sensemaking 0.19.1 → 0.20.0

package/dist/esm/store/sqlite/open.js

@@ -1,23 +1,21 @@
 // The Node floor (>=22.20) is explained here and nowhere else: 22.20 is the first release with
 // both FTS5 and row-returning INSERT ... RETURNING. Raise it only for a load-bearing capability.
-import { mkdirSync } from 'node:fs';
-import { join } from 'node:path';
 import { DatabaseSync } from 'node:sqlite';
-import { contentTokenize, featureSignature, STATE_DIR } from '../../config/index.js';
-import { rekeyChunkText } from '../../embed/handoff.js';
+import { contentTokenize, featureSignature } from '../../config/index.js';
 import { SenseError } from '../../errors.js';
 import { activeFeatures, FEATURES } from '../../features/index.js';
-import { clearCache } from '../cache.js';
+import { listFiles, parseFile } from '../../scan/index.js';
+import { openWithDialect } from '../open.js';
 import { getMeta, setMeta } from '../shared.js';
 import { BEGIN_WRITE, withTransaction } from '../transaction.js';
 import { createConnection } from './connection.js';
-import { changedSignatureKeys, embedIdentityAdopted, rebuildContentTable, reconcile, signatureDiff } from './reconcile.js';
+import { contentRow, INSERT_CONTENT_SQL, sqliteDialect } from './reconcile.js';
 import { registerFunctions } from './sql-functions.js';
 import { createStore } from './store.js';
 export const DB_FILENAME = 'cache.db';
 // Cache shape version, independent of the config's own `version`. Bumping it rebuilds
 // existing trees on first query.
-export const SCHEMA_VERSION = '18';
+export const SCHEMA_VERSION = '19';
 // unicode61 splits on spaces, so a language written without them indexes a whole run as one
 // token and word search finds nothing; `content.tokenize` is how such a tree picks trigram.
 const DEFAULT_TOKENIZE = 'porter unicode61';
@@ -51,7 +49,7 @@ async function createContentTable(conn, tokenize) {
 }
 // Content is a separate table (not a column on frontmatter) so `SELECT * FROM frontmatter`
 // can't dump file text into context. Features add their own tables after the core ones.
-async function ensureSchema(conn, cfg, tokenize) {
+async function ensureSchemaTables(conn, cfg, tokenize) {
     await conn.exec(`CREATE TABLE IF NOT EXISTS frontmatter ("path" TEXT PRIMARY KEY, "_mtime" REAL, "_ctime" REAL, "_size" INTEGER, "_parse_error" TEXT)`);
     // IF NOT EXISTS is safe against a tokenizer change: open() compares the table's own DDL
     // against the resolved tokenizer before this runs, so a stale table is already gone by now.
@@ -78,18 +76,9 @@ function setJournalWal(db) {
         }
     }
 }
-async function connect(cfg) {
-    const stateDir = join(cfg.baseDir, STATE_DIR);
-    mkdirSync(stateDir, {
-        recursive: true
-    });
-    const dbPath = join(stateDir, DB_FILENAME);
+async function connect(dbPath, cfg) {
     const db = new DatabaseSync(dbPath);
-    // A throw below must release this handle, or on Windows the leaked WAL db is undeletable and
-    // scratch cleanup fails with EPERM/EBUSY. `closed` keeps the catch from double-closing.
-    let closed = false;
     try {
-        var _ref;
         // Before journal_mode, not after: converting a fresh database to WAL takes a brief exclusive
         // lock, and with no timeout set yet a second process opening the same tree fails in 1ms.
         db.exec('PRAGMA busy_timeout = 30000');
@@ -98,107 +87,88 @@ async function connect(cfg) {
         setJournalWal(db);
         registerFunctions(db, contentTokenize(cfg) === undefined);
         const conn = createConnection(db);
-        db.exec('CREATE TABLE IF NOT EXISTS meta (key TEXT PRIMARY KEY, value TEXT)');
         // Before the rebuild branch below, never after: that branch deletes the cache, so a typo'd
         // tokenizer validated later would cost a full re-index (and re-embed) to reach its own error.
         const tokenize = resolveTokenize(db, cfg);
-        // Schema-version or feature-set mismatch: reconcile only reparses changed files, so an
-        // old cache can't be patched incrementally -- rebuild instead (cheap: nothing expensive lives here).
-        const version = await getMeta(conn, 'schema_version');
-        const features = await getMeta(conn, 'features');
-        const wantFeatures = featureSignature(cfg, FEATURES);
-        let tokenizeOnlyRebuild = false;
-        if (version !== null && version !== SCHEMA_VERSION || features !== null && features !== wantFeatures) {
-            // Indexing derives from presets, so a config edit rebuilding the cache must say so and
-            // name what changed -- silent rebuilds make derived indexing look like a hang or a bug.
-            if (version !== null && version !== SCHEMA_VERSION) {
-                console.error('sense: cache format changed (new sensemaking version); rebuilding the index');
-                closed = true;
-                db.close();
-                clearCache(cfg);
-                return connect(cfg);
-            }
-            const changedKeys = changedSignatureKeys(features !== null && features !== void 0 ? features : '', wantFeatures);
-            // Only the tokenizer moved, and frontmatter, links, sections and embeddings are file-derived
-            // and tokenizer-independent. Any other signature change takes the full clear/reopen below.
-            if (changedKeys.size === 1 && changedKeys.has('tokenize')) {
-                console.error('sense: config change (content tokenizer) rebuilds the text index; vectors, links, and sections are kept');
-                // One transaction, so a crash before COMMIT rolls back to the old tokenizer's table rather
-                // than to no table. IF EXISTS makes a retry after such a crash a no-op, not a raw error.
-                await withTransaction(conn, async ()=>{
-                    await conn.exec('DROP TABLE IF EXISTS content');
-                    await createContentTable(conn, tokenize);
-                }, BEGIN_WRITE);
-                tokenizeOnlyRebuild = true;
-            } else if (changedKeys.size === 1 && changedKeys.has('embed') && embedIdentityAdopted(features !== null && features !== void 0 ? features : '', wantFeatures)) {
-                // First sight of a resolved weight identity: the model itself hasn't changed, so
-                // adopt it into meta with no rebuild and no re-embed, mirroring the tokenize precedent.
-                console.error("sense: recorded the embedding model's resolved identity; vectors are unaffected");
-                await setMeta(conn, 'features', wantFeatures);
-            } else {
-                const changed = signatureDiff(features !== null && features !== void 0 ? features : '', wantFeatures);
-                console.error(`sense: config change (${changed}) rebuilds the index`);
-                closed = true;
-                db.close();
-                clearCache(cfg);
-                return connect(cfg);
-            }
-        }
-        // Meta can lie after a crash between table creation and the signature write; the table's own
-        // DDL cannot, so a mismatch here rebuilds whatever meta says.
-        const stored = storedTokenize(db);
-        if (stored !== null && stored !== tokenize) {
-            console.error('sense: cache was built with a different content tokenizer; rebuilding the index');
-            closed = true;
-            db.close();
-            clearCache(cfg);
-            return connect(cfg);
-        }
-        // One writer at a time: the feature hooks check a column then add it, so two cold opens racing
-        // here both see it missing and the second ALTER fails with a duplicate column.
-        await withTransaction(conn, ()=>ensureSchema(conn, cfg, tokenize), BEGIN_WRITE);
-        let rebuildWarnings = [];
-        if (tokenizeOnlyRebuild) {
-            rebuildWarnings = await rebuildContentTable(conn, cfg, cfg.baseDir);
-            await setMeta(conn, 'features', wantFeatures);
-        }
-        // 3x the largest reconcile this cache has recorded, floored at 30s and capped at 10min.
-        // Installed before reconcile() -- that call is the one that races a watcher's transaction.
-        const recordedMaxMs = Number((_ref = await getMeta(conn, 'reconcile_max_ms')) !== null && _ref !== void 0 ? _ref : '0');
-        db.exec(`PRAGMA busy_timeout = ${Math.min(Math.max(30000, 3 * recordedMaxMs), 600000)}`);
-        const { parsed, warnings } = await reconcile(conn, cfg, cfg.baseDir);
         return {
-            db,
-            conn,
-            cfg,
-            dbPath,
-            parsed,
-            warnings: [
-                ...rebuildWarnings,
-                ...warnings
-            ]
+            handle: {
+                db,
+                tokenize
+            },
+            conn
         };
     } catch (err) {
-        if (!closed) db.close();
+        // A throw here must release this handle, or on Windows the leaked WAL db is undeletable and
+        // scratch cleanup fails with EPERM/EBUSY.
+        db.close();
         throw err;
     }
 }
-// The sqlite store's open: connects synchronously (see connect() above), then wraps the
-// resulting connection in the async Store interface.
-export async function openSqlite(cfg) {
-    const { db, conn, cfg: resolvedCfg, dbPath, parsed, warnings } = await connect(cfg);
-    const store = createStore(db, conn, resolvedCfg, resolvedCfg.baseDir);
-    // reconcile ran before this object existed, so its chunk text is keyed by the connection.
-    rekeyChunkText(conn, store);
-    return {
-        store: store,
-        cfg: resolvedCfg,
-        dbPath,
-        parsed,
-        warnings
-    };
+async function close(handle) {
+    handle.db.close();
+}
+// Only the tokenizer moved, and frontmatter, links, sections and embeddings are file-derived
+// and tokenizer-independent. Any other signature change takes the full clear/reopen in open.ts.
+async function partialRebuild(handle, conn, _cfg, changedKeys) {
+    if (!(changedKeys.size === 1 && changedKeys.has('tokenize'))) return false;
+    console.error('sense: config change (content tokenizer) rebuilds the text index; vectors, links, and sections are kept');
+    // One transaction, so a crash before COMMIT rolls back to the old tokenizer's table rather
+    // than to no table. IF EXISTS makes a retry after such a crash a no-op, not a raw error.
+    await withTransaction(conn, async ()=>{
+        await conn.exec('DROP TABLE IF EXISTS content');
+        await createContentTable(conn, handle.tokenize);
+    }, BEGIN_WRITE);
+    return true;
 }
-export async function docCount(store) {
-    const stmt = await store.prepare('SELECT COUNT(*) AS n FROM frontmatter');
-    return (await stmt.get()).n;
+// Tokenize-only rebuild: only `content` is repopulated (the other tables are tokenizer-independent)
+// and no feature extractors run. Warnings come from the reparse, since mtimes stay untouched and reconcile would never reparse them.
+async function postSchemaRebuild(_handle, conn, cfg, baseDir, wantFeatures) {
+    const stmt = await conn.prepare('SELECT "path" FROM frontmatter');
+    const known = new Set((await stmt.all()).map((r)=>r.path));
+    const files = listFiles(cfg, baseDir).filter((f)=>known.has(f.relPath));
+    const segmenting = contentTokenize(cfg) === undefined;
+    const warnings = [];
+    const rows = [];
+    for (const file of files){
+        const { doc, warnings: fileWarnings } = parseFile(file);
+        warnings.push(...fileWarnings);
+        rows.push(contentRow(doc, segmenting));
+    }
+    await withTransaction(conn, async ()=>{
+        if (rows.length > 0) await conn.runBatch(INSERT_CONTENT_SQL, rows);
+    }, BEGIN_WRITE);
+    await setMeta(conn, 'features', wantFeatures);
+    return warnings;
+}
+// Meta can lie after a crash between table creation and the signature write; the table's own
+// DDL cannot, so a mismatch here rebuilds whatever meta says.
+function extraRebuildReason(handle) {
+    const stored = storedTokenize(handle.db);
+    if (stored !== null && stored !== handle.tokenize) return 'cache was built with a different content tokenizer';
+    return null;
+}
+async function ensureSchema(handle, conn, cfg) {
+    // One writer at a time: the feature hooks check a column then add it, so two cold opens racing
+    // here both see it missing and the second ALTER fails with a duplicate column.
+    await withTransaction(conn, ()=>ensureSchemaTables(conn, cfg, handle.tokenize), BEGIN_WRITE);
+}
+async function setDerivedBusyTimeout(handle, _conn, ms) {
+    handle.db.exec(`PRAGMA busy_timeout = ${ms}`);
+}
+// This store's dialect (types.ts's OpenDialect) for the shared orchestration in store/open.ts.
+export const sqliteOpenDialect = {
+    filename: DB_FILENAME,
+    schemaVersion: SCHEMA_VERSION,
+    reconcileDialect: sqliteDialect,
+    connect,
+    close,
+    partialRebuild,
+    postSchemaRebuild,
+    extraRebuildReason,
+    ensureSchema,
+    setDerivedBusyTimeout,
+    createStore: (handle, conn, cfg)=>createStore(handle.db, conn, cfg)
+};
+export async function openSqlite(cfg) {
+    return openWithDialect(cfg, sqliteOpenDialect);
 }

package/dist/esm/store/sqlite/open.js.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/store/sqlite/open.ts"],"sourcesContent":["// The Node floor (>=22.20) is explained here and nowhere else: 22.20 is the first release with\n// both FTS5 and row-returning INSERT ... RETURNING. Raise it only for a load-bearing capability.\nimport { mkdirSync } from 'node:fs';\nimport { join } from 'node:path';\nimport { DatabaseSync } from 'node:sqlite';\nimport type { Config, ResolvedConfig } from '../../config/index.ts';\nimport { contentTokenize, featureSignature, STATE_DIR } from '../../config/index.ts';\nimport { rekeyChunkText } from '../../embed/handoff.ts';\nimport { SenseError } from '../../errors.ts';\nimport { activeFeatures, FEATURES } from '../../features/index.ts';\nimport { clearCache } from '../cache.ts';\nimport { getMeta, setMeta } from '../shared.ts';\nimport { BEGIN_WRITE, withTransaction } from '../transaction.ts';\nimport type { Connection, Store } from '../types.ts';\nimport { createConnection } from './connection.ts';\nimport { changedSignatureKeys, embedIdentityAdopted, rebuildContentTable, reconcile, signatureDiff } from './reconcile.ts';\nimport { registerFunctions } from './sql-functions.ts';\nimport { createStore } from './store.ts';\n\nexport const DB_FILENAME = 'cache.db';\n// Cache shape version, independent of the config's own `version`. Bumping it rebuilds\n// existing trees on first query.\nexport const SCHEMA_VERSION = '18';\n\ninterface ConnectResult {\n  db: DatabaseSync;\n  conn: Connection;\n  cfg: ResolvedConfig;\n  dbPath: string;\n  parsed: number;\n  warnings: string[];\n}\n\nexport interface OpenResult {\n  store: Store;\n  cfg: ResolvedConfig;\n  dbPath: string;\n  parsed: number;\n  warnings: string[];\n}\n\n// unicode61 splits on spaces, so a language written without them indexes a whole run as one\n// token and word search finds nothing; `content.tokenize` is how such a tree picks trigram.\nconst DEFAULT_TOKENIZE = 'porter unicode61';\n\n// FTS5 takes its tokenizer as a DDL string literal where nothing can bind, so the configured\n// value is concatenated; probing a throwaway table first is both the safety and the validation.\nfunction resolveTokenize(db: DatabaseSync, cfg: Config): string {\n  const configured = contentTokenize(cfg);\n  if (configured === undefined) return DEFAULT_TOKENIZE;\n  const literal = configured.replace(/'/g, \"''\");\n  try {\n    db.exec('DROP TABLE IF EXISTS temp.sense_tokenize_probe');\n    db.exec(`CREATE VIRTUAL TABLE temp.sense_tokenize_probe USING fts5(x, tokenize = '${literal}')`);\n    db.exec('DROP TABLE IF EXISTS temp.sense_tokenize_probe');\n  } catch (err) {\n    throw new SenseError('CONFIG_INVALID', `content.tokenize \"${configured}\" is not a tokenizer this SQLite accepts (${(err as Error).message}); the built-in choices are unicode61, ascii, porter, and trigram, each with their own options`);\n  }\n  return literal;\n}\n\n// The tokenizer the content table was actually built with, from its own DDL -- the one\n// record that cannot desynchronize from the table. NULL when the table does not exist yet.\nfunction storedTokenize(db: DatabaseSync): string | null {\n  const row = db.prepare(`SELECT sql FROM sqlite_master WHERE name = 'content'`).get() as { sql: string } | undefined;\n  if (!row) return null;\n  const m = row.sql.match(/tokenize = '((?:[^']|'')*)'/);\n  return m ? m[1] : null;\n}\n\n// The `_seg` sidecars are appended after path, never inserted: bm25() and snippet(content, 2)\n// are documented against the first three columns. Each holds its field's exploded unspaced runs.\nasync function createContentTable(conn: Connection, tokenize: string): Promise<void> {\n  await conn.exec(`CREATE VIRTUAL TABLE IF NOT EXISTS content USING fts5(title, summary, text, path UNINDEXED, title_seg, summary_seg, text_seg, tokenize = '${tokenize}')`);\n}\n\n// Content is a separate table (not a column on frontmatter) so `SELECT * FROM frontmatter`\n// can't dump file text into context. Features add their own tables after the core ones.\nasync function ensureSchema(conn: Connection, cfg: Config, tokenize: string): Promise<void> {\n  await conn.exec(`CREATE TABLE IF NOT EXISTS frontmatter (\"path\" TEXT PRIMARY KEY, \"_mtime\" REAL, \"_ctime\" REAL, \"_size\" INTEGER, \"_parse_error\" TEXT)`);\n  // IF NOT EXISTS is safe against a tokenizer change: open() compares the table's own DDL\n  // against the resolved tokenizer before this runs, so a stale table is already gone by now.\n  await createContentTable(conn, tokenize);\n  // Coverage, not ownership: a path can appear under several presets. path leads the PK so the\n  // per-doc delete is an index hit -- keyed the other way, cold builds went quadratic.\n  await conn.exec(`CREATE TABLE IF NOT EXISTS preset_files (\"path\" TEXT, preset TEXT, PRIMARY KEY (\"path\", preset))`);\n  await conn.exec('CREATE INDEX IF NOT EXISTS preset_files_preset ON preset_files(preset)');\n  for (const feature of activeFeatures(cfg)) await feature.schema(conn);\n  if ((await getMeta(conn, 'schema_version')) === null) await setMeta(conn, 'schema_version', SCHEMA_VERSION);\n  if ((await getMeta(conn, 'features')) === null) await setMeta(conn, 'features', featureSignature(cfg, FEATURES));\n}\n\n// Two processes opening the same fresh tree both try to convert it, and the loser gets SQLITE_BUSY\n// with no busy handler behind it. Bounded because a lock held past this is a real problem, not a race.\nfunction setJournalWal(db: DatabaseSync): void {\n  const deadline = Date.now() + 5000;\n  for (;;) {\n    try {\n      db.exec('PRAGMA journal_mode = WAL');\n      return;\n    } catch (err) {\n      if (Date.now() >= deadline || !/database is locked|busy/i.test((err as Error).message)) throw err;\n      Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, 20);\n    }\n  }\n}\n\nasync function connect(cfg: ResolvedConfig): Promise<ConnectResult> {\n  const stateDir = join(cfg.baseDir, STATE_DIR);\n  mkdirSync(stateDir, { recursive: true });\n  const dbPath = join(stateDir, DB_FILENAME);\n\n  const db = new DatabaseSync(dbPath);\n  // A throw below must release this handle, or on Windows the leaked WAL db is undeletable and\n  // scratch cleanup fails with EPERM/EBUSY. `closed` keeps the catch from double-closing.\n  let closed = false;\n  try {\n    // Before journal_mode, not after: converting a fresh database to WAL takes a brief exclusive\n    // lock, and with no timeout set yet a second process opening the same tree fails in 1ms.\n    db.exec('PRAGMA busy_timeout = 30000');\n    // busy_timeout does not cover the WAL conversion itself: SQLite does not invoke the busy\n    // handler for it, so a concurrent cold open needs its own bounded wait.\n    setJournalWal(db);\n    registerFunctions(db, contentTokenize(cfg) === undefined);\n    const conn = createConnection(db);\n\n    db.exec('CREATE TABLE IF NOT EXISTS meta (key TEXT PRIMARY KEY, value TEXT)');\n\n    // Before the rebuild branch below, never after: that branch deletes the cache, so a typo'd\n    // tokenizer validated later would cost a full re-index (and re-embed) to reach its own error.\n    const tokenize = resolveTokenize(db, cfg);\n\n    // Schema-version or feature-set mismatch: reconcile only reparses changed files, so an\n    // old cache can't be patched incrementally -- rebuild instead (cheap: nothing expensive lives here).\n    const version = await getMeta(conn, 'schema_version');\n    const features = await getMeta(conn, 'features');\n    const wantFeatures = featureSignature(cfg, FEATURES);\n    let tokenizeOnlyRebuild = false;\n    if ((version !== null && version !== SCHEMA_VERSION) || (features !== null && features !== wantFeatures)) {\n      // Indexing derives from presets, so a config edit rebuilding the cache must say so and\n      // name what changed -- silent rebuilds make derived indexing look like a hang or a bug.\n      if (version !== null && version !== SCHEMA_VERSION) {\n        console.error('sense: cache format changed (new sensemaking version); rebuilding the index');\n        closed = true;\n        db.close();\n        clearCache(cfg);\n        return connect(cfg);\n      }\n      const changedKeys = changedSignatureKeys(features ?? '', wantFeatures);\n      // Only the tokenizer moved, and frontmatter, links, sections and embeddings are file-derived\n      // and tokenizer-independent. Any other signature change takes the full clear/reopen below.\n      if (changedKeys.size === 1 && changedKeys.has('tokenize')) {\n        console.error('sense: config change (content tokenizer) rebuilds the text index; vectors, links, and sections are kept');\n        // One transaction, so a crash before COMMIT rolls back to the old tokenizer's table rather\n        // than to no table. IF EXISTS makes a retry after such a crash a no-op, not a raw error.\n        await withTransaction(\n          conn,\n          async () => {\n            await conn.exec('DROP TABLE IF EXISTS content');\n            await createContentTable(conn, tokenize);\n          },\n          BEGIN_WRITE\n        );\n        tokenizeOnlyRebuild = true;\n      } else if (changedKeys.size === 1 && changedKeys.has('embed') && embedIdentityAdopted(features ?? '', wantFeatures)) {\n        // First sight of a resolved weight identity: the model itself hasn't changed, so\n        // adopt it into meta with no rebuild and no re-embed, mirroring the tokenize precedent.\n        console.error(\"sense: recorded the embedding model's resolved identity; vectors are unaffected\");\n        await setMeta(conn, 'features', wantFeatures);\n      } else {\n        const changed = signatureDiff(features ?? '', wantFeatures);\n        console.error(`sense: config change (${changed}) rebuilds the index`);\n        closed = true;\n        db.close();\n        clearCache(cfg);\n        return connect(cfg);\n      }\n    }\n\n    // Meta can lie after a crash between table creation and the signature write; the table's own\n    // DDL cannot, so a mismatch here rebuilds whatever meta says.\n    const stored = storedTokenize(db);\n    if (stored !== null && stored !== tokenize) {\n      console.error('sense: cache was built with a different content tokenizer; rebuilding the index');\n      closed = true;\n      db.close();\n      clearCache(cfg);\n      return connect(cfg);\n    }\n\n    // One writer at a time: the feature hooks check a column then add it, so two cold opens racing\n    // here both see it missing and the second ALTER fails with a duplicate column.\n    await withTransaction(conn, () => ensureSchema(conn, cfg, tokenize), BEGIN_WRITE);\n\n    let rebuildWarnings: string[] = [];\n    if (tokenizeOnlyRebuild) {\n      rebuildWarnings = await rebuildContentTable(conn, cfg, cfg.baseDir);\n      await setMeta(conn, 'features', wantFeatures);\n    }\n\n    // 3x the largest reconcile this cache has recorded, floored at 30s and capped at 10min.\n    // Installed before reconcile() -- that call is the one that races a watcher's transaction.\n    const recordedMaxMs = Number((await getMeta(conn, 'reconcile_max_ms')) ?? '0');\n    db.exec(`PRAGMA busy_timeout = ${Math.min(Math.max(30000, 3 * recordedMaxMs), 600_000)}`);\n\n    const { parsed, warnings } = await reconcile(conn, cfg, cfg.baseDir);\n\n    return { db, conn, cfg, dbPath, parsed, warnings: [...rebuildWarnings, ...warnings] };\n  } catch (err) {\n    if (!closed) db.close();\n    throw err;\n  }\n}\n\n// The sqlite store's open: connects synchronously (see connect() above), then wraps the\n// resulting connection in the async Store interface.\nexport async function openSqlite(cfg: ResolvedConfig): Promise<OpenResult> {\n  const { db, conn, cfg: resolvedCfg, dbPath, parsed, warnings } = await connect(cfg);\n  const store = createStore(db, conn, resolvedCfg, resolvedCfg.baseDir);\n  // reconcile ran before this object existed, so its chunk text is keyed by the connection.\n  rekeyChunkText(conn, store);\n  return { store: store, cfg: resolvedCfg, dbPath, parsed, warnings };\n}\n\nexport async function docCount(store: Store): Promise<number> {\n  const stmt = await store.prepare('SELECT COUNT(*) AS n FROM frontmatter');\n  return ((await stmt.get()) as { n: number }).n;\n}\n"],"names":["mkdirSync","join","DatabaseSync","contentTokenize","featureSignature","STATE_DIR","rekeyChunkText","SenseError","activeFeatures","FEATURES","clearCache","getMeta","setMeta","BEGIN_WRITE","withTransaction","createConnection","changedSignatureKeys","embedIdentityAdopted","rebuildContentTable","reconcile","signatureDiff","registerFunctions","createStore","DB_FILENAME","SCHEMA_VERSION","DEFAULT_TOKENIZE","resolveTokenize","db","cfg","configured","undefined","literal","replace","exec","err","message","storedTokenize","row","prepare","get","m","sql","match","createContentTable","conn","tokenize","ensureSchema","feature","schema","setJournalWal","deadline","Date","now","test","Atomics","wait","Int32Array","SharedArrayBuffer","connect","stateDir","baseDir","recursive","dbPath","closed","version","features","wantFeatures","tokenizeOnlyRebuild","console","error","close","changedKeys","size","has","changed","stored","rebuildWarnings","recordedMaxMs","Number","Math","min","max","parsed","warnings","openSqlite","resolvedCfg","store","docCount","stmt","n"],"mappings":"AAAA,+FAA+F;AAC/F,iGAAiG;AACjG,SAASA,SAAS,QAAQ,UAAU;AACpC,SAASC,IAAI,QAAQ,YAAY;AACjC,SAASC,YAAY,QAAQ,cAAc;AAE3C,SAASC,eAAe,EAAEC,gBAAgB,EAAEC,SAAS,QAAQ,wBAAwB;AACrF,SAASC,cAAc,QAAQ,yBAAyB;AACxD,SAASC,UAAU,QAAQ,kBAAkB;AAC7C,SAASC,cAAc,EAAEC,QAAQ,QAAQ,0BAA0B;AACnE,SAASC,UAAU,QAAQ,cAAc;AACzC,SAASC,OAAO,EAAEC,OAAO,QAAQ,eAAe;AAChD,SAASC,WAAW,EAAEC,eAAe,QAAQ,oBAAoB;AAEjE,SAASC,gBAAgB,QAAQ,kBAAkB;AACnD,SAASC,oBAAoB,EAAEC,oBAAoB,EAAEC,mBAAmB,EAAEC,SAAS,EAAEC,aAAa,QAAQ,iBAAiB;AAC3H,SAASC,iBAAiB,QAAQ,qBAAqB;AACvD,SAASC,WAAW,QAAQ,aAAa;AAEzC,OAAO,MAAMC,cAAc,WAAW;AACtC,sFAAsF;AACtF,iCAAiC;AACjC,OAAO,MAAMC,iBAAiB,KAAK;AAmBnC,4FAA4F;AAC5F,4FAA4F;AAC5F,MAAMC,mBAAmB;AAEzB,6FAA6F;AAC7F,gGAAgG;AAChG,SAASC,gBAAgBC,EAAgB,EAAEC,GAAW;IACpD,MAAMC,aAAa1B,gBAAgByB;IACnC,IAAIC,eAAeC,WAAW,OAAOL;IACrC,MAAMM,UAAUF,WAAWG,OAAO,CAAC,MAAM;IACzC,IAAI;QACFL,GAAGM,IAAI,CAAC;QACRN,GAAGM,IAAI,CAAC,CAAC,yEAAyE,EAAEF,QAAQ,EAAE,CAAC;QAC/FJ,GAAGM,IAAI,CAAC;IACV,EAAE,OAAOC,KAAK;QACZ,MAAM,IAAI3B,WAAW,kBAAkB,CAAC,kBAAkB,EAAEsB,WAAW,0CAA0C,EAAE,AAACK,IAAcC,OAAO,CAAC,8FAA8F,CAAC;IAC3O;IACA,OAAOJ;AACT;AAEA,uFAAuF;AACvF,2FAA2F;AAC3F,SAASK,eAAeT,EAAgB;IACtC,MAAMU,MAAMV,GAAGW,OAAO,CAAC,CAAC,oDAAoD,CAAC,EAAEC,GAAG;IAClF,IAAI,CAACF,KAAK,OAAO;IACjB,MAAMG,IAAIH,IAAII,GAAG,CAACC,KAAK,CAAC;IACxB,OAAOF,IAAIA,CAAC,CAAC,EAAE,GAAG;AACpB;AAEA,8FAA8F;AAC9F,iGAAiG;AACjG,eAAeG,mBAAmBC,IAAgB,EAAEC,QAAgB;IAClE,MAAMD,KAAKX,IAAI,CAAC,CAAC,0IAA0I,EAAEY,SAAS,EAAE,CAAC;AAC3K;AAEA,2FAA2F;AAC3F,wFAAwF;AACxF,eAAeC,aAAaF,IAAgB,EAAEhB,GAAW,EAAEiB,QAAgB;IACzE,MAAMD,KAAKX,IAAI,CAAC,CAAC,oIAAoI,CAAC;IACtJ,wFAAwF;IACxF,4FAA4F;IAC5F,MAAMU,mBAAmBC,MAAMC;IAC/B,6FAA6F;IAC7F,qFAAqF;IACrF,MAAMD,KAAKX,IAAI,CAAC,CAAC,gGAAgG,CAAC;IAClH,MAAMW,KAAKX,IAAI,CAAC;IAChB,KAAK,MAAMc,WAAWvC,eAAeoB,KAAM,MAAMmB,QAAQC,MAAM,CAACJ;IAChE,IAAI,AAAC,MAAMjC,QAAQiC,MAAM,sBAAuB,MAAM,MAAMhC,QAAQgC,MAAM,kBAAkBpB;IAC5F,IAAI,AAAC,MAAMb,QAAQiC,MAAM,gBAAiB,MAAM,MAAMhC,QAAQgC,MAAM,YAAYxC,iBAAiBwB,KAAKnB;AACxG;AAEA,mGAAmG;AACnG,uGAAuG;AACvG,SAASwC,cAActB,EAAgB;IACrC,MAAMuB,WAAWC,KAAKC,GAAG,KAAK;IAC9B,OAAS;QACP,IAAI;YACFzB,GAAGM,IAAI,CAAC;YACR;QACF,EAAE,OAAOC,KAAK;YACZ,IAAIiB,KAAKC,GAAG,MAAMF,YAAY,CAAC,2BAA2BG,IAAI,CAAC,AAACnB,IAAcC,OAAO,GAAG,MAAMD;YAC9FoB,QAAQC,IAAI,CAAC,IAAIC,WAAW,IAAIC,kBAAkB,KAAK,GAAG,GAAG;QAC/D;IACF;AACF;AAEA,eAAeC,QAAQ9B,GAAmB;IACxC,MAAM+B,WAAW1D,KAAK2B,IAAIgC,OAAO,EAAEvD;IACnCL,UAAU2D,UAAU;QAAEE,WAAW;IAAK;IACtC,MAAMC,SAAS7D,KAAK0D,UAAUpC;IAE9B,MAAMI,KAAK,IAAIzB,aAAa4D;IAC5B,6FAA6F;IAC7F,wFAAwF;IACxF,IAAIC,SAAS;IACb,IAAI;YAsF4B;QArF9B,6FAA6F;QAC7F,yFAAyF;QACzFpC,GAAGM,IAAI,CAAC;QACR,yFAAyF;QACzF,wEAAwE;QACxEgB,cAActB;QACdN,kBAAkBM,IAAIxB,gBAAgByB,SAASE;QAC/C,MAAMc,OAAO7B,iBAAiBY;QAE9BA,GAAGM,IAAI,CAAC;QAER,2FAA2F;QAC3F,8FAA8F;QAC9F,MAAMY,WAAWnB,gBAAgBC,IAAIC;QAErC,uFAAuF;QACvF,qGAAqG;QACrG,MAAMoC,UAAU,MAAMrD,QAAQiC,MAAM;QACpC,MAAMqB,WAAW,MAAMtD,QAAQiC,MAAM;QACrC,MAAMsB,eAAe9D,iBAAiBwB,KAAKnB;QAC3C,IAAI0D,sBAAsB;QAC1B,IAAI,AAACH,YAAY,QAAQA,YAAYxC,kBAAoByC,aAAa,QAAQA,aAAaC,cAAe;YACxG,uFAAuF;YACvF,wFAAwF;YACxF,IAAIF,YAAY,QAAQA,YAAYxC,gBAAgB;gBAClD4C,QAAQC,KAAK,CAAC;gBACdN,SAAS;gBACTpC,GAAG2C,KAAK;gBACR5D,WAAWkB;gBACX,OAAO8B,QAAQ9B;YACjB;YACA,MAAM2C,cAAcvD,qBAAqBiD,qBAAAA,sBAAAA,WAAY,IAAIC;YACzD,6FAA6F;YAC7F,2FAA2F;YAC3F,IAAIK,YAAYC,IAAI,KAAK,KAAKD,YAAYE,GAAG,CAAC,aAAa;gBACzDL,QAAQC,KAAK,CAAC;gBACd,2FAA2F;gBAC3F,yFAAyF;gBACzF,MAAMvD,gBACJ8B,MACA;oBACE,MAAMA,KAAKX,IAAI,CAAC;oBAChB,MAAMU,mBAAmBC,MAAMC;gBACjC,GACAhC;gBAEFsD,sBAAsB;YACxB,OAAO,IAAII,YAAYC,IAAI,KAAK,KAAKD,YAAYE,GAAG,CAAC,YAAYxD,qBAAqBgD,qBAAAA,sBAAAA,WAAY,IAAIC,eAAe;gBACnH,iFAAiF;gBACjF,wFAAwF;gBACxFE,QAAQC,KAAK,CAAC;gBACd,MAAMzD,QAAQgC,MAAM,YAAYsB;YAClC,OAAO;gBACL,MAAMQ,UAAUtD,cAAc6C,qBAAAA,sBAAAA,WAAY,IAAIC;gBAC9CE,QAAQC,KAAK,CAAC,CAAC,sBAAsB,EAAEK,QAAQ,oBAAoB,CAAC;gBACpEX,SAAS;gBACTpC,GAAG2C,KAAK;gBACR5D,WAAWkB;gBACX,OAAO8B,QAAQ9B;YACjB;QACF;QAEA,6FAA6F;QAC7F,8DAA8D;QAC9D,MAAM+C,SAASvC,eAAeT;QAC9B,IAAIgD,WAAW,QAAQA,WAAW9B,UAAU;YAC1CuB,QAAQC,KAAK,CAAC;YACdN,SAAS;YACTpC,GAAG2C,KAAK;YACR5D,WAAWkB;YACX,OAAO8B,QAAQ9B;QACjB;QAEA,+FAA+F;QAC/F,+EAA+E;QAC/E,MAAMd,gBAAgB8B,MAAM,IAAME,aAAaF,MAAMhB,KAAKiB,WAAWhC;QAErE,IAAI+D,kBAA4B,EAAE;QAClC,IAAIT,qBAAqB;YACvBS,kBAAkB,MAAM1D,oBAAoB0B,MAAMhB,KAAKA,IAAIgC,OAAO;YAClE,MAAMhD,QAAQgC,MAAM,YAAYsB;QAClC;QAEA,wFAAwF;QACxF,2FAA2F;QAC3F,MAAMW,gBAAgBC,QAAQ,OAAA,MAAMnE,QAAQiC,MAAM,iCAApB,kBAAA,OAA4C;QAC1EjB,GAAGM,IAAI,CAAC,CAAC,sBAAsB,EAAE8C,KAAKC,GAAG,CAACD,KAAKE,GAAG,CAAC,OAAO,IAAIJ,gBAAgB,SAAU;QAExF,MAAM,EAAEK,MAAM,EAAEC,QAAQ,EAAE,GAAG,MAAMhE,UAAUyB,MAAMhB,KAAKA,IAAIgC,OAAO;QAEnE,OAAO;YAAEjC;YAAIiB;YAAMhB;YAAKkC;YAAQoB;YAAQC,UAAU;mBAAIP;mBAAoBO;aAAS;QAAC;IACtF,EAAE,OAAOjD,KAAK;QACZ,IAAI,CAAC6B,QAAQpC,GAAG2C,KAAK;QACrB,MAAMpC;IACR;AACF;AAEA,wFAAwF;AACxF,qDAAqD;AACrD,OAAO,eAAekD,WAAWxD,GAAmB;IAClD,MAAM,EAAED,EAAE,EAAEiB,IAAI,EAAEhB,KAAKyD,WAAW,EAAEvB,MAAM,EAAEoB,MAAM,EAAEC,QAAQ,EAAE,GAAG,MAAMzB,QAAQ9B;IAC/E,MAAM0D,QAAQhE,YAAYK,IAAIiB,MAAMyC,aAAaA,YAAYzB,OAAO;IACpE,0FAA0F;IAC1FtD,eAAesC,MAAM0C;IACrB,OAAO;QAAEA,OAAOA;QAAO1D,KAAKyD;QAAavB;QAAQoB;QAAQC;IAAS;AACpE;AAEA,OAAO,eAAeI,SAASD,KAAY;IACzC,MAAME,OAAO,MAAMF,MAAMhD,OAAO,CAAC;IACjC,OAAO,AAAE,CAAA,MAAMkD,KAAKjD,GAAG,EAAC,EAAqBkD,CAAC;AAChD"}
+{"version":3,"sources":["/Users/kevin/Dev/OpenSource/ai/sensemaking/src/store/sqlite/open.ts"],"sourcesContent":["// The Node floor (>=22.20) is explained here and nowhere else: 22.20 is the first release with\n// both FTS5 and row-returning INSERT ... RETURNING. Raise it only for a load-bearing capability.\nimport { DatabaseSync } from 'node:sqlite';\nimport type { Config, ResolvedConfig } from '../../config/index.ts';\nimport { contentTokenize, featureSignature } from '../../config/index.ts';\nimport { SenseError } from '../../errors.ts';\nimport { activeFeatures, FEATURES } from '../../features/index.ts';\nimport { listFiles, parseFile } from '../../scan/index.ts';\nimport type { OpenResult } from '../open.ts';\nimport { openWithDialect } from '../open.ts';\nimport { getMeta, setMeta } from '../shared.ts';\nimport { BEGIN_WRITE, withTransaction } from '../transaction.ts';\nimport type { Connection, OpenDialect } from '../types.ts';\nimport { createConnection } from './connection.ts';\nimport { contentRow, INSERT_CONTENT_SQL, sqliteDialect } from './reconcile.ts';\nimport { registerFunctions } from './sql-functions.ts';\nimport { createStore } from './store.ts';\n\nexport const DB_FILENAME = 'cache.db';\n// Cache shape version, independent of the config's own `version`. Bumping it rebuilds\n// existing trees on first query.\nexport const SCHEMA_VERSION = '19';\n\nexport type { OpenResult };\n\n// unicode61 splits on spaces, so a language written without them indexes a whole run as one\n// token and word search finds nothing; `content.tokenize` is how such a tree picks trigram.\nconst DEFAULT_TOKENIZE = 'porter unicode61';\n\n// The connection plus this store's own extra state (types.ts's OpenDialect<Handle>): the\n// resolved tokenizer, threaded through connect, the tokenize-only rebuild, and ensureSchema.\ninterface SqliteHandle {\n  db: DatabaseSync;\n  tokenize: string;\n}\n\n// FTS5 takes its tokenizer as a DDL string literal where nothing can bind, so the configured\n// value is concatenated; probing a throwaway table first is both the safety and the validation.\nfunction resolveTokenize(db: DatabaseSync, cfg: Config): string {\n  const configured = contentTokenize(cfg);\n  if (configured === undefined) return DEFAULT_TOKENIZE;\n  const literal = configured.replace(/'/g, \"''\");\n  try {\n    db.exec('DROP TABLE IF EXISTS temp.sense_tokenize_probe');\n    db.exec(`CREATE VIRTUAL TABLE temp.sense_tokenize_probe USING fts5(x, tokenize = '${literal}')`);\n    db.exec('DROP TABLE IF EXISTS temp.sense_tokenize_probe');\n  } catch (err) {\n    throw new SenseError('CONFIG_INVALID', `content.tokenize \"${configured}\" is not a tokenizer this SQLite accepts (${(err as Error).message}); the built-in choices are unicode61, ascii, porter, and trigram, each with their own options`);\n  }\n  return literal;\n}\n\n// The tokenizer the content table was actually built with, from its own DDL -- the one\n// record that cannot desynchronize from the table. NULL when the table does not exist yet.\nfunction storedTokenize(db: DatabaseSync): string | null {\n  const row = db.prepare(`SELECT sql FROM sqlite_master WHERE name = 'content'`).get() as { sql: string } | undefined;\n  if (!row) return null;\n  const m = row.sql.match(/tokenize = '((?:[^']|'')*)'/);\n  return m ? m[1] : null;\n}\n\n// The `_seg` sidecars are appended after path, never inserted: bm25() and snippet(content, 2)\n// are documented against the first three columns. Each holds its field's exploded unspaced runs.\nasync function createContentTable(conn: Connection, tokenize: string): Promise<void> {\n  await conn.exec(`CREATE VIRTUAL TABLE IF NOT EXISTS content USING fts5(title, summary, text, path UNINDEXED, title_seg, summary_seg, text_seg, tokenize = '${tokenize}')`);\n}\n\n// Content is a separate table (not a column on frontmatter) so `SELECT * FROM frontmatter`\n// can't dump file text into context. Features add their own tables after the core ones.\nasync function ensureSchemaTables(conn: Connection, cfg: Config, tokenize: string): Promise<void> {\n  await conn.exec(`CREATE TABLE IF NOT EXISTS frontmatter (\"path\" TEXT PRIMARY KEY, \"_mtime\" REAL, \"_ctime\" REAL, \"_size\" INTEGER, \"_parse_error\" TEXT)`);\n  // IF NOT EXISTS is safe against a tokenizer change: open() compares the table's own DDL\n  // against the resolved tokenizer before this runs, so a stale table is already gone by now.\n  await createContentTable(conn, tokenize);\n  // Coverage, not ownership: a path can appear under several presets. path leads the PK so the\n  // per-doc delete is an index hit -- keyed the other way, cold builds went quadratic.\n  await conn.exec(`CREATE TABLE IF NOT EXISTS preset_files (\"path\" TEXT, preset TEXT, PRIMARY KEY (\"path\", preset))`);\n  await conn.exec('CREATE INDEX IF NOT EXISTS preset_files_preset ON preset_files(preset)');\n  for (const feature of activeFeatures(cfg)) await feature.schema(conn);\n  if ((await getMeta(conn, 'schema_version')) === null) await setMeta(conn, 'schema_version', SCHEMA_VERSION);\n  if ((await getMeta(conn, 'features')) === null) await setMeta(conn, 'features', featureSignature(cfg, FEATURES));\n}\n\n// Two processes opening the same fresh tree both try to convert it, and the loser gets SQLITE_BUSY\n// with no busy handler behind it. Bounded because a lock held past this is a real problem, not a race.\nfunction setJournalWal(db: DatabaseSync): void {\n  const deadline = Date.now() + 5000;\n  for (;;) {\n    try {\n      db.exec('PRAGMA journal_mode = WAL');\n      return;\n    } catch (err) {\n      if (Date.now() >= deadline || !/database is locked|busy/i.test((err as Error).message)) throw err;\n      Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, 20);\n    }\n  }\n}\n\nasync function connect(dbPath: string, cfg: ResolvedConfig): Promise<{ handle: SqliteHandle; conn: Connection }> {\n  const db = new DatabaseSync(dbPath);\n  try {\n    // Before journal_mode, not after: converting a fresh database to WAL takes a brief exclusive\n    // lock, and with no timeout set yet a second process opening the same tree fails in 1ms.\n    db.exec('PRAGMA busy_timeout = 30000');\n    // busy_timeout does not cover the WAL conversion itself: SQLite does not invoke the busy\n    // handler for it, so a concurrent cold open needs its own bounded wait.\n    setJournalWal(db);\n    registerFunctions(db, contentTokenize(cfg) === undefined);\n    const conn = createConnection(db);\n    // Before the rebuild branch below, never after: that branch deletes the cache, so a typo'd\n    // tokenizer validated later would cost a full re-index (and re-embed) to reach its own error.\n    const tokenize = resolveTokenize(db, cfg);\n    return { handle: { db, tokenize }, conn };\n  } catch (err) {\n    // A throw here must release this handle, or on Windows the leaked WAL db is undeletable and\n    // scratch cleanup fails with EPERM/EBUSY.\n    db.close();\n    throw err;\n  }\n}\n\nasync function close(handle: SqliteHandle): Promise<void> {\n  handle.db.close();\n}\n\n// Only the tokenizer moved, and frontmatter, links, sections and embeddings are file-derived\n// and tokenizer-independent. Any other signature change takes the full clear/reopen in open.ts.\nasync function partialRebuild(handle: SqliteHandle, conn: Connection, _cfg: Config, changedKeys: Set<string>): Promise<boolean> {\n  if (!(changedKeys.size === 1 && changedKeys.has('tokenize'))) return false;\n  console.error('sense: config change (content tokenizer) rebuilds the text index; vectors, links, and sections are kept');\n  // One transaction, so a crash before COMMIT rolls back to the old tokenizer's table rather\n  // than to no table. IF EXISTS makes a retry after such a crash a no-op, not a raw error.\n  await withTransaction(\n    conn,\n    async () => {\n      await conn.exec('DROP TABLE IF EXISTS content');\n      await createContentTable(conn, handle.tokenize);\n    },\n    BEGIN_WRITE\n  );\n  return true;\n}\n\n// Tokenize-only rebuild: only `content` is repopulated (the other tables are tokenizer-independent)\n// and no feature extractors run. Warnings come from the reparse, since mtimes stay untouched and reconcile would never reparse them.\nasync function postSchemaRebuild(_handle: SqliteHandle, conn: Connection, cfg: Config, baseDir: string, wantFeatures: string): Promise<string[]> {\n  const stmt = await conn.prepare('SELECT \"path\" FROM frontmatter');\n  const known = new Set(((await stmt.all()) as Array<{ path: string }>).map((r) => r.path));\n  const files = listFiles(cfg, baseDir).filter((f) => known.has(f.relPath));\n  const segmenting = contentTokenize(cfg) === undefined;\n  const warnings: string[] = [];\n  const rows: unknown[][] = [];\n  for (const file of files) {\n    const { doc, warnings: fileWarnings } = parseFile(file);\n    warnings.push(...fileWarnings);\n    rows.push(contentRow(doc, segmenting));\n  }\n  await withTransaction(\n    conn,\n    async () => {\n      if (rows.length > 0) await conn.runBatch(INSERT_CONTENT_SQL, rows);\n    },\n    BEGIN_WRITE\n  );\n  await setMeta(conn, 'features', wantFeatures);\n  return warnings;\n}\n\n// Meta can lie after a crash between table creation and the signature write; the table's own\n// DDL cannot, so a mismatch here rebuilds whatever meta says.\nfunction extraRebuildReason(handle: SqliteHandle): string | null {\n  const stored = storedTokenize(handle.db);\n  if (stored !== null && stored !== handle.tokenize) return 'cache was built with a different content tokenizer';\n  return null;\n}\n\nasync function ensureSchema(handle: SqliteHandle, conn: Connection, cfg: Config): Promise<void> {\n  // One writer at a time: the feature hooks check a column then add it, so two cold opens racing\n  // here both see it missing and the second ALTER fails with a duplicate column.\n  await withTransaction(conn, () => ensureSchemaTables(conn, cfg, handle.tokenize), BEGIN_WRITE);\n}\n\nasync function setDerivedBusyTimeout(handle: SqliteHandle, _conn: Connection, ms: number): Promise<void> {\n  handle.db.exec(`PRAGMA busy_timeout = ${ms}`);\n}\n\n// This store's dialect (types.ts's OpenDialect) for the shared orchestration in store/open.ts.\nexport const sqliteOpenDialect: OpenDialect<SqliteHandle> = {\n  filename: DB_FILENAME,\n  schemaVersion: SCHEMA_VERSION,\n  reconcileDialect: sqliteDialect,\n  connect,\n  close,\n  partialRebuild,\n  postSchemaRebuild,\n  extraRebuildReason,\n  ensureSchema,\n  setDerivedBusyTimeout,\n  createStore: (handle, conn, cfg) => createStore(handle.db, conn, cfg),\n};\n\nexport async function openSqlite(cfg: ResolvedConfig): Promise<OpenResult> {\n  return openWithDialect(cfg, sqliteOpenDialect);\n}\n"],"names":["DatabaseSync","contentTokenize","featureSignature","SenseError","activeFeatures","FEATURES","listFiles","parseFile","openWithDialect","getMeta","setMeta","BEGIN_WRITE","withTransaction","createConnection","contentRow","INSERT_CONTENT_SQL","sqliteDialect","registerFunctions","createStore","DB_FILENAME","SCHEMA_VERSION","DEFAULT_TOKENIZE","resolveTokenize","db","cfg","configured","undefined","literal","replace","exec","err","message","storedTokenize","row","prepare","get","m","sql","match","createContentTable","conn","tokenize","ensureSchemaTables","feature","schema","setJournalWal","deadline","Date","now","test","Atomics","wait","Int32Array","SharedArrayBuffer","connect","dbPath","handle","close","partialRebuild","_cfg","changedKeys","size","has","console","error","postSchemaRebuild","_handle","baseDir","wantFeatures","stmt","known","Set","all","map","r","path","files","filter","f","relPath","segmenting","warnings","rows","file","doc","fileWarnings","push","length","runBatch","extraRebuildReason","stored","ensureSchema","setDerivedBusyTimeout","_conn","ms","sqliteOpenDialect","filename","schemaVersion","reconcileDialect","openSqlite"],"mappings":"AAAA,+FAA+F;AAC/F,iGAAiG;AACjG,SAASA,YAAY,QAAQ,cAAc;AAE3C,SAASC,eAAe,EAAEC,gBAAgB,QAAQ,wBAAwB;AAC1E,SAASC,UAAU,QAAQ,kBAAkB;AAC7C,SAASC,cAAc,EAAEC,QAAQ,QAAQ,0BAA0B;AACnE,SAASC,SAAS,EAAEC,SAAS,QAAQ,sBAAsB;AAE3D,SAASC,eAAe,QAAQ,aAAa;AAC7C,SAASC,OAAO,EAAEC,OAAO,QAAQ,eAAe;AAChD,SAASC,WAAW,EAAEC,eAAe,QAAQ,oBAAoB;AAEjE,SAASC,gBAAgB,QAAQ,kBAAkB;AACnD,SAASC,UAAU,EAAEC,kBAAkB,EAAEC,aAAa,QAAQ,iBAAiB;AAC/E,SAASC,iBAAiB,QAAQ,qBAAqB;AACvD,SAASC,WAAW,QAAQ,aAAa;AAEzC,OAAO,MAAMC,cAAc,WAAW;AACtC,sFAAsF;AACtF,iCAAiC;AACjC,OAAO,MAAMC,iBAAiB,KAAK;AAInC,4FAA4F;AAC5F,4FAA4F;AAC5F,MAAMC,mBAAmB;AASzB,6FAA6F;AAC7F,gGAAgG;AAChG,SAASC,gBAAgBC,EAAgB,EAAEC,GAAW;IACpD,MAAMC,aAAaxB,gBAAgBuB;IACnC,IAAIC,eAAeC,WAAW,OAAOL;IACrC,MAAMM,UAAUF,WAAWG,OAAO,CAAC,MAAM;IACzC,IAAI;QACFL,GAAGM,IAAI,CAAC;QACRN,GAAGM,IAAI,CAAC,CAAC,yEAAyE,EAAEF,QAAQ,EAAE,CAAC;QAC/FJ,GAAGM,IAAI,CAAC;IACV,EAAE,OAAOC,KAAK;QACZ,MAAM,IAAI3B,WAAW,kBAAkB,CAAC,kBAAkB,EAAEsB,WAAW,0CAA0C,EAAE,AAACK,IAAcC,OAAO,CAAC,8FAA8F,CAAC;IAC3O;IACA,OAAOJ;AACT;AAEA,uFAAuF;AACvF,2FAA2F;AAC3F,SAASK,eAAeT,EAAgB;IACtC,MAAMU,MAAMV,GAAGW,OAAO,CAAC,CAAC,oDAAoD,CAAC,EAAEC,GAAG;IAClF,IAAI,CAACF,KAAK,OAAO;IACjB,MAAMG,IAAIH,IAAII,GAAG,CAACC,KAAK,CAAC;IACxB,OAAOF,IAAIA,CAAC,CAAC,EAAE,GAAG;AACpB;AAEA,8FAA8F;AAC9F,iGAAiG;AACjG,eAAeG,mBAAmBC,IAAgB,EAAEC,QAAgB;IAClE,MAAMD,KAAKX,IAAI,CAAC,CAAC,0IAA0I,EAAEY,SAAS,EAAE,CAAC;AAC3K;AAEA,2FAA2F;AAC3F,wFAAwF;AACxF,eAAeC,mBAAmBF,IAAgB,EAAEhB,GAAW,EAAEiB,QAAgB;IAC/E,MAAMD,KAAKX,IAAI,CAAC,CAAC,oIAAoI,CAAC;IACtJ,wFAAwF;IACxF,4FAA4F;IAC5F,MAAMU,mBAAmBC,MAAMC;IAC/B,6FAA6F;IAC7F,qFAAqF;IACrF,MAAMD,KAAKX,IAAI,CAAC,CAAC,gGAAgG,CAAC;IAClH,MAAMW,KAAKX,IAAI,CAAC;IAChB,KAAK,MAAMc,WAAWvC,eAAeoB,KAAM,MAAMmB,QAAQC,MAAM,CAACJ;IAChE,IAAI,AAAC,MAAM/B,QAAQ+B,MAAM,sBAAuB,MAAM,MAAM9B,QAAQ8B,MAAM,kBAAkBpB;IAC5F,IAAI,AAAC,MAAMX,QAAQ+B,MAAM,gBAAiB,MAAM,MAAM9B,QAAQ8B,MAAM,YAAYtC,iBAAiBsB,KAAKnB;AACxG;AAEA,mGAAmG;AACnG,uGAAuG;AACvG,SAASwC,cAActB,EAAgB;IACrC,MAAMuB,WAAWC,KAAKC,GAAG,KAAK;IAC9B,OAAS;QACP,IAAI;YACFzB,GAAGM,IAAI,CAAC;YACR;QACF,EAAE,OAAOC,KAAK;YACZ,IAAIiB,KAAKC,GAAG,MAAMF,YAAY,CAAC,2BAA2BG,IAAI,CAAC,AAACnB,IAAcC,OAAO,GAAG,MAAMD;YAC9FoB,QAAQC,IAAI,CAAC,IAAIC,WAAW,IAAIC,kBAAkB,KAAK,GAAG,GAAG;QAC/D;IACF;AACF;AAEA,eAAeC,QAAQC,MAAc,EAAE/B,GAAmB;IACxD,MAAMD,KAAK,IAAIvB,aAAauD;IAC5B,IAAI;QACF,6FAA6F;QAC7F,yFAAyF;QACzFhC,GAAGM,IAAI,CAAC;QACR,yFAAyF;QACzF,wEAAwE;QACxEgB,cAActB;QACdN,kBAAkBM,IAAItB,gBAAgBuB,SAASE;QAC/C,MAAMc,OAAO3B,iBAAiBU;QAC9B,2FAA2F;QAC3F,8FAA8F;QAC9F,MAAMkB,WAAWnB,gBAAgBC,IAAIC;QACrC,OAAO;YAAEgC,QAAQ;gBAAEjC;gBAAIkB;YAAS;YAAGD;QAAK;IAC1C,EAAE,OAAOV,KAAK;QACZ,4FAA4F;QAC5F,0CAA0C;QAC1CP,GAAGkC,KAAK;QACR,MAAM3B;IACR;AACF;AAEA,eAAe2B,MAAMD,MAAoB;IACvCA,OAAOjC,EAAE,CAACkC,KAAK;AACjB;AAEA,6FAA6F;AAC7F,gGAAgG;AAChG,eAAeC,eAAeF,MAAoB,EAAEhB,IAAgB,EAAEmB,IAAY,EAAEC,WAAwB;IAC1G,IAAI,CAAEA,CAAAA,YAAYC,IAAI,KAAK,KAAKD,YAAYE,GAAG,CAAC,WAAU,GAAI,OAAO;IACrEC,QAAQC,KAAK,CAAC;IACd,2FAA2F;IAC3F,yFAAyF;IACzF,MAAMpD,gBACJ4B,MACA;QACE,MAAMA,KAAKX,IAAI,CAAC;QAChB,MAAMU,mBAAmBC,MAAMgB,OAAOf,QAAQ;IAChD,GACA9B;IAEF,OAAO;AACT;AAEA,oGAAoG;AACpG,qIAAqI;AACrI,eAAesD,kBAAkBC,OAAqB,EAAE1B,IAAgB,EAAEhB,GAAW,EAAE2C,OAAe,EAAEC,YAAoB;IAC1H,MAAMC,OAAO,MAAM7B,KAAKN,OAAO,CAAC;IAChC,MAAMoC,QAAQ,IAAIC,IAAI,AAAE,CAAA,MAAMF,KAAKG,GAAG,EAAC,EAA+BC,GAAG,CAAC,CAACC,IAAMA,EAAEC,IAAI;IACvF,MAAMC,QAAQtE,UAAUkB,KAAK2C,SAASU,MAAM,CAAC,CAACC,IAAMR,MAAMR,GAAG,CAACgB,EAAEC,OAAO;IACvE,MAAMC,aAAa/E,gBAAgBuB,SAASE;IAC5C,MAAMuD,WAAqB,EAAE;IAC7B,MAAMC,OAAoB,EAAE;IAC5B,KAAK,MAAMC,QAAQP,MAAO;QACxB,MAAM,EAAEQ,GAAG,EAAEH,UAAUI,YAAY,EAAE,GAAG9E,UAAU4E;QAClDF,SAASK,IAAI,IAAID;QACjBH,KAAKI,IAAI,CAACxE,WAAWsE,KAAKJ;IAC5B;IACA,MAAMpE,gBACJ4B,MACA;QACE,IAAI0C,KAAKK,MAAM,GAAG,GAAG,MAAM/C,KAAKgD,QAAQ,CAACzE,oBAAoBmE;IAC/D,GACAvE;IAEF,MAAMD,QAAQ8B,MAAM,YAAY4B;IAChC,OAAOa;AACT;AAEA,6FAA6F;AAC7F,8DAA8D;AAC9D,SAASQ,mBAAmBjC,MAAoB;IAC9C,MAAMkC,SAAS1D,eAAewB,OAAOjC,EAAE;IACvC,IAAImE,WAAW,QAAQA,WAAWlC,OAAOf,QAAQ,EAAE,OAAO;IAC1D,OAAO;AACT;AAEA,eAAekD,aAAanC,MAAoB,EAAEhB,IAAgB,EAAEhB,GAAW;IAC7E,+FAA+F;IAC/F,+EAA+E;IAC/E,MAAMZ,gBAAgB4B,MAAM,IAAME,mBAAmBF,MAAMhB,KAAKgC,OAAOf,QAAQ,GAAG9B;AACpF;AAEA,eAAeiF,sBAAsBpC,MAAoB,EAAEqC,KAAiB,EAAEC,EAAU;IACtFtC,OAAOjC,EAAE,CAACM,IAAI,CAAC,CAAC,sBAAsB,EAAEiE,IAAI;AAC9C;AAEA,+FAA+F;AAC/F,OAAO,MAAMC,oBAA+C;IAC1DC,UAAU7E;IACV8E,eAAe7E;IACf8E,kBAAkBlF;IAClBsC;IACAG;IACAC;IACAO;IACAwB;IACAE;IACAC;IACA1E,aAAa,CAACsC,QAAQhB,MAAMhB,MAAQN,YAAYsC,OAAOjC,EAAE,EAAEiB,MAAMhB;AACnE,EAAE;AAEF,OAAO,eAAe2E,WAAW3E,GAAmB;IAClD,OAAOhB,gBAAgBgB,KAAKuE;AAC9B"}

package/dist/esm/store/sqlite/reconcile.d.ts

@@ -1,10 +1,5 @@
-import type { Config } from '../../config/index.js';
-import type { Connection } from '../types.js';
-export declare function reconcile(conn: Connection, cfg: Config, baseDir: string): Promise<{
-    parsed: number;
-    warnings: string[];
-}>;
-export declare function changedSignatureKeys(before: string, after: string): Set<string>;
-export declare function embedIdentityAdopted(before: string, after: string): boolean;
-export declare function signatureDiff(before: string, after: string): string;
-export declare function rebuildContentTable(conn: Connection, cfg: Config, baseDir: string): Promise<string[]>;
+import type { ParsedDoc } from '../../scan/index.js';
+import type { ReconcileDialect } from '../types.js';
+export declare const INSERT_CONTENT_SQL = "INSERT INTO content (rowid, title, summary, text, \"path\", title_seg, summary_seg, text_seg) VALUES ((SELECT rowid FROM frontmatter WHERE \"path\" = ?), ?, ?, ?, ?, ?, ?, ?)";
+export declare function contentRow(doc: ParsedDoc, segmenting: boolean): unknown[];
+export declare const sqliteDialect: ReconcileDialect;

package/dist/esm/store/sqlite/reconcile.js

@@ -1,29 +1,18 @@
 import { contentTokenize } from '../../config/index.js';
 import { SenseError } from '../../errors.js';
-import { activeFeatures } from '../../features/index.js';
-import { progress } from '../../output/progress.js';
-import { listFiles, parseFile, RESERVED_COLUMNS } from '../../scan/index.js';
-import { reparseFiles } from '../../scan/reparse.js';
 import { segmentField } from '../../text/segment.js';
-import { getColumns, getMeta, quoteIdent, setMeta } from '../shared.js';
-import { BEGIN_WRITE, withTransaction } from '../transaction.js';
-// Feature-owned columns (`_rank`) must stay out of the upsert: a reparse would null the last
-// computed value on every touch, not just the reconciles that recompute it.
-const CORE_FRONTMATTER_COLUMNS = new Set([
-    'path',
-    '_mtime',
-    '_ctime',
-    '_size',
-    '_parse_error'
-]);
+import { recordReconcileDuration } from '../shared.js';
+import { BEGIN_WRITE } from '../transaction.js';
+// This store's dialect (types.ts's ReconcileDialect) for the shared orchestration in
+// store/reconcile.ts. content is FTS5-virtual with "_seg" sidecars for lexical.ts.
 // SQLite's compile-time SQLITE_MAX_COLUMN, default 2000 (https://www.sqlite.org/limits.html).
 const MAX_FRONTMATTER_COLUMNS = 2000;
-// Shared by reconcile() and the tokenize-only rebuild in open(), so both prepare the same
-// literal instead of two copies drifting apart.
-const INSERT_CONTENT_SQL = `INSERT INTO content (rowid, title, summary, text, "path", title_seg, summary_seg, text_seg) VALUES ((SELECT rowid FROM frontmatter WHERE "path" = ?), ?, ?, ?, ?, ?, ?, ?)`;
+// Shared by reconcileContent() and the tokenize-only rebuild in open.ts, so both prepare the
+// same literal instead of two copies drifting apart.
+export const INSERT_CONTENT_SQL = `INSERT INTO content (rowid, title, summary, text, "path", title_seg, summary_seg, text_seg) VALUES ((SELECT rowid FROM frontmatter WHERE "path" = ?), ?, ?, ?, ?, ?, ?, ?)`;
 // A single content row's param tuple, matching INSERT_CONTENT_SQL's placeholder order.
 // Assumes the frontmatter row for doc.relPath already exists (rowid subquery).
-function contentRow(doc, segmenting) {
+export function contentRow(doc, segmenting) {
     return [
         doc.relPath,
         doc.search.title,
@@ -35,188 +24,29 @@ function contentRow(doc, segmenting) {
         segmenting ? segmentField(doc.search.text) : ''
     ];
 }
-export async function reconcile(conn, cfg, baseDir) {
-    const files = listFiles(cfg, baseDir);
-    const currentSet = new Set(files.map((f)=>f.relPath));
-    const existingStmt = await conn.prepare('SELECT "path", "_mtime", "_size" FROM frontmatter');
-    const existingRows = await existingStmt.all();
-    const existing = new Map(existingRows.map((r)=>[
-            r.path,
-            r
+async function reconcileContent(conn, touched, docs, _delta, cfg) {
+    // FTS5 has no upsert, so delete-before-insert, coupled to the frontmatter rowid (indexed via
+    // its PRIMARY KEY) rather than the UNINDEXED `path` column, which a per-row DELETE would scan to find.
+    if (touched.length > 0) await conn.runBatch('DELETE FROM content WHERE rowid = (SELECT rowid FROM frontmatter WHERE "path" = ?)', touched.map((p)=>[
+            p
         ]));
-    const vanished = existingRows.filter((r)=>!currentSet.has(r.path)).map((r)=>r.path);
-    const toReparse = files.filter((f)=>{
-        const row = existing.get(f.relPath);
-        return !row || row._mtime !== f.mtimeMs || row._size !== f.size;
-    });
-    if (vanished.length === 0 && toReparse.length === 0) return {
-        parsed: 0,
-        warnings: []
-    };
-    const features = activeFeatures(cfg);
-    const seenColumns = await getColumns(conn);
-    // Bulk reparses (a sync, a cold build) are the long silences a query can hit; short
-    // reconciles stay silent (progress() has a threshold).
-    const report = progress('reparsing files', toReparse.length);
-    const { docs: parsedDocs, warnings, newColumns } = await reparseFiles(toReparse, features, cfg, seenColumns, report.tick);
-    report.finish();
-    for (const col of newColumns)seenColumns.add(col);
-    const allColumns = [
-        ...seenColumns
-    ];
-    // Fence before ALTERing: SQLite's own failure past this point is a raw
-    // "too many columns on sqlite_altertab_frontmatter" with no indication of the boundary or the levers.
-    if (allColumns.length > MAX_FRONTMATTER_COLUMNS) {
-        throw new SenseError('COLUMN_LIMIT', `frontmatter would need ${allColumns.length} columns, crossing SQLite's compile-time SQLITE_MAX_COLUMN limit (default ${MAX_FRONTMATTER_COLUMNS}; see https://www.sqlite.org/limits.html). Narrow the presets' include globs so fewer/other files are indexed, or fix whatever is generating unbounded frontmatter keys.`);
+    if (docs.length > 0) {
+        // A non-default tokenizer means the tree has chosen its own scheme; a phrase query over
+        // grapheme runs would be nonsense against trigram, so the sidecars stay empty.
+        const segmenting = contentTokenize(cfg) === undefined;
+        await conn.runBatch(INSERT_CONTENT_SQL, docs.map((doc)=>contentRow(doc, segmenting)));
     }
-    // Columns the frontmatter upsert actually writes: core + parsed frontmatter keys, never a
-    // feature-owned reserved column (see CORE_FRONTMATTER_COLUMNS above).
-    const writableColumns = allColumns.filter((c)=>CORE_FRONTMATTER_COLUMNS.has(c) || !RESERVED_COLUMNS.has(c));
-    // ON CONFLICT UPDATE (not OR REPLACE) keeps the row's rowid stable across reparses --
-    // content rows are coupled to that rowid below.
-    const insertSql = `INSERT INTO frontmatter (${writableColumns.map(quoteIdent).join(', ')}) VALUES (${writableColumns.map(()=>'?').join(', ')}) ON CONFLICT("path") DO UPDATE SET ${writableColumns.filter((c)=>c !== 'path').map((c)=>`${quoteIdent(c)} = excluded.${quoteIdent(c)}`).join(', ')}`;
-    const added = toReparse.filter((f)=>!existing.has(f.relPath)).map((f)=>f.relPath);
-    const delta = {
-        files,
-        reparsed: parsedDocs.map((d)=>d.relPath),
-        added,
-        vanished
-    };
-    const addedSet = new Set(added);
-    const reparsedExisting = parsedDocs.map((d)=>d.relPath).filter((p)=>!addedSet.has(p));
-    const txStart = Date.now();
-    await withTransaction(conn, async ()=>{
-        var _feature_remove, _feature_afterReconcile;
-        // Re-read inside the write transaction: newColumns came from a read taken before it opened, so a
-        // concurrent reconcile may have added some of them since. ALTER has no IF NOT EXISTS.
-        const present = await getColumns(conn);
-        for (const col of newColumns)if (!present.has(col)) await conn.exec(`ALTER TABLE frontmatter ADD COLUMN ${quoteIdent(col)}`);
-        // content's rowid lookup depends on its coupled frontmatter row, so the content
-        // deletes/inserts run while every path still has one -- the vanished frontmatter delete comes last.
-        if (vanished.length > 0) {
-            await conn.runBatch('DELETE FROM content WHERE rowid = (SELECT rowid FROM frontmatter WHERE "path" = ?)', vanished.map((p)=>[
-                    p
-                ]));
-        }
-        if (reparsedExisting.length > 0) {
-            // FTS5 has no upsert, so delete-before-insert, coupled to the frontmatter rowid (indexed via
-            // its PRIMARY KEY) rather than the UNINDEXED `path` column, which a per-row DELETE would scan to find.
-            await conn.runBatch('DELETE FROM content WHERE rowid = (SELECT rowid FROM frontmatter WHERE "path" = ?)', reparsedExisting.map((p)=>[
-                    p
-                ]));
-        }
-        if (parsedDocs.length > 0) {
-            const rows = parsedDocs.map((doc)=>writableColumns.map((col)=>{
-                    var _doc_data_col;
-                    if (col === 'path') return doc.relPath;
-                    if (col === '_mtime') return doc.mtimeMs;
-                    if (col === '_ctime') return doc.ctimeMs;
-                    if (col === '_size') return doc.size;
-                    // Written per parse, unlike _rank, which a feature pass owns and the upsert skips.
-                    if (col === '_parse_error') return doc.parseError;
-                    return (_doc_data_col = doc.data[col]) !== null && _doc_data_col !== void 0 ? _doc_data_col : null;
-                }));
-            await conn.runBatch(insertSql, rows);
-            // A non-default tokenizer means the tree has chosen its own scheme; a phrase query over
-            // grapheme runs would be nonsense against trigram, so the sidecars stay empty.
-            const segmenting = contentTokenize(cfg) === undefined;
-            await conn.runBatch(INSERT_CONTENT_SQL, parsedDocs.map((doc)=>contentRow(doc, segmenting)));
-        }
-        if (vanished.length > 0) await conn.runBatch('DELETE FROM frontmatter WHERE "path" = ?', vanished.map((p)=>[
-                p
-            ]));
-        // A preset edit forces a full rebuild, so an unchanged doc's coverage is already correct;
-        // new docs have nothing to clear, which keeps cold builds linear.
-        const presetTouched = [
-            ...vanished,
-            ...reparsedExisting
-        ];
-        if (presetTouched.length > 0) await conn.runBatch('DELETE FROM preset_files WHERE "path" = ?', presetTouched.map((p)=>[
-                p
-            ]));
-        const presetRows = [];
-        for (const doc of parsedDocs)for (const presetName of doc.presets)presetRows.push([
-            doc.relPath,
-            presetName
-        ]);
-        if (presetRows.length > 0) await conn.runBatch('INSERT INTO preset_files ("path", preset) VALUES (?, ?)', presetRows);
-        const removedPaths = [
-            ...vanished,
-            ...reparsedExisting
-        ];
-        if (removedPaths.length > 0) for (const feature of features)await ((_feature_remove = feature.remove) === null || _feature_remove === void 0 ? void 0 : _feature_remove.call(feature, conn, removedPaths, delta));
-        for (const feature of features){
-            var _feature_store;
-            const docsForFeature = parsedDocs.map((doc)=>({
-                    path: doc.relPath,
-                    extracted: doc.extracted[feature.name]
-                }));
-            await ((_feature_store = feature.store) === null || _feature_store === void 0 ? void 0 : _feature_store.call(feature, conn, docsForFeature, delta));
-        }
-        for (const feature of features)await ((_feature_afterReconcile = feature.afterReconcile) === null || _feature_afterReconcile === void 0 ? void 0 : _feature_afterReconcile.call(feature, conn, delta));
-    }, BEGIN_WRITE);
-    // Reconcile's own write-transaction duration, for open()'s derived busy_timeout (F):
-    // keep the observed max so a big watcher reconcile's lock hold is what the next open bounds its wait against.
-    const durationMs = Date.now() - txStart;
-    const prevRaw = await getMeta(conn, 'reconcile_max_ms');
-    // -1, not 0, so a genuinely 0ms first reconcile (sub-millisecond, common on a tiny tree)
-    // still gets recorded instead of losing to the "nothing recorded yet" default.
-    const prevMax = prevRaw === null ? -1 : Number(prevRaw);
-    if (durationMs > prevMax) await setMeta(conn, 'reconcile_max_ms', String(durationMs));
-    return {
-        parsed: parsedDocs.length,
-        warnings
-    };
-}
-// Segment keys that moved between two feature signatures (see config.featureSignature's
-// format: global features, embed provider, tokenize, then one segment per preset).
-export function changedSignatureKeys(before, after) {
-    const keyOf = (part)=>part.startsWith('preset:') ? part.split(':').slice(0, 2).join(':') : part.split(':')[0];
-    const parse = (sig)=>new Map(sig.split('|').map((part)=>[
-                keyOf(part),
-                part
-            ]));
-    const a = parse(before);
-    const b = parse(after);
-    const changed = new Set();
-    for (const [key, val] of b)if (a.get(key) !== val) changed.add(key);
-    for (const key of a.keys())if (!b.has(key)) changed.add(key);
-    return changed;
-}
-// Whether the embed segment only gained its resolved weight identity: same provider and
-// model, no identity recorded before, one now -- adopted into meta without a rebuild.
-export function embedIdentityAdopted(before, after) {
-    const embedPart = (sig)=>sig.split('|').find((p)=>p.startsWith('embed:'));
-    const b = embedPart(before);
-    const a = embedPart(after);
-    if (b === undefined || a === undefined) return false;
-    const at = a.indexOf('@');
-    return b.indexOf('@') === -1 && at !== -1 && a.slice(0, at) === b;
-}
-// Names what moved, for the rebuild notice.
-export function signatureDiff(before, after) {
-    const changed = changedSignatureKeys(before, after);
-    const label = (key)=>key === 'embed' ? 'embed settings' : key === 'tokenize' ? 'content tokenizer' : key.startsWith('preset:') ? `preset "${key.slice(7)}"` : 'features';
-    return changed.size === 0 ? 'features' : [
-        ...changed
-    ].map(label).join(', ');
-}
-// Tokenize-only rebuild: only `content` is repopulated (the other tables are tokenizer-independent)
-// and no feature extractors run. Returns per-file warnings, since mtimes stay untouched and reconcile would never reparse them.
-export async function rebuildContentTable(conn, cfg, baseDir) {
-    const stmt = await conn.prepare('SELECT "path" FROM frontmatter');
-    const known = new Set((await stmt.all()).map((r)=>r.path));
-    const files = listFiles(cfg, baseDir).filter((f)=>known.has(f.relPath));
-    const segmenting = contentTokenize(cfg) === undefined;
-    const warnings = [];
-    const rows = [];
-    for (const file of files){
-        const { doc, warnings: fileWarnings } = parseFile(file);
-        warnings.push(...fileWarnings);
-        rows.push(contentRow(doc, segmenting));
-    }
-    await withTransaction(conn, async ()=>{
-        if (rows.length > 0) await conn.runBatch(INSERT_CONTENT_SQL, rows);
-    }, BEGIN_WRITE);
-    return warnings;
 }
+export const sqliteDialect = {
+    beginMode: ()=>BEGIN_WRITE,
+    checkColumnLimit (count) {
+        // Fence before ALTERing: SQLite's own failure past this point is a raw
+        // "too many columns on sqlite_altertab_frontmatter" with no indication of the boundary or the levers.
+        if (count > MAX_FRONTMATTER_COLUMNS) {
+            throw new SenseError('COLUMN_LIMIT', `frontmatter would need ${count} columns, crossing SQLite's compile-time SQLITE_MAX_COLUMN limit (default ${MAX_FRONTMATTER_COLUMNS}; see https://www.sqlite.org/limits.html). Narrow the presets' include globs so fewer/other files are indexed, or fix whatever is generating unbounded frontmatter keys.`);
+        }
+    },
+    columnTypeSuffix: ()=>'',
+    reconcileContent,
+    recordDuration: recordReconcileDuration
+};