short.",
-    "Jigsaw strategy: edges first, then the sky, and give the lamp a clean bulb for the blue hours.",
-    "The museum's clock gallery is closed for restoration until the pendulum clock returns.",
-    "Preserving basil: pesto freezes flat in bags; the leaves kept in water go black within a week.",
-    "Draft excluder: the brush strip on the letterbox cut the hallway draught more than the door seal did.",
-    "The organ recital opened with a toccata nobody knew and closed with everyone humming it.",
-    "Train-set wiring: one feeder per rail section, or the far loop slows to a crawl.",
-    "The pottery class moved on to wheel work; my first cylinder collapsed into a bowl with charm.",
-];
-const DIR_NEIGHBOR_CONTENTS = [
-    "Vaccine appointment bookkeeping: the clinic sends both a text and a letter, and only the letter carries the time.",
-    "Reading glasses: the cheap pair from the station shop outlasted the optician's by two years.",
-    "The new debit card needs the reader for anything over the limit; the shop keeps the old machine as backup.",
-    "Laundering the down jacket: tennis balls in the dryer, low heat, and patience.",
-];
-const hubNeighbors = (prefix, contents) => contents.map((_, i) => `${prefix}_n${i}`);
-const LINKED_FALLBACK_ROWS = [
-    // --- saturation pair: identical content + metadata, degrees 16 vs 20 ---
-    {
-        key: "hub_sat_16",
-        content: HUB_CONTENT,
-        createdAt: "2026-05-10T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.6,
-        accessCount: 0,
-        lastAccessedDaysAgo: 120,
-        linksTo: hubNeighbors("huba", HUB_A_NEIGHBOR_CONTENTS),
-    },
-    {
-        key: "hub_sat_20",
-        content: HUB_CONTENT,
-        createdAt: "2026-05-10T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.6,
-        accessCount: 0,
-        lastAccessedDaysAgo: 120,
-        linksTo: hubNeighbors("hubb", HUB_B_NEIGHBOR_CONTENTS),
-    },
-    // --- direction pair: equal metadata; similarity gap engineered ~0.20 ---
-    // MEASURED (bge-small-en-v1.5, 2026-09-17): unlinked 0.8233 vs linked
-    // 0.6022 — gap 0.2212. The linked row deliberately carries the plural
-    // "guests" (an FTS-DISTINCT token from the query's "guest" under the
-    // unicode61 tokenizer) so it stays single-channel: the pin isolates
-    // similarity vs connections credit, not the both-channel boost.
-    {
-        key: "dir_unlinked",
-        content: "Home anchorage: Anchorholme, east of the skerries. The guest mooring lies by the old timber jetty — " +
-            "arrive at slack water and mind the ferry wake inside the bay.",
-        createdAt: "2026-07-01T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.6,
-        accessCount: 0,
-        lastAccessedDaysAgo: 90,
-        linksTo: [],
-    },
-    {
-        key: "dir_linked",
-        content: "Anchoring for the night behind the skerries: lay out chain at least four times the depth in the crowded " +
-            "bay, take one of the buoys kept for guests if the harbour has room, and check the swing radius around " +
-            "the mooring field before dark.",
-        createdAt: "2026-07-01T09:00:00.000Z",
-        memoryType: "knowledge",
-        baseStrength: 0.6,
-        accessCount: 0,
-        lastAccessedDaysAgo: 90,
-        linksTo: ["dirn_n0", "dirn_n1", "dirn_n2", "dirn_n3"],
-    },
-    // --- link carriers: 40 disjoint off-domain filler rows (degrees all 1) ---
-    ...fillerRows("huba", HUB_A_NEIGHBOR_CONTENTS),
-    ...fillerRows("hubb", HUB_B_NEIGHBOR_CONTENTS),
-    ...fillerRows("dirn", DIR_NEIGHBOR_CONTENTS),
-];
-/** Declared degrees: hubs 16/20, direction-linked 4, every neighbor 1. */
-function linkedFallbackDegrees() {
-    const degrees = {
-        hub_sat_16: 16,
-        hub_sat_20: 20,
-        dir_linked: 4,
-        dir_unlinked: 0,
-    };
-    for (const prefix of ["huba", "hubb", "dirn"]) {
-        const count = prefix === "huba"
-            ? HUB_A_NEIGHBOR_CONTENTS.length
-            : prefix === "hubb"
-                ? HUB_B_NEIGHBOR_CONTENTS.length
-                : DIR_NEIGHBOR_CONTENTS.length;
-        for (let i = 0; i < count; i++)
-            degrees[`${prefix}_n${i}`] = 1;
-    }
-    return degrees;
-}
-exports.LINKED_FALLBACK_FIXTURE = {
-    key: "linked_fallback",
-    description: "#449 (items 1+3): the connections reshape — a 16-link and a 20-link identical-content hub " +
-        "must score byte-equal post-change (both saturate at K=16), and an unlinked row ~0.20 more " +
-        "similar must outrank a 4-link rival (similarity dominates the saturated connections credit).",
-    rows: LINKED_FALLBACK_ROWS,
-    query: "Anchorholme guest anchorage",
-    distinctiveToken: "anchorholme",
-    expectedTop1: "dir_unlinked",
-    expectedUndirectedDegree: linkedFallbackDegrees(),
-    linkedExpectations: {
-        saturationPair: { hubA: "hub_sat_16", hubB: "hub_sat_20" },
-        directionPair: { unlinked: "dir_unlinked", linked: "dir_linked" },
-    },
-};
-// ---------------------------------------------------------------------------
-// Fixture invariants (checked by ranking-eval at run time AND pinned in
-// tests/ranking-eval-tooling.test.ts — a corpus that silently drifts from
-// its declared shape would fake a pass/fail)
-// ---------------------------------------------------------------------------
-/**
- * Undirected degree per row key from the declared linksTo arrays (#449):
- * a row's own linksTo entries PLUS the times other rows reference it — each
- * linksTo entry is one link row, so a mutual pair counts twice, mirroring
- * storage.getLinks(id, "both").length.
- */
-function undirectedDegrees(rows) {
-    const degree = new Map(rows.map((r) => [r.key, 0]));
-    for (const r of rows) {
-        for (const to of r.linksTo) {
-            degree.set(r.key, (degree.get(r.key) ?? 0) + 1);
-            degree.set(to, (degree.get(to) ?? 0) + 1);
-        }
-    }
-    return degree;
-}
-function checkFixtureInvariants(f) {
-    const problems = [];
-    const keys = new Set(f.rows.map((r) => r.key));
-    if (keys.size !== f.rows.length)
-        problems.push("duplicate row keys");
-    for (const r of f.rows) {
-        for (const to of r.linksTo) {
-            if (!keys.has(to))
-                problems.push(`${r.key} links to unknown key ${to}`);
-            if (to === r.key)
-                problems.push(`${r.key} links to itself`);
-        }
-    }
-    if (f.expectedTop1 && !keys.has(f.expectedTop1)) {
-        problems.push(`expectedTop1 ${f.expectedTop1} is not a row key`);
-    }
-    if (f.linkedExpectations) {
-        const declared = [
-            ["hubA", f.linkedExpectations.saturationPair.hubA],
-            ["hubB", f.linkedExpectations.saturationPair.hubB],
-            ["unlinked", f.linkedExpectations.directionPair.unlinked],
-            ["linked", f.linkedExpectations.directionPair.linked],
-        ];
-        for (const [role, key] of declared) {
-            if (!keys.has(key)) {
-                problems.push(`linkedExpectations ${role} ${key} is not a row key`);
-            }
-        }
-        const hubs = f.rows.filter((r) => r.key === f.linkedExpectations.saturationPair.hubA ||
-            r.key === f.linkedExpectations.saturationPair.hubB);
-        if (hubs.length === 2) {
-            const [a, b] = hubs;
-            const metaEqual = a.content === b.content &&
-                a.createdAt === b.createdAt &&
-                a.memoryType === b.memoryType &&
-                a.baseStrength === b.baseStrength &&
-                a.accessCount === b.accessCount &&
-                a.lastAccessedDaysAgo === b.lastAccessedDaysAgo;
-            if (!metaEqual) {
-                problems.push("saturation-pair hubs differ beyond their links (content/metadata must be identical)");
-            }
-        }
-    }
-    if (f.expectedUndirectedDegree) {
-        const degree = undirectedDegrees(f.rows);
-        for (const [key, expected] of Object.entries(f.expectedUndirectedDegree)) {
-            if (!keys.has(key)) {
-                problems.push(`expectedUndirectedDegree names unknown key ${key}`);
-                continue;
-            }
-            const actual = degree.get(key) ?? 0;
-            if (actual !== expected) {
-                problems.push(`${key} undirected degree is ${actual}, declared ${expected}`);
-            }
-        }
-    }
-    if (f.distinctiveToken) {
-        const token = f.distinctiveToken.toLowerCase();
-        const carriers = f.rows.filter((r) => r.content.toLowerCase().includes(token));
-        if (carriers.length !== 1) {
-            problems.push(`distinctive token "${f.distinctiveToken}" carried by ${carriers.length} rows (expected exactly 1)`);
-        }
-        else if (f.expectedTop1 && carriers[0].key !== f.expectedTop1) {
-            problems.push(`distinctive token carried by ${carriers[0].key}, not expectedTop1 ${f.expectedTop1}`);
-        }
-        if (!f.query.toLowerCase().includes(token)) {
-            problems.push(`query does not contain the distinctive token`);
-        }
-    }
-    return problems;
-}

package/dist/eval/recall-sweep.d.ts

@@ -1,87 +0,0 @@
-#!/usr/bin/env node
-/**
- * Recall@k sweep eval for session-intent keying (#192, commit 0867d6c).
- *
- * SHIP GATE: does blending the prompt with the per-session EMA centroid
- * (retrieval.blendQueryVector) improve recall across rephrased turns, AND
- * what does it COST when a session shifts topics mid-stream? Picks the
- * default `sessionIntentWeight`.
- *
- * Two scenario families, same corpus:
- *
- *   FOCUSED — 6 sessions (one per topic), 4 turns each of GENUINE REPHRASES
- *   of one intent. Tests the UPSIDE: does the centroid pull drifting
- *   rephrases back on-topic? (recall@5 / p@5 should rise with w.)
- *
- *   SHIFT — 5 sessions, each 2 turns on topic A then 2 turns on topic B (a
- *   clean mid-session topic change). Tests the DOWNSIDE that focused-only
- *   can't see: a higher weight leaves the centroid partly A right after the
- *   shift, so the first turn(s) on B may LAG (lower recall@5 for B). The
- *   load-bearing read-out is post-shift recovery at turn-3 (first B turn)
- *   and turn-4, plus turns-to-recover.
- *
- * Method (both families):
- *   - Synthetic corpus of 30 memories across 6 well-separated topics (5 each),
- *     embedded with the REAL bge-small-en-v1.5 model (the production embedder).
- *   - For each weight w in {0.0, 0.2, 0.4, 0.5, 0.6, 0.8}: fresh
- *     SessionRecallRegistry per session; per turn, embed the prompt ONCE,
- *     blend with the live centroid, pass the blended vector to retrieve() via
- *     queryEmbedding, record recall@5 + p@5 for the turn's CURRENT topic,
- *     then fold the prompt into the centroid (EMA α=0.4).
- *
- * Fairness controls (load-bearing — the eval is useless if these slip):
- *   - `noStrengthen: true` on every retrieve(): the DB stays STATIC across all
- *     retrieve calls. Strengthening would mutate effective_strength/access_count
- *     between runs and contaminate cross-weight comparisons.
- *   - Uniform memory metadata: every memory has base_strength=0.5, created_at
- *     ≈ now, access_count=0, and NO links. So effective_strength, recency,
- *     connections, and the freshness boost are all uniform → the ONLY
- *     discriminator is vector cosine + RRF rank. That isolates the blend.
- *   - Cold-exposure slots are a no-op here (all candidates equally cold),
- *     so the top-k is the plain score order.
- *   - The centroid update is weight-INDEPENDENT (EMA of prompts at α=0.4), so
- *     the per-turn centroid is identical across all weights for a given
- *     session; only the blend differs. Embeddings are precomputed once.
- *
- * Metrics (per turn, for the turn's CURRENT topic):
- *   - recall@5  — did >=1 same-topic memory surface in the top-5? (coarse; the
- *     requested ship-gate metric — saturates on a small corpus.)
- *   - p@5       — on-topic count in the top-5 (0..5, finer).
- *   - recall@1  — was the SINGLE top result on-topic? (finest; focused only.)
- *
- * Honest by construction: it prints whatever the numbers are, including
- * w=0 winning, the blend hurting a focused scenario, or high weights lagging
- * badly post-shift. Report written to data/eval-recall-sweep/report.md and
- * printed to stdout.
- *
- * ## Section 4 — SCOPE dimension (#203 soft project affinity)
- *
- * A THIRD family on a SEPARATE project-labeled corpus (the blend-sweep corpus
- * above has no project labels). Goal: prove #203's soft project affinity drops
- * cross-scope noise (the boat/battery case — a "hardware" query surfacing
- * "marine" memories on a shared token like "battery") below the cap WITHOUT a
- * hard filter, and WITHOUT losing same-scope recall.
- *
- *   Corpus: 6 hardware + 6 marine memories. 3 marine memories are
- *   CONTAMINATION SEEDS (share a token with the hardware queries); 3 are
- *   marine-only fillers (control).
- *
- *   Each hardware query runs TWICE on the same static DB:
- *     scope OFF — no `project` sent (byte-identical to pre-#203).
- *     scope ON  — `project: "hardware"` (computeScore adds +scopeAffinity 0.15
- *                 to hardware memories; marine gets 0; no hard filter).
- *
- *   Metrics: contamination@5 (marine in top-5 / 5 — LOWER is better),
- *   recall@5 (gold hardware memory surfaced), sameScope@5 (hardware in top-5).
- *   Scope is orthogonal to the blend, so the blend weight is held at 0 here —
- *   `project` is the ONLY variable. Same invariants (noStrengthen, real
- *   embedder, uniform metadata, neverCalledEmbed self-check).
- *
- * Run:  npm run eval:recall-sweep [--now <ISO>]   (==  node dist/eval/recall-sweep.js)
- *
- * `--now <ISO>` (#458) pins the eval clock: the planted corpora's created_at
- * AND every retrieve() call score against the SAME instant, so before/after
- * runs are wall-clock-independent. Default: the live clock (the pre-#458
- * behavior); the report header records which clock produced it.
- */
-export {};