@hviana/sema 0.2.6 → 0.2.8

package/dist/src/store-sqlite.js

@@ -0,0 +1,942 @@
+// store-sqlite.ts — SQLite + rabitq-ivf persistence adapter.
+//
+// SQliteStore extends AbstractStore and implements only the bare essentials
+// required for database communication: SQL schema, prepared statements, and
+// VectorDatabase management. All domain logic — caching, dedup/merge
+// decisions, interior-node indexing, geometric halo scheduling, buffer
+// management — lives in the AbstractStore base class (store.ts).
+//
+// Two clearly separated concerns, each with its OWN vector index:
+//
+//   • STORAGE  — the content-addressed DAG (nodes, edges, metadata) in SQLite,
+//     plus a VectorDatabase for node gists (resonant lookup), in its OWN
+//     SQLite file (`<path>.content.vec`).
+//   • CACHE    — the distributional "company" of each node (its halo), a
+//     derived/regenerable concept layer, in a SEPARATE VectorDatabase,
+//     in its own SQLite file (`<path>.halo.vec`).
+//
+// Write discipline:
+//
+//   • Integer node ids: a dense 0,1,2,… counter. SQLite stores them inline in
+//     the rowid B-tree (no secondary index), child lists pack to 4 bytes each,
+//     and the vector index keys on them as a 4-byte int32.
+//   • One deferred WRITE TRANSACTION: node rows, edges and halo updates are
+//     committed in batches on the flush cadence (and on close/snapshot) rather
+//     than one implicit transaction per node.
+//   • Each of the three SQLite databases (main DAG, content vectors, halo
+//     vectors) is an independent file. The vector databases persist to their
+//     own files on every write — no separate serialisation step is needed.
+import { DatabaseSync } from "node:sqlite";
+import { AbstractStore, flatBytesKids, packKids, unpackKids, } from "./store.js";
+import { DEFAULT_CONFIG } from "./config.js";
+import { VectorDatabase } from "./rabitq-ivf/src/index.js";
+const SCHEMA = `
+-- Content-addressed lookup goes through ONE small integer index: h is the
+-- FNV-1a hash of the row's content key (leaf bytes, or the packed kid ids of a
+-- mixed branch). Indexing the hash instead of the blob itself means the lookup
+-- index stores ~9 bytes per node rather than a full copy of every leaf and kid
+-- blob (which doubled the table); the point query fetches the h-candidates and
+-- verifies the actual blob, so collisions cost a fetch, never a wrong id.
+CREATE TABLE IF NOT EXISTS node (
+  id   INTEGER PRIMARY KEY,
+  leaf BLOB,
+  kids BLOB,
+  h    INTEGER NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_node_h ON node(h);
+-- Negative-lookup BLOOM FILTER over node.h — the RAM structure that lets the
+-- content-addressed dedup probes (findLeaf/findBranch, the training hot path)
+-- answer "definitely new content" without an idx_node_h descent.  Persisted
+-- here at checkpoint/close with upto = the next node id NOT covered; on
+-- open, rows with id >= upto (a crash's tail) are re-added by an incremental
+-- rowid-range scan, so the filter can NEVER be missing an inserted hash — a
+-- false negative would silently break hash-consing (duplicate mints), while
+-- a false positive only costs the SQLite probe that was paid on every call
+-- before the filter existed.
+CREATE TABLE IF NOT EXISTS bloom (
+  id   INTEGER PRIMARY KEY CHECK (id = 1),
+  bits BLOB NOT NULL,
+  n    INTEGER NOT NULL,
+  upto INTEGER NOT NULL
+);
+-- The reverse structural edge child→parent. A WITHOUT ROWID table clustered on
+-- (child, parent) IS the lookup index: parents(child) is one B-tree descent to a
+-- contiguous run, with no separate rowid heap and no secondary index to maintain
+-- (~2x smaller than a heap table + idx_kid_child). The composite primary key also
+-- gives free (child,parent) dedup, so a child repeated within one branch's kids
+-- never writes a duplicate row.
+CREATE TABLE IF NOT EXISTS kid (
+  child  INTEGER NOT NULL,
+  parent INTEGER NOT NULL,
+  PRIMARY KEY (child, parent)
+) WITHOUT ROWID;
+-- seq is a plain INTEGER PRIMARY KEY: rowids are assigned max+1 and edges are
+-- never deleted, so insertion order is preserved without AUTOINCREMENT's
+-- sqlite_sequence bookkeeping table.
+CREATE TABLE IF NOT EXISTS edge (
+  src INTEGER NOT NULL,
+  dst INTEGER NOT NULL,
+  seq INTEGER PRIMARY KEY
+);
+CREATE UNIQUE INDEX IF NOT EXISTS idx_edge ON edge(src, dst);
+CREATE INDEX IF NOT EXISTS idx_edge_dst ON edge(dst);
+CREATE TABLE IF NOT EXISTS halo (
+  id   INTEGER PRIMARY KEY,
+  vec  BLOB NOT NULL,
+  mass REAL NOT NULL DEFAULT 0
+);
+-- CONTAINMENT: parent's byte content contains child's, for children that are
+-- not among any parent's kids list (sub-span flat branches inside leaf-parent
+-- chunks). Kept SEPARATE from kid so parents() — the structural climb — is
+-- untouched; containers() serves the climbing surface of orphan flat branches
+-- durably (a session-local map would be lost on restart or ingest-cache replay).
+-- LOG-STRUCTURED PACKED PAGES: a child's parents are int32-packed blobs split
+-- across (child, seq) rows.  A flush APPENDS one page holding only the NEW
+-- parents (never reading the list back), then geometrically merges adjacent
+-- pages (byte concat; full dedup when a merge reaches the seq-0 base page), so
+-- page count stays O(log fan-in) and total merge I/O is amortized linear.
+-- This keeps the packed format's ~4 B/parent density (contain is ~11% of a
+-- trained store; one B-tree cell per pair measured 2.3× larger) while fixing
+-- the single-blob row's flaw: merging new parents into ONE row re-wrote the
+-- child's WHOLE list per flush — quadratic in a hot window's corpus-sized
+-- fan-in.  Transient duplicates may exist across pages until a base merge;
+-- readers dedup (containers()) or tolerate repeats by contract (slices).
+-- The (child, seq) key is packed into the INTEGER PRIMARY KEY (rowid) as
+-- child·2^8 + seq — exact in a JS double for every valid id, and a child's
+-- pages are one contiguous rowid range.  The geometric merge bounds live page
+-- counts (and thus seq) to O(log fan-in) ≪ 2^8, and a narrow band keeps the
+-- rowid varint within ~1 byte of the bare child id.  A rowid table, NOT a WITHOUT ROWID one: index B-tree cells cap at
+-- ~¼ page, so at 1 KiB pages any child beyond ~55 parents would spill each
+-- page row onto a mostly-empty overflow page (measured +26% table size).
+CREATE TABLE IF NOT EXISTS contain (
+  id      INTEGER PRIMARY KEY,
+  parents BLOB NOT NULL
+);
+-- CANONICAL-FORM index (Store.canonAdd/canonFind): h is the 32-bit hash of a
+-- node's CANONICAL content key (the modality's canonicalizer output — case-
+-- folded, whitespace-collapsed text, etc.; the store never sees the key
+-- itself, only its hash).  Same hash-then-verify discipline as idx_node_h:
+-- the caller re-canonicalizes each candidate's bytes before trusting it, so
+-- a collision costs a read, never a wrong id.  WITHOUT ROWID on (h, id):
+-- the composite key IS the lookup index and gives free dedup.
+CREATE TABLE IF NOT EXISTS canon (
+  h  INTEGER NOT NULL,
+  id INTEGER NOT NULL,
+  PRIMARY KEY (h, id)
+) WITHOUT ROWID;
+CREATE TABLE IF NOT EXISTS snapshot (
+  id INTEGER PRIMARY KEY CHECK (id = 1),
+  data BLOB NOT NULL
+);
+CREATE TABLE IF NOT EXISTS meta (
+  key TEXT PRIMARY KEY,
+  val TEXT NOT NULL
+);
+`;
+/** Blocked Bloom filter over 32-bit content hashes.  k=4 bit positions per
+ *  key via double hashing; at the maintained ≥12 bits/key the false-positive
+ *  rate stays under ~1%, so ≥99% of miss-probes skip SQLite entirely.
+ *  No deletions (nodes are never deleted), so growth is a rebuild. */
+class NodeBloom {
+    bits;
+    mask;
+    n = 0;
+    constructor(log2bits) {
+        this.bits = new Uint8Array(1 << (log2bits - 3));
+        this.mask = (1 << log2bits) - 1 >>> 0;
+    }
+    add(h) {
+        const h2 = (Math.imul(h, 0x9e3779b1) | 1) >>> 0;
+        let x = h >>> 0;
+        const bits = this.bits;
+        const mask = this.mask;
+        for (let i = 0; i < 4; i++) {
+            const b = x & mask;
+            bits[b >> 3] |= 1 << (b & 7);
+            x = (x + h2) >>> 0;
+        }
+        this.n++;
+    }
+    mightContain(h) {
+        const h2 = (Math.imul(h, 0x9e3779b1) | 1) >>> 0;
+        let x = h >>> 0;
+        const bits = this.bits;
+        const mask = this.mask;
+        for (let i = 0; i < 4; i++) {
+            const b = x & mask;
+            if ((bits[b >> 3] & (1 << (b & 7))) === 0)
+                return false;
+            x = (x + h2) >>> 0;
+        }
+        return true;
+    }
+    /** Whether the filter is due for a bigger rebuild (< 12 bits/key). */
+    get saturated() {
+        return this.n * 12 > this.bits.length * 8;
+    }
+}
+/** log2 of the bloom bit count sized for `n` keys at 16 bits/key, clamped to
+ *  [17, 30] — 16 KiB floor, 128 MiB ceiling (past ~90M nodes the FPR is
+ *  allowed to drift up rather than RAM). */
+function bloomLog2For(n) {
+    let log2 = 17;
+    while (log2 < 30 && (1 << log2) < n * 16)
+        log2++;
+    return log2;
+}
+/** Below this node count the filter is NOT persisted: a rebuild scan at open
+ *  is instant at this size, and the blob would dominate a small store's
+ *  on-disk footprint. */
+const BLOOM_PERSIST_MIN_NODES = 65_536;
+export class SQliteStore extends AbstractStore {
+    opts;
+    vectorCacheMb;
+    sqliteCacheMb;
+    vectorSeed;
+    // ── the two SEPARATE vector indices ───────────────────────────────────
+    content = null; // STORAGE: node gists
+    halos = null; // CACHE:   halo gists
+    sqlite = null;
+    // Deferred write transaction guard.
+    _inTx = false;
+    // ── negative dedup filter ──────────────────────────────────────────────
+    /** Bloom filter over every node row's content hash `h`.  The dedup probes
+     *  consult it first: "definitely absent" answers the training hot path's
+     *  dominant miss case in RAM (profiled: the SQLite probes were 38.7% of
+     *  ingest wall at 2.9M nodes).  INVARIANT: every inserted h is added
+     *  before the probe that could see it — a false negative would mint a
+     *  duplicate node and silently break hash-consing. */
+    _bloom = null;
+    /** Dedup probes answered by the filter alone this session (observability). */
+    bloomSkips = 0;
+    _insertNode = null;
+    _insertKid = null;
+    _selContain = null;
+    _selParents = null;
+    _selHalo = null;
+    _upHalo = null;
+    _selLeaf = null;
+    _selFlat = null;
+    _selKids = null;
+    _selNode = null;
+    _insertEdge = null;
+    _selNext = null;
+    _selPrev = null;
+    _setMeta = null;
+    _getMeta = null;
+    _insCanon = null;
+    _selCanon = null;
+    _cntCanon = null;
+    _selContentFrom = null;
+    _delMeta = null;
+    _insSnapshot = null;
+    _selSnapshot = null;
+    constructor(opts = {}) {
+        const d = DEFAULT_CONFIG.store;
+        // Resolve config from opts, falling back to defaults.
+        const config = {
+            minHaloMass: opts.minHaloMass ?? d.minHaloMass,
+            efSearch: opts.efSearch ?? d.efSearch,
+            compactEveryNWrites: opts.compactEveryNWrites ?? d.compactEveryNWrites,
+            overfetch: opts.overfetch ?? d.overfetch,
+            batchSize: opts.batchSize ?? d.batchSize,
+            dedupCacheMax: opts.dedupCacheMax ?? d.dedupCacheMax,
+            bytesCacheMax: opts.bytesCacheMax ?? d.bytesCacheMax,
+            recCacheBytes: opts.recCacheBytes ?? d.recCacheBytes,
+            ingestCacheBytes: opts.ingestCacheBytes ?? d.ingestCacheBytes,
+            pendingGistBytes: opts.pendingGistBytes ?? d.pendingGistBytes,
+            haloCacheBytes: opts.haloCacheBytes ?? d.haloCacheBytes,
+            vectorCacheMb: opts.vectorCacheMb ?? d.vectorCacheMb,
+            sqliteCacheMb: opts.sqliteCacheMb ?? d.sqliteCacheMb,
+            coveredIdsMax: opts.coveredIdsMax ?? d.coveredIdsMax,
+            chainCacheBytes: opts.chainCacheBytes ?? d.chainCacheBytes,
+        };
+        const D = opts.D ?? 1024;
+        const maxGroup = opts.maxGroup ?? DEFAULT_CONFIG.geometry.maxGroup;
+        super(config, D, maxGroup);
+        this.opts = opts;
+        this.vectorCacheMb = opts.vectorCacheMb ?? d.vectorCacheMb;
+        this.sqliteCacheMb = opts.sqliteCacheMb ?? d.sqliteCacheMb;
+        this.vectorSeed = (0x51f15e ^ 0x9e3779b9) >>> 0;
+        this._ready = this._dbOpen();
+    }
+    // ── Vector-DB paths ───────────────────────────────────────────────────
+    vectorDbPath(name) {
+        const stem = this.opts.path ?? ":memory:";
+        if (stem === ":memory:")
+            return ":memory:";
+        return `${stem}.${name}.vec`;
+    }
+    openVectorDB(name) {
+        return new VectorDatabase({
+            dbPath: this.vectorDbPath(name),
+            dim: this.D,
+            efSearch: this.efSearch,
+            // Query-side estimator precision.  8 bits: 4 bits measurably misranks
+            // tight gist clusters (mixture recall@10 37.5% vs 39.0%, self-recall
+            // 69% vs 76%; rabitq test 2a), and the codes on disk are independent of
+            // it, so existing stores adopt the sharper setting on reopen.  The one
+            // decision the sharper ranking exposed — a saturated hub winning a
+            // sub-noise rank tie and silencing its region — is handled by the
+            // tie-band saturation fallback in attention.ts, which is derived from
+            // the estimator's own margin noise rather than tuned to a bit width.
+            queryBits: this.opts.queryBits ?? 8,
+            seed: this.vectorSeed,
+            cacheSizeMb: this.vectorCacheMb,
+        });
+    }
+    // ── Abstract method implementations ───────────────────────────────────
+    // -- Lifecycle --
+    async _dbOpen() {
+        const stem = this.opts.path ?? ":memory:";
+        const sp = stem === ":memory:" ? ":memory:" : `${stem}.sqlite`;
+        this.sqlite = new DatabaseSync(sp);
+        // Page size matched to the store's row grain BEFORE any table exists (a
+        // no-op on a non-empty database). Every table here holds small rows —
+        // packed kid/contain pairs, byte-packed flat branches, quantized halos —
+        // so SQLite's default 4 KiB pages carry mostly slack; 1 KiB pages keep
+        // leaf-page fill high at a negligible extra tree depth.
+        this.sqlite.exec("PRAGMA page_size = 1024;");
+        // WAL + NORMAL sync: crash-safe (WAL commits are atomic) without a full
+        // fsync per commit — the same discipline the vector databases use.
+        this.sqlite.exec("PRAGMA journal_mode = WAL; PRAGMA synchronous = NORMAL;");
+        // SQLite's autocheckpoint default is 1000 PAGES — 1 MiB at this store's
+        // 1 KiB page size — so at training write rates nearly every batch commit
+        // crossed the threshold and paid a synchronous checkpoint: random writes
+        // into a GB-scale main file, with the hottest B-tree pages copied out on
+        // every commit instead of being coalesced across commits in the WAL.
+        // Profiled on a 7.5M-node store: ~114 ms PER COMMIT, ~19% of ingest
+        // wall-clock.  16 MiB of WAL between checkpoints coalesces hot-page
+        // rewrites and amortizes the checkpoint's fixed cost; crash-safety is
+        // unchanged (WAL recovery replays it), the only trade is a bounded,
+        // fixed-size WAL file.
+        this.sqlite.exec("PRAGMA wal_autocheckpoint = 65536;");
+        // Page cache for the DAG tables.  Content-addressed dedup (findLeaf /
+        // findBranch), parent probes and contain appends are millions of point
+        // queries per session against a file that outgrows SQLite's ~2 MiB
+        // default cache almost immediately; every miss is a file read on the
+        // ingest hot path.  Negative = KiB budget (size-based, page-size
+        // independent).  Latency only — results identical at any value.
+        const cacheKb = Math.max(0, Math.floor(this.sqliteCacheMb * 1024));
+        if (cacheKb > 0)
+            this.sqlite.exec(`PRAGMA cache_size = -${cacheKb};`);
+        this.sqlite.exec(SCHEMA);
+        // Recover D from a previous run so loadFromStore() can bootstrap with any
+        // default — the real value is read from meta before vector indices load.
+        // The meta table always exists here (SCHEMA above creates it), so these
+        // reads run bare: a throw is a genuine storage fault, and swallowing it
+        // would let the store proceed with a default D/maxGroup that silently
+        // disagrees with what training persisted.
+        {
+            const row = this.sqlite.prepare("SELECT val FROM meta WHERE key = 'train.D'").get();
+            if (row) {
+                const d = Number(row.val);
+                if (Number.isInteger(d) && d > 0)
+                    this._D = d;
+            }
+        }
+        // Recover maxGroup from meta so indexSubtree uses the training-time value.
+        {
+            const row = this.sqlite.prepare("SELECT val FROM meta WHERE key = 'geometry.maxGroup'").get();
+            if (row) {
+                const g = Number(row.val);
+                if (Number.isInteger(g) && g > 0)
+                    this._maxGroup = g;
+            }
+        }
+        // Persist maxGroup to meta when opening a FRESH store (no rows yet) so
+        // indexSubtree always sees the training-time value even when the store is
+        // accessed without a Mind / full snapshot.
+        if (this._nextId === 0) {
+            this.sqlite.prepare("INSERT OR IGNORE INTO meta (key, val) VALUES ('geometry.maxGroup', ?)").run(String(this._maxGroup));
+        }
+        this.content = this.openVectorDB("content");
+        this.halos = this.openVectorDB("halo");
+        // Ids are dense (0,1,2,…) and never deleted, so MAX(id)+1 IS the next id
+        // — one rightmost B-tree descent on the rowid, O(log n).  (COUNT(*) gives
+        // the same value but scans every leaf page, so open time grew linearly
+        // with the store; a per-row scan into a Set was worse still.)
+        this._nextId = (this.sqlite.prepare("SELECT MAX(id) AS m FROM node").get().m ?? -1) + 1;
+        this._bloomLoad();
+    }
+    // ── negative dedup filter: load / rebuild / persist ────────────────────
+    /** Restore the persisted filter and top up the rows a crash left uncovered
+     *  (id ≥ stored `upto`); a missing/undersized blob forces a full rebuild.
+     *  Clean-close sessions restart with a 0-row top-up. */
+    _bloomLoad() {
+        const row = this.sqlite.prepare("SELECT bits, n, upto FROM bloom WHERE id = 1").get();
+        const wantLog2 = bloomLog2For(this._nextId);
+        if (row && row.bits.length * 8 >= (1 << wantLog2)) {
+            const log2 = 31 - Math.clz32(row.bits.length * 8);
+            const b = new NodeBloom(log2);
+            b.bits.set(row.bits);
+            b.n = row.n;
+            this._bloom = b;
+            this._bloomTopUp(row.upto);
+        }
+        else {
+            this._bloomRebuild();
+        }
+    }
+    /** Re-add the hashes of rows with id ≥ `from` (the tail a crash cut off).
+     *  Re-adding a hash twice only re-sets its bits — never a correctness
+     *  issue — so overlap with the persisted coverage is harmless. */
+    _bloomTopUp(from) {
+        if (from >= this._nextId)
+            return;
+        const b = this._bloom;
+        const scan = this.sqlite.prepare("SELECT h FROM node WHERE id >= ?");
+        scan.setReturnArrays(true);
+        for (const r of scan.iterate(from)) {
+            b.add(r[0]);
+        }
+    }
+    /** Build a fresh filter sized for the current node count from one
+     *  sequential scan of the h column.  Called at open when no usable blob is
+     *  stored, and again whenever growth saturates the bits-per-key budget. */
+    _bloomRebuild() {
+        const b = new NodeBloom(bloomLog2For(this._nextId));
+        b.n = 0;
+        if (this._nextId > 0) {
+            const scan = this.sqlite.prepare("SELECT h FROM node");
+            scan.setReturnArrays(true);
+            for (const r of scan.iterate()) {
+                // add() counts n itself.
+                b.add(r[0]);
+            }
+        }
+        this._bloom = b;
+    }
+    /** Persist the filter with its coverage watermark.  Runs inside the
+     *  caller's transaction when one is open (checkpoint/close both flush
+     *  first). */
+    _bloomPersist() {
+        if (!this._bloom || !this.sqlite)
+            return;
+        if (this._nextId < BLOOM_PERSIST_MIN_NODES)
+            return;
+        this.sqlite.prepare("INSERT INTO bloom (id, bits, n, upto) VALUES (1, ?, ?, ?) " +
+            "ON CONFLICT(id) DO UPDATE SET bits = excluded.bits, " +
+            "n = excluded.n, upto = excluded.upto").run(this._bloom.bits, this._bloom.n, this._nextId);
+    }
+    _dbClose() {
+        if (this.sqlite)
+            this._bloomPersist();
+        if (this.content) {
+            this.content.close();
+            this.content = null;
+        }
+        if (this.halos) {
+            this.halos.close();
+            this.halos = null;
+        }
+        if (this.sqlite) {
+            this.sqlite.close();
+            this.sqlite = null;
+        }
+    }
+    // -- Transaction --
+    _dbBeginTx() {
+        if (this._inTx || !this.sqlite)
+            return;
+        this.sqlite.exec("BEGIN");
+        this._inTx = true;
+    }
+    _dbCommitTx() {
+        if (!this._inTx || !this.sqlite)
+            return;
+        this._inTx = false; // clear first so a throw can't wedge us mid-commit
+        this.sqlite.exec("COMMIT");
+    }
+    // -- Node CRUD --
+    _dbInsertNode(id, leaf, kids, h) {
+        if (!this._insertNode) {
+            this._insertNode = this.sqlite.prepare("INSERT INTO node (id, leaf, kids, h) VALUES (?, ?, ?, ?)");
+        }
+        this._insertNode.run(id, leaf, kids, h);
+        // The filter must see every inserted hash BEFORE any probe could — the
+        // add sits in the same synchronous call as the row insert.
+        const b = this._bloom;
+        b.add(h);
+        if (b.saturated)
+            this._bloomRebuild();
+    }
+    _dbGetNode(id) {
+        if (!this._selNode) {
+            this._selNode = this.sqlite.prepare("SELECT id, leaf, kids FROM node WHERE id = ?");
+        }
+        const r = this._selNode.get(id);
+        if (!r)
+            return null;
+        // A zero-length kids blob marks a FLAT branch: the leaf column holds its
+        // bytes and the kid list is derived, one implicit leaf per byte.
+        const flat = r.kids !== null && r.kids.byteLength === 0;
+        return {
+            id: r.id,
+            leaf: r.leaf && !flat ? new Uint8Array(r.leaf) : null,
+            kids: flat
+                ? flatBytesKids(r.leaf)
+                : (r.kids ? unpackKids(r.kids) : null),
+        };
+    }
+    _dbFindLeaf(h, bytes) {
+        if (this._bloom && !this._bloom.mightContain(h)) {
+            this.bloomSkips++;
+            return null;
+        }
+        if (!this._selLeaf) {
+            this._selLeaf = this.sqlite.prepare("SELECT id FROM node WHERE h = ? AND leaf = ? AND kids IS NULL LIMIT 1");
+        }
+        const row = this._selLeaf.get(h, bytes);
+        return row ? row.id : null;
+    }
+    _dbFindBranchByLeaf(h, bytes) {
+        if (this._bloom && !this._bloom.mightContain(h)) {
+            this.bloomSkips++;
+            return null;
+        }
+        if (!this._selFlat) {
+            this._selFlat = this.sqlite.prepare("SELECT id FROM node WHERE h = ? AND leaf = ? AND kids IS NOT NULL LIMIT 1");
+        }
+        const row = this._selFlat.get(h, bytes);
+        return row ? row.id : null;
+    }
+    _dbFindBranchByKids(h, packed) {
+        if (this._bloom && !this._bloom.mightContain(h)) {
+            this.bloomSkips++;
+            return null;
+        }
+        if (!this._selKids) {
+            this._selKids = this.sqlite.prepare("SELECT id FROM node WHERE h = ? AND kids = ? LIMIT 1");
+        }
+        const row = this._selKids.get(h, packed);
+        return row ? row.id : null;
+    }
+    // -- Kid (structural parent) edges --
+    _dbInsertKid(child, parent) {
+        if (!this._insertKid) {
+            this._insertKid = this.sqlite.prepare("INSERT OR IGNORE INTO kid (child, parent) VALUES (?, ?)");
+        }
+        this._insertKid.run(child, parent);
+    }
+    _dbGetParents(id) {
+        if (!this._selParents) {
+            this._selParents = this.sqlite.prepare("SELECT parent FROM kid WHERE child = ?");
+        }
+        return this._selParents.all(id).map((r) => r.parent);
+    }
+    // -- Containment --
+    _selParentsFirst = null;
+    _dbGetParentsFirst(id, limit) {
+        if (!this._selParentsFirst) {
+            this._selParentsFirst = this.sqlite.prepare("SELECT parent FROM kid WHERE child = ? LIMIT ?");
+        }
+        return this._selParentsFirst.all(id, limit)
+            .map((r) => r.parent);
+    }
+    _selChain = null;
+    /** {@link Store.chainRun}'s walk as ONE recursive CTE: the whole
+     *  transparent chain (no edge in or out, exactly one parent) is descended
+     *  inside SQLite — per-node work is the same three indexed probes as the
+     *  base class's loop, but without a JS↔SQLite round trip per node, which
+     *  dominates on the deep single-structure scaffolding this read exists
+     *  for. */
+    _chainWalk(id, cap) {
+        if (!this._selChain) {
+            this._selChain = this.sqlite.prepare(`WITH RECURSIVE chain(n, d) AS (
+           SELECT ?, 0
+           UNION ALL
+           SELECT (SELECT parent FROM kid WHERE child = chain.n LIMIT 1),
+                  chain.d + 1
+           FROM chain
+           WHERE chain.d < ?
+             AND NOT EXISTS (SELECT 1 FROM edge WHERE src = chain.n)
+             AND NOT EXISTS (SELECT 1 FROM edge WHERE dst = chain.n)
+             AND (SELECT count(*)
+                    FROM (SELECT 1 FROM kid WHERE child = chain.n LIMIT 2)
+                 ) = 1
+         )
+         SELECT n FROM chain ORDER BY d`);
+        }
+        return this._selChain.all(id, cap - 1)
+            .map((r) => r.n);
+    }
+    /** Width of the seq band inside the packed contain rowid: rowid =
+     *  child·SEQ_SPAN + seq.  Page sizes are geometric, so a chain of k live
+     *  pages needs a base of ≥ 2^k · 4 bytes — seq stays ≤ ~50 for any
+     *  physically possible list; the append path guards the band anyway. */
+    static SEQ_SPAN = 1 << 8;
+    _selContainExists = null;
+    _dbContainExists(child) {
+        if (!this._selContainExists) {
+            // A probe of the child's rowid range — one descent.
+            this._selContainExists = this.sqlite.prepare("SELECT 1 AS one FROM contain WHERE id >= ? AND id < ? LIMIT 1");
+        }
+        const base = child * SQliteStore.SEQ_SPAN;
+        return this._selContainExists.get(base, base + SQliteStore.SEQ_SPAN) !==
+            undefined;
+    }
+    _selContainPages = null;
+    /** The child's page directory — (seq, byte length) per page, O(log fan-in)
+     *  rows, each read from the cell header without loading the blob. */
+    containPages(child) {
+        if (!this._selContainPages) {
+            this._selContainPages = this.sqlite.prepare("SELECT id, length(parents) AS len FROM contain " +
+                "WHERE id >= ? AND id < ? ORDER BY id");
+        }
+        const base = child * SQliteStore.SEQ_SPAN;
+        return this._selContainPages.all(base, base + SQliteStore.SEQ_SPAN)
+            .map((r) => ({ seq: r.id - base, len: r.len }));
+    }
+    _selContainPageSub = null;
+    _selContainPage = null;
+    containPage(child, seq) {
+        if (!this._selContainPage) {
+            this._selContainPage = this.sqlite.prepare("SELECT parents FROM contain WHERE id = ?");
+        }
+        return this._selContainPage.get(child * SQliteStore.SEQ_SPAN + seq).parents;
+    }
+    _dbGetContainParentsSlice(child, offset, limit) {
+        if (!this._selContainPageSub) {
+            // substr on a BLOB returns bytes: unpack only the requested span.
+            this._selContainPageSub = this.sqlite.prepare("SELECT substr(parents, 1 + ? * 4, ? * 4) AS page " +
+                "FROM contain WHERE id = ?");
+        }
+        const out = [];
+        let skip = offset;
+        for (const seg of this.containPages(child)) {
+            const n = seg.len >>> 2;
+            if (skip >= n) {
+                skip -= n;
+                continue;
+            }
+            const take = Math.min(limit - out.length, n - skip);
+            const row = this._selContainPageSub.get(skip, take, child * SQliteStore.SEQ_SPAN + seg.seq);
+            if (row && row.page.length > 0)
+                out.push(...unpackKids(row.page));
+            skip = 0;
+            if (out.length >= limit)
+                break;
+        }
+        return out;
+    }
+    _selContainCount = null;
+    _dbGetContainCount(child) {
+        if (!this._selContainCount) {
+            // Stored ENTRY count (transient duplicates included) — consistent with
+            // what the slice streams, which is all the seam math needs.
+            this._selContainCount = this.sqlite.prepare("SELECT COALESCE(SUM(length(parents)), 0) / 4 AS n " +
+                "FROM contain WHERE id >= ? AND id < ?");
+        }
+        const base = child * SQliteStore.SEQ_SPAN;
+        return this._selContainCount.get(base, base + SQliteStore.SEQ_SPAN).n;
+    }
+    _dbGetContainParents(child) {
+        if (!this._selContain) {
+            this._selContain = this.sqlite.prepare("SELECT parents FROM contain WHERE id >= ? AND id < ? ORDER BY id");
+        }
+        const base = child * SQliteStore.SEQ_SPAN;
+        const rows = this._selContain.all(base, base + SQliteStore.SEQ_SPAN);
+        if (rows.length === 0)
+            return [];
+        // Dedup across pages (a pair re-added after its page merged into the base
+        // may repeat in the tail until the next base merge), preserving order.
+        const seen = new Set();
+        const out = [];
+        for (const r of rows) {
+            for (const p of unpackKids(r.parents)) {
+                if (!seen.has(p)) {
+                    seen.add(p);
+                    out.push(p);
+                }
+            }
+        }
+        return out;
+    }
+    _upsContainPage = null;
+    _delContainPage = null;
+    _dbAppendContain(child, parents) {
+        if (parents.length === 0)
+            return;
+        if (!this._upsContainPage) {
+            this._upsContainPage = this.sqlite.prepare("INSERT INTO contain (id, parents) VALUES (?, ?) " +
+                "ON CONFLICT(id) DO UPDATE SET parents = excluded.parents");
+            this._delContainPage = this.sqlite.prepare("DELETE FROM contain WHERE id = ?");
+        }
+        // Fold stored tail pages into the new page IN MEMORY, then write ONCE.
+        // Merge rule per fold: below a 256-byte floor always merge (most children
+        // have a handful of parents — two tiny rows would double their per-row
+        // overhead for no amortization benefit); above it, merge only while the
+        // stored page is ≤ 2× the accumulated one, which bounds total re-writing
+        // to O(fan-in · log fan-in) bytes and keeps the page directory
+        // logarithmic.  A fold that reaches the seq-0 base page dedups, squeezing
+        // out duplicates on the same geometric schedule.  Writing the fold's
+        // result as one upsert (an in-place row replace in the common
+        // small-child case) rather than delete+insert churn keeps B-tree leaf
+        // fill at the packed single-row format's level — measured: churn alone
+        // cost ~26% extra pages on the same payload.
+        const segs = this.containPages(child);
+        let cur = packKids(parents);
+        let seq = segs.length > 0 ? segs[segs.length - 1].seq + 1 : 0;
+        if (seq >= SQliteStore.SEQ_SPAN) {
+            throw new Error(`contain page seq overflow for child ${child}`);
+        }
+        const folded = [];
+        while (segs.length > 0) {
+            const a = segs[segs.length - 1];
+            if (a.len + cur.byteLength > 256 && a.len > 2 * cur.byteLength)
+                break;
+            const pa = this.containPage(child, a.seq);
+            if (a.seq === 0) {
+                const seen = new Set();
+                const ids = [];
+                for (const p of unpackKids(pa)) {
+                    if (!seen.has(p)) {
+                        seen.add(p);
+                        ids.push(p);
+                    }
+                }
+                for (const p of unpackKids(cur)) {
+                    if (!seen.has(p)) {
+                        seen.add(p);
+                        ids.push(p);
+                    }
+                }
+                cur = packKids(ids);
+            }
+            else {
+                const m = new Uint8Array(pa.byteLength + cur.byteLength);
+                m.set(pa, 0);
+                m.set(cur, pa.byteLength);
+                cur = m;
+            }
+            folded.push(a.seq);
+            seq = a.seq;
+            segs.pop();
+        }
+        const base = child * SQliteStore.SEQ_SPAN;
+        for (const s of folded) {
+            if (s !== seq)
+                this._delContainPage.run(base + s);
+        }
+        this._upsContainPage.run(base + seq, cur);
+    }
+    // -- Continuation edges --
+    _dbInsertEdge(src, dst) {
+        if (!this._insertEdge) {
+            this._insertEdge = this.sqlite.prepare("INSERT OR IGNORE INTO edge (src, dst) VALUES (?, ?)");
+        }
+        this._insertEdge.run(src, dst);
+    }
+    _dbGetNextEdges(id) {
+        if (!this._selNext) {
+            this._selNext = this.sqlite.prepare("SELECT dst FROM edge WHERE src = ? ORDER BY seq ASC");
+        }
+        return this._selNext.all(id).map((r) => r.dst);
+    }
+    _dbGetPrevEdges(id) {
+        if (!this._selPrev) {
+            this._selPrev = this.sqlite.prepare("SELECT src FROM edge WHERE dst = ? ORDER BY seq DESC");
+        }
+        return this._selPrev.all(id).map((r) => r.src);
+    }
+    _selNextFirst = null;
+    _dbGetNextEdgesFirst(id, limit) {
+        if (!this._selNextFirst) {
+            this._selNextFirst = this.sqlite.prepare("SELECT dst FROM edge WHERE src = ? ORDER BY seq ASC LIMIT ?");
+        }
+        return this._selNextFirst.all(id, limit)
+            .map((r) => r.dst);
+    }
+    _selPrevFirst = null;
+    _dbGetPrevEdgesFirst(id, limit) {
+        if (!this._selPrevFirst) {
+            this._selPrevFirst = this.sqlite.prepare("SELECT src FROM edge WHERE dst = ? ORDER BY seq DESC LIMIT ?");
+        }
+        return this._selPrevFirst.all(id, limit)
+            .map((r) => r.src);
+    }
+    _selPrevCount = null;
+    /** {@link Store.prevCount} — one indexed COUNT over idx_edge_dst, never a
+     *  row materialisation (a common continuation's reverse fan-in is
+     *  corpus-sized). */
+    prevCount(id) {
+        if (!this._selPrevCount) {
+            this._selPrevCount = this.sqlite.prepare("SELECT COUNT(*) AS n FROM edge WHERE dst = ?");
+        }
+        return this._selPrevCount.get(id).n;
+    }
+    _selSrcExists = null;
+    _dbEdgeSrcExists(src) {
+        if (!this._selSrcExists) {
+            // A prefix probe of idx_edge(src, dst) — one B-tree descent.
+            this._selSrcExists = this.sqlite.prepare("SELECT 1 AS one FROM edge WHERE src = ? LIMIT 1");
+        }
+        return this._selSrcExists.get(src) !== undefined;
+    }
+    _dbEdgeDistinctSrcCount() {
+        return this.sqlite.prepare("SELECT COUNT(*) AS n FROM (SELECT DISTINCT src FROM edge)").get().n;
+    }
+    // -- Halos (durable 2-bit quantized rows) --
+    _dbGetHalo(id) {
+        if (!this._selHalo) {
+            this._selHalo = this.sqlite.prepare("SELECT vec, mass FROM halo WHERE id = ?");
+        }
+        const r = this._selHalo.get(id);
+        return r ?? null;
+    }
+    _dbUpsertHalo(id, encodedVec, mass) {
+        if (!this._upHalo) {
+            this._upHalo = this.sqlite.prepare("INSERT INTO halo (id, vec, mass) VALUES (?, ?, ?) " +
+                "ON CONFLICT(id) DO UPDATE SET vec = excluded.vec, mass = excluded.mass");
+        }
+        this._upHalo.run(id, encodedVec, mass);
+    }
+    // -- Meta --
+    _dbGetMeta(key) {
+        if (!this._getMeta) {
+            this._getMeta = this.sqlite.prepare("SELECT val FROM meta WHERE key = ?");
+        }
+        const row = this._getMeta.get(key);
+        return row ? row.val : null;
+    }
+    _dbSetMeta(key, val) {
+        if (!this._setMeta) {
+            this._setMeta = this.sqlite.prepare("INSERT INTO meta (key, val) VALUES (?, ?) " +
+                "ON CONFLICT(key) DO UPDATE SET val = excluded.val");
+        }
+        this._setMeta.run(key, val);
+    }
+    _dbDeleteMeta(key) {
+        if (!this._delMeta) {
+            this._delMeta = this.sqlite.prepare("DELETE FROM meta WHERE key = ?");
+        }
+        this._delMeta.run(key);
+    }
+    // -- Canonical-form index (Store optional capability) --
+    canonAdd(h, id) {
+        if (!this._insCanon) {
+            this._insCanon = this.sqlite.prepare("INSERT OR IGNORE INTO canon (h, id) VALUES (?, ?)");
+        }
+        // Join the deferred write transaction (committed by flush/commit), so a
+        // bulk index build coalesces instead of paying autocommit per row.
+        this._dbBeginTx();
+        this._insCanon.run(h, id);
+    }
+    canonFind(h) {
+        if (!this._selCanon) {
+            this._selCanon = this.sqlite.prepare("SELECT id FROM canon WHERE h = ?");
+        }
+        return this._selCanon.all(h).map((r) => r.id);
+    }
+    canonCount() {
+        if (!this._cntCanon) {
+            this._cntCanon = this.sqlite.prepare("SELECT count(*) AS c FROM canon");
+        }
+        return this._cntCanon.get().c;
+    }
+    eachContent(cb, fromId = 0) {
+        if (!this._selContentFrom) {
+            // Content-bearing nodes are FLAT branches: leaf present, kids an
+            // empty (zero-length) blob — the same population PART 2 of the
+            // diagnostics calls "distinct content spans".
+            this._selContentFrom = this.sqlite.prepare("SELECT id, leaf FROM node " +
+                "WHERE id >= ? AND leaf IS NOT NULL " +
+                "AND (kids IS NULL OR length(kids) = 0)");
+        }
+        for (const row of this._selContentFrom.iterate(fromId)) {
+            const r = row;
+            cb(r.id, new Uint8Array(r.leaf));
+        }
+    }
+    // -- Snapshot --
+    _dbSaveSnapshot(bytes) {
+        if (!this._insSnapshot) {
+            this._insSnapshot = this.sqlite.prepare("INSERT INTO snapshot (id, data) VALUES (1, ?) " +
+                "ON CONFLICT(id) DO UPDATE SET data = excluded.data");
+        }
+        this._insSnapshot.run(bytes);
+        // Checkpoint cadence: persisting the dedup filter here keeps the crash
+        // top-up (the id ≥ upto rescan at next open) bounded by one checkpoint
+        // interval instead of the whole session.
+        this._bloomPersist();
+    }
+    _dbLoadSnapshot() {
+        if (!this._selSnapshot) {
+            this._selSnapshot = this.sqlite.prepare("SELECT data FROM snapshot WHERE id = 1");
+        }
+        const r = this._selSnapshot.get();
+        return r ? new Uint8Array(r.data) : null;
+    }
+    // -- Vector DB: content (gist) index --
+    _vecContentUpsert(entries) {
+        this.content.upsertMany(entries);
+    }
+    _vecContentQuery(v, k, ef) {
+        return this.content.query(v, k, { ef });
+    }
+    _vecContentHas(id) {
+        return this.content ? this.content.has(id) : false;
+    }
+    _vecContentSize() {
+        return this.content ? this.content.size : 0;
+    }
+    _vecContentLastReads() {
+        return this.content ? this.content.lastQueryStorageReads : 0;
+    }
+    _vecContentPhysicalSize() {
+        return this.content ? this.content.physicalSize : 0;
+    }
+    _vecContentClusterCount() {
+        return this.content ? this.content.clusterCount : 0;
+    }
+    _vecContentCompact() {
+        this.content.compact();
+    }
+    _vecContentDeleteMany(ids) {
+        this.content.deleteMany(ids);
+    }
+    *_vecContentEntriesSince(after) {
+        if (this.content)
+            yield* this.content.keysSince(after);
+    }
+    /** {@link AbstractStore._dbEdgeOrHaloIds} — one C-side scan; UNION dedups
+     *  and (with the ORDER BY) sorts, so the result is ready for the binary-
+     *  search membership probes and the ascending-id maintenance walks. */
+    _dbEdgeOrHaloIds() {
+        const rows = this.sqlite.prepare("SELECT src AS id FROM edge UNION SELECT dst FROM edge " +
+            "UNION SELECT id FROM halo ORDER BY 1").all();
+        const out = new Array(rows.length);
+        for (let i = 0; i < rows.length; i++)
+            out[i] = rows[i].id;
+        return out;
+    }
+    // -- Vector DB: halo index --
+    _vecHaloUpsert(entries) {
+        this.halos.upsertMany(entries);
+    }
+    _vecHaloQuery(v, k, ef) {
+        return this.halos.query(v, k, { ef });
+    }
+    _vecHaloSize() {
+        return this.halos ? this.halos.size : 0;
+    }
+    _vecHaloPhysicalSize() {
+        return this.halos ? this.halos.physicalSize : 0;
+    }
+    _vecHaloClusterCount() {
+        return this.halos ? this.halos.clusterCount : 0;
+    }
+    _vecHaloCompact() {
+        this.halos.compact();
+    }
+    /** Pre-fill both vector indices' RAM caches with sequential scans (up to
+     *  their budget caps) — seconds of streaming instead of the minutes of
+     *  random point reads a cold training/query session otherwise pays while
+     *  warming.  Optional; call once after open before sustained work.
+     *  Returns rows warmed. */
+    async warmVectorCaches() {
+        await this._ensureReady();
+        return (this.content?.warmCache() ?? 0) + (this.halos?.warmCache() ?? 0);
+    }
+}