@jigyasudham/veto 3.0.0 → 3.1.1

package/dist/transcripts/schema.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"schema.d.ts","sourceRoot":"","sources":["../../src/transcripts/schema.ts"],"names":[],"mappings":"AAUA,eAAO,MAAM,0BAA0B,IAAI,CAAC;AAG5C,eAAO,MAAM,WAAW,4GAEd,CAAC;AACX,MAAM,MAAM,SAAS,GAAG,CAAC,OAAO,WAAW,CAAC,CAAC,MAAM,CAAC,CAAC;AAErD,MAAM,MAAM,SAAS,GAAG;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,CAAC;AAExD,eAAO,MAAM,UAAU,EAAE,SAAS,EA+HjC,CAAC;AAIF,eAAO,MAAM,gBAAgB,EAAE,SAAS,EAAsE,CAAC;AAE/G,MAAM,MAAM,UAAU,GAAG;IACvB,EAAE,EAAE,MAAM,CAAC;IACX,MAAM,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,MAAM,CAAC;IAC1B,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,MAAM,CAAC;IACvB,YAAY,EAAE,MAAM,CAAC;IACrB,aAAa,EAAE,MAAM,CAAC;IACtB,kBAAkB,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;CACpB,CAAC;AAEF,MAAM,MAAM,aAAa,GAAG;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,MAAM,CAAC;IAC1B,eAAe,EAAE,MAAM,CAAC;IACxB,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;CACtB,CAAC;AAEF,MAAM,MAAM,YAAY,GAAG;IACzB,EAAE,EAAE,MAAM,CAAC;IACX,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,GAAG,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,SAAS,CAAC;IAChB,GAAG,EAAE,MAAM,CAAC;CACb,CAAC;AAEF,MAAM,MAAM,gBAAgB,GAAG;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,EAAE,EAAE,MAAM,CAAC;CACZ,CAAC;AAEF,MAAM,MAAM,QAAQ,GAAG;IACrB,EAAE,EAAE,MAAM,CAAC;IACX,UAAU,EAAE,MAAM,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,GAAG,EAAE,MAAM,CAAC;IACZ,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE,SAAS,CAAC;IAChB,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;CACpB,CAAC"}
+{"version":3,"file":"schema.d.ts","sourceRoot":"","sources":["../../src/transcripts/schema.ts"],"names":[],"mappings":"AAUA,eAAO,MAAM,0BAA0B,IAAI,CAAC;AAG5C,eAAO,MAAM,WAAW,4GAEd,CAAC;AACX,MAAM,MAAM,SAAS,GAAG,CAAC,OAAO,WAAW,CAAC,CAAC,MAAM,CAAC,CAAC;AAErD,MAAM,MAAM,SAAS,GAAG;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,CAAC;AAExD,eAAO,MAAM,UAAU,EAAE,SAAS,EA+JjC,CAAC;AAIF,eAAO,MAAM,gBAAgB,EAAE,SAAS,EAAsE,CAAC;AAE/G,MAAM,MAAM,UAAU,GAAG;IACvB,EAAE,EAAE,MAAM,CAAC;IACX,MAAM,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,MAAM,CAAC;IAC1B,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,eAAe,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,MAAM,CAAC;IACvB,YAAY,EAAE,MAAM,CAAC;IACrB,aAAa,EAAE,MAAM,CAAC;IACtB,kBAAkB,EAAE,MAAM,GAAG,IAAI,CAAC;IAClC,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,WAAW,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,iFAAiF;IACjF,eAAe,EAAE,MAAM,CAAC;IACxB,qEAAqE;IACrE,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;CAC5B,CAAC;AAEF,MAAM,MAAM,eAAe,GAAG;IAC5B,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,GAAG,EAAE,UAAU,CAAC;IAChB,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAEF,MAAM,MAAM,aAAa,GAAG;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf,iBAAiB,EAAE,MAAM,CAAC;IAC1B,eAAe,EAAE,MAAM,CAAC;IACxB,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;CACtB,CAAC;AAEF,MAAM,MAAM,YAAY,GAAG;IACzB,EAAE,EAAE,MAAM,CAAC;IACX,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,GAAG,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,SAAS,CAAC;IAChB,GAAG,EAAE,MAAM,CAAC;CACb,CAAC;AAEF,MAAM,MAAM,gBAAgB,GAAG;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,EAAE,EAAE,MAAM,CAAC;CACZ,CAAC;AAEF,MAAM,MAAM,QAAQ,GAAG;IACrB,EAAE,EAAE,MAAM,CAAC;IACX,UAAU,EAAE,MAAM,CAAC;IACnB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,GAAG,EAAE,MAAM,CAAC;IACZ,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE,SAAS,CAAC;IAChB,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,WAAW,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,MAAM,GAAG,IAAI,CAAC;IACzB,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,EAAE,MAAM,CAAC;CACpB,CAAC"}

package/dist/transcripts/schema.js

@@ -7,7 +7,7 @@
 // rather than editing an existing one, so the schema grows without churn and
 // re-open is idempotent.
 // Highest migration version defined below.
-export const TRANSCRIPTS_SCHEMA_VERSION = 3;
+export const TRANSCRIPTS_SCHEMA_VERSION = 4;
 // The common vocabulary every source normalizes into.
 export const EVENT_KINDS = [
     'user_message', 'assistant_message', 'tool_call', 'tool_result', 'reasoning', 'meta', 'unknown',
@@ -17,39 +17,39 @@ export const MIGRATIONS = [
         // v1 — the two foundational anchor tables. Events / segments / FTS get their
         // own later migrations (Steps 6/9/10) so we don't big-bang the schema.
         version: 1,
-        up: `
-      -- L0 archive index: one row per captured source-CLI session.
-      CREATE TABLE IF NOT EXISTS archives (
-        id                  TEXT PRIMARY KEY,
-        source              TEXT NOT NULL DEFAULT 'claude',   -- claude|codex|gemini
-        source_session_id   TEXT NOT NULL,
-        project_dir         TEXT,                              -- normalized (lowercase drive)
-        veto_session_id     TEXT,                              -- link to veto.db sessions.id when known
-        archive_path        TEXT NOT NULL,                     -- absolute path to the .gz on disk
-        content_sha256      TEXT NOT NULL,                     -- of raw source bytes (dedup key)
-        source_bytes        INTEGER NOT NULL DEFAULT 0,        -- uncompressed size
-        archive_bytes       INTEGER NOT NULL DEFAULT 0,        -- gz size on disk
-        source_format_hint  TEXT,                              -- e.g. 'claude-jsonl'
-        parser_version      INTEGER NOT NULL DEFAULT 0,
-        indexed_through_seq INTEGER NOT NULL DEFAULT 0,        -- lazy-index watermark
-        captured_at         TEXT NOT NULL,
-        updated_at          TEXT NOT NULL,
-        UNIQUE(source, source_session_id)
-      );
-      CREATE INDEX IF NOT EXISTS idx_archives_project ON archives(project_dir);
-      CREATE INDEX IF NOT EXISTS idx_archives_sha     ON archives(content_sha256);
-
-      -- Live host-session -> transcript-file mapping, UPSERTed fire-and-forget by
-      -- the statusline (Step 4) so save-time capture knows which file to read.
-      CREATE TABLE IF NOT EXISTS session_map (
-        source            TEXT NOT NULL DEFAULT 'claude',
-        source_session_id TEXT NOT NULL,
-        transcript_path   TEXT NOT NULL,
-        project_dir       TEXT,
-        last_seen_at      TEXT NOT NULL,
-        PRIMARY KEY (source, source_session_id)
-      );
-      CREATE INDEX IF NOT EXISTS idx_session_map_project ON session_map(project_dir);
+        up: `
+      -- L0 archive index: one row per captured source-CLI session.
+      CREATE TABLE IF NOT EXISTS archives (
+        id                  TEXT PRIMARY KEY,
+        source              TEXT NOT NULL DEFAULT 'claude',   -- claude|codex|gemini
+        source_session_id   TEXT NOT NULL,
+        project_dir         TEXT,                              -- normalized (lowercase drive)
+        veto_session_id     TEXT,                              -- link to veto.db sessions.id when known
+        archive_path        TEXT NOT NULL,                     -- absolute path to the .gz on disk
+        content_sha256      TEXT NOT NULL,                     -- of raw source bytes (dedup key)
+        source_bytes        INTEGER NOT NULL DEFAULT 0,        -- uncompressed size
+        archive_bytes       INTEGER NOT NULL DEFAULT 0,        -- gz size on disk
+        source_format_hint  TEXT,                              -- e.g. 'claude-jsonl'
+        parser_version      INTEGER NOT NULL DEFAULT 0,
+        indexed_through_seq INTEGER NOT NULL DEFAULT 0,        -- lazy-index watermark
+        captured_at         TEXT NOT NULL,
+        updated_at          TEXT NOT NULL,
+        UNIQUE(source, source_session_id)
+      );
+      CREATE INDEX IF NOT EXISTS idx_archives_project ON archives(project_dir);
+      CREATE INDEX IF NOT EXISTS idx_archives_sha     ON archives(content_sha256);
+
+      -- Live host-session -> transcript-file mapping, UPSERTed fire-and-forget by
+      -- the statusline (Step 4) so save-time capture knows which file to read.
+      CREATE TABLE IF NOT EXISTS session_map (
+        source            TEXT NOT NULL DEFAULT 'claude',
+        source_session_id TEXT NOT NULL,
+        transcript_path   TEXT NOT NULL,
+        project_dir       TEXT,
+        last_seen_at      TEXT NOT NULL,
+        PRIMARY KEY (source, source_session_id)
+      );
+      CREATE INDEX IF NOT EXISTS idx_session_map_project ON session_map(project_dir);
     `,
     },
     {
@@ -58,33 +58,33 @@ export const MIGRATIONS = [
         // UNCOMPRESSED L0 bytes for exact expansion (Step 11). Unknown/unparsed lines
         // are stored+ordered (kind='unknown') so format drift never loses data.
         version: 2,
-        up: `
-      CREATE TABLE IF NOT EXISTS events (
-        id                TEXT PRIMARY KEY,
-        archive_id        TEXT NOT NULL,
-        source_session_id TEXT NOT NULL,
-        seq               INTEGER NOT NULL,        -- monotonic per archive, per event
-        line_index        INTEGER NOT NULL,        -- source JSONL line (0-based)
-        block_index       INTEGER NOT NULL DEFAULT 0,
-        kind              TEXT NOT NULL,           -- EVENT_KINDS
-        source_type       TEXT,                    -- verbatim source 'type' (user/assistant/system/…)
-        role              TEXT,
-        tool_name         TEXT,                    -- for tool_call/tool_result
-        text              TEXT,                    -- normalized + MASKED content for FTS
-        secret_count      INTEGER NOT NULL DEFAULT 0,
-        event_uuid        TEXT,                    -- native uuid (Claude)
-        parent_uuid       TEXT,                    -- native parentUuid (tree)
-        is_sidechain      INTEGER NOT NULL DEFAULT 0,
-        ts_source         TEXT,                    -- verbatim timestamp string
-        ts_utc            TEXT,                    -- normalized ISO
-        raw_offset        INTEGER NOT NULL DEFAULT 0,
-        raw_length        INTEGER NOT NULL DEFAULT 0,
-        UNIQUE(archive_id, seq),
-        FOREIGN KEY (archive_id) REFERENCES archives(id)
-      );
-      CREATE INDEX IF NOT EXISTS idx_events_archive ON events(archive_id, seq);
-      CREATE INDEX IF NOT EXISTS idx_events_session ON events(source_session_id);
-      CREATE INDEX IF NOT EXISTS idx_events_kind    ON events(kind);
+        up: `
+      CREATE TABLE IF NOT EXISTS events (
+        id                TEXT PRIMARY KEY,
+        archive_id        TEXT NOT NULL,
+        source_session_id TEXT NOT NULL,
+        seq               INTEGER NOT NULL,        -- monotonic per archive, per event
+        line_index        INTEGER NOT NULL,        -- source JSONL line (0-based)
+        block_index       INTEGER NOT NULL DEFAULT 0,
+        kind              TEXT NOT NULL,           -- EVENT_KINDS
+        source_type       TEXT,                    -- verbatim source 'type' (user/assistant/system/…)
+        role              TEXT,
+        tool_name         TEXT,                    -- for tool_call/tool_result
+        text              TEXT,                    -- normalized + MASKED content for FTS
+        secret_count      INTEGER NOT NULL DEFAULT 0,
+        event_uuid        TEXT,                    -- native uuid (Claude)
+        parent_uuid       TEXT,                    -- native parentUuid (tree)
+        is_sidechain      INTEGER NOT NULL DEFAULT 0,
+        ts_source         TEXT,                    -- verbatim timestamp string
+        ts_utc            TEXT,                    -- normalized ISO
+        raw_offset        INTEGER NOT NULL DEFAULT 0,
+        raw_length        INTEGER NOT NULL DEFAULT 0,
+        UNIQUE(archive_id, seq),
+        FOREIGN KEY (archive_id) REFERENCES archives(id)
+      );
+      CREATE INDEX IF NOT EXISTS idx_events_archive ON events(archive_id, seq);
+      CREATE INDEX IF NOT EXISTS idx_events_session ON events(source_session_id);
+      CREATE INDEX IF NOT EXISTS idx_events_kind    ON events(kind);
     `,
     },
     {
@@ -104,39 +104,71 @@ export const MIGRATIONS = [
         // loads its module, which throws on exactly the Nodes this redo supports.
         // Databases that ran the old dev-only v3 keep an inert orphan table.
         version: 3,
-        up: `
-      -- One row per indexed event. project_dir/session denormalized so the
-      -- hot query is a single join (same shape events_fts had). The INTEGER id
-      -- exists so postings never carry a 36-char uuid — measured, that uuid
-      -- (stored twice per posting: PK + purge index) was ~3x the entire
-      -- index's justified size.
-      CREATE TABLE IF NOT EXISTS search_docs (
-        id                INTEGER PRIMARY KEY,
-        event_id          TEXT NOT NULL UNIQUE,
-        archive_id        TEXT NOT NULL,
-        source_session_id TEXT NOT NULL,
-        project_dir       TEXT,
-        seq               INTEGER NOT NULL,
-        kind              TEXT NOT NULL,
-        len               INTEGER NOT NULL   -- token count, for BM25 length norm
-      );
-      CREATE INDEX IF NOT EXISTS idx_search_docs_archive ON search_docs(archive_id);
-      CREATE INDEX IF NOT EXISTS idx_search_docs_scope
-        ON search_docs(project_dir, source_session_id);
-
-      -- Term dictionary: term text stored once, compact integer join key.
-      CREATE TABLE IF NOT EXISTS search_terms (
-        id   INTEGER PRIMARY KEY,
-        term TEXT NOT NULL UNIQUE
-      );
-
-      CREATE TABLE IF NOT EXISTS search_postings (
-        term_id INTEGER NOT NULL,
-        doc_id  INTEGER NOT NULL,
-        tf      INTEGER NOT NULL,
-        PRIMARY KEY (term_id, doc_id)
-      ) WITHOUT ROWID;
-      CREATE INDEX IF NOT EXISTS idx_postings_doc ON search_postings(doc_id);
+        up: `
+      -- One row per indexed event. project_dir/session denormalized so the
+      -- hot query is a single join (same shape events_fts had). The INTEGER id
+      -- exists so postings never carry a 36-char uuid — measured, that uuid
+      -- (stored twice per posting: PK + purge index) was ~3x the entire
+      -- index's justified size.
+      CREATE TABLE IF NOT EXISTS search_docs (
+        id                INTEGER PRIMARY KEY,
+        event_id          TEXT NOT NULL UNIQUE,
+        archive_id        TEXT NOT NULL,
+        source_session_id TEXT NOT NULL,
+        project_dir       TEXT,
+        seq               INTEGER NOT NULL,
+        kind              TEXT NOT NULL,
+        len               INTEGER NOT NULL   -- token count, for BM25 length norm
+      );
+      CREATE INDEX IF NOT EXISTS idx_search_docs_archive ON search_docs(archive_id);
+      CREATE INDEX IF NOT EXISTS idx_search_docs_scope
+        ON search_docs(project_dir, source_session_id);
+
+      -- Term dictionary: term text stored once, compact integer join key.
+      CREATE TABLE IF NOT EXISTS search_terms (
+        id   INTEGER PRIMARY KEY,
+        term TEXT NOT NULL UNIQUE
+      );
+
+      CREATE TABLE IF NOT EXISTS search_postings (
+        term_id INTEGER NOT NULL,
+        doc_id  INTEGER NOT NULL,
+        tf      INTEGER NOT NULL,
+        PRIMARY KEY (term_id, doc_id)
+      ) WITHOUT ROWID;
+      CREATE INDEX IF NOT EXISTS idx_postings_doc ON search_postings(doc_id);
+    `,
+    },
+    {
+        // v4 — chunk vectors for the semantic layer (Phase B).
+        //
+        // One row per WINDOW, not per event: whole-event vectors fail the
+        // paraphrase gate to mean-pooling dilution (spec §11 amendment). Keyed by
+        // the same INTEGER doc_id the postings use — storing the 36-char event
+        // uuid here would repeat the mistake that made the lexical index 3x too
+        // big. Chunks are internal (C6): nothing outside search.ts sees them.
+        //
+        // No `scale` column: cosine similarity is invariant to a positive
+        // per-vector scalar, so the dequantization scale cancels entirely. The
+        // stored `norm` is the L2 norm of the int8 vector itself, precomputed so a
+        // query is one dot product and one divide.
+        //
+        // Vectors live in their own lifecycle: `archives.chunker_version` and
+        // `archives.embed_model` gate re-derivation independently of
+        // parser_version, so ingest never depends on the model being installed and
+        // installing it later backfills without re-parsing anything.
+        version: 4,
+        up: `
+      CREATE TABLE IF NOT EXISTS search_vectors (
+        doc_id      INTEGER NOT NULL,
+        chunk_index INTEGER NOT NULL,
+        vec         BLOB NOT NULL,     -- int8, one byte per dimension
+        norm        REAL NOT NULL,     -- L2 norm of the int8 vector
+        PRIMARY KEY (doc_id, chunk_index)
+      );
+
+      ALTER TABLE archives ADD COLUMN chunker_version INTEGER NOT NULL DEFAULT 0;
+      ALTER TABLE archives ADD COLUMN embed_model TEXT;
     `,
     },
 ];

package/dist/transcripts/schema.js.map

@@ -1 +1 @@
-{"version":3,"file":"schema.js","sourceRoot":"","sources":["../../src/transcripts/schema.ts"],"names":[],"mappings":"AAAA,oFAAoF;AACpF,EAAE;AACF,6EAA6E;AAC7E,8EAA8E;AAC9E,2EAA2E;AAC3E,2EAA2E;AAC3E,6EAA6E;AAC7E,yBAAyB;AAEzB,2CAA2C;AAC3C,MAAM,CAAC,MAAM,0BAA0B,GAAG,CAAC,CAAC;AAE5C,sDAAsD;AACtD,MAAM,CAAC,MAAM,WAAW,GAAG;IACzB,cAAc,EAAE,mBAAmB,EAAE,WAAW,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,EAAE,SAAS;CACvF,CAAC;AAKX,MAAM,CAAC,MAAM,UAAU,GAAgB;IACrC;QACE,6EAA6E;QAC7E,uEAAuE;QACvE,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAiCH;KACF;IACD;QACE,2EAA2E;QAC3E,uEAAuE;QACvE,8EAA8E;QAC9E,wEAAwE;QACxE,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;KA2BH;KACF;IACD;QACE,uEAAuE;QACvE,mEAAmE;QACnE,wEAAwE;QACxE,uEAAuE;QACvE,mEAAmE;QACnE,EAAE;QACF,0EAA0E;QAC1E,yEAAyE;QACzE,0EAA0E;QAC1E,sEAAsE;QACtE,iEAAiE;QACjE,EAAE;QACF,wEAAwE;QACxE,0EAA0E;QAC1E,qEAAqE;QACrE,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAiCH;KACF;CACF,CAAC;AAEF,0EAA0E;AAC1E,oDAAoD;AACpD,MAAM,CAAC,MAAM,gBAAgB,GAAgB,CAAC,cAAc,EAAE,mBAAmB,EAAE,WAAW,EAAE,aAAa,CAAC,CAAC"}
+{"version":3,"file":"schema.js","sourceRoot":"","sources":["../../src/transcripts/schema.ts"],"names":[],"mappings":"AAAA,oFAAoF;AACpF,EAAE;AACF,6EAA6E;AAC7E,8EAA8E;AAC9E,2EAA2E;AAC3E,2EAA2E;AAC3E,6EAA6E;AAC7E,yBAAyB;AAEzB,2CAA2C;AAC3C,MAAM,CAAC,MAAM,0BAA0B,GAAG,CAAC,CAAC;AAE5C,sDAAsD;AACtD,MAAM,CAAC,MAAM,WAAW,GAAG;IACzB,cAAc,EAAE,mBAAmB,EAAE,WAAW,EAAE,aAAa,EAAE,WAAW,EAAE,MAAM,EAAE,SAAS;CACvF,CAAC;AAKX,MAAM,CAAC,MAAM,UAAU,GAAgB;IACrC;QACE,6EAA6E;QAC7E,uEAAuE;QACvE,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAiCH;KACF;IACD;QACE,2EAA2E;QAC3E,uEAAuE;QACvE,8EAA8E;QAC9E,wEAAwE;QACxE,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;KA2BH;KACF;IACD;QACE,uEAAuE;QACvE,mEAAmE;QACnE,wEAAwE;QACxE,uEAAuE;QACvE,mEAAmE;QACnE,EAAE;QACF,0EAA0E;QAC1E,yEAAyE;QACzE,0EAA0E;QAC1E,sEAAsE;QACtE,iEAAiE;QACjE,EAAE;QACF,wEAAwE;QACxE,0EAA0E;QAC1E,qEAAqE;QACrE,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAiCH;KACF;IACD;QACE,uDAAuD;QACvD,EAAE;QACF,kEAAkE;QAClE,0EAA0E;QAC1E,uEAAuE;QACvE,wEAAwE;QACxE,sEAAsE;QACtE,EAAE;QACF,kEAAkE;QAClE,uEAAuE;QACvE,2EAA2E;QAC3E,2CAA2C;QAC3C,EAAE;QACF,sEAAsE;QACtE,6DAA6D;QAC7D,2EAA2E;QAC3E,6DAA6D;QAC7D,OAAO,EAAE,CAAC;QACV,EAAE,EAAE;;;;;;;;;;;KAWH;KACF;CACF,CAAC;AAEF,0EAA0E;AAC1E,oDAAoD;AACpD,MAAM,CAAC,MAAM,gBAAgB,GAAgB,CAAC,cAAc,EAAE,mBAAmB,EAAE,WAAW,EAAE,aAAa,CAAC,CAAC"}

package/dist/transcripts/search.d.ts

@@ -19,4 +19,39 @@ export type SearchOptions = {
  */
 export declare function buildSnippet(text: string, terms: Set<string>): string;
 export declare function searchEvents(query: string, opts?: SearchOptions): SearchHit[];
+/**
+ * One ranker's contribution for a 0-based rank.
+ *
+ * Exported so the property the fix turns on can be asserted directly rather
+ * than through retrieval, where it is not reliably constructible: with a
+ * bag-of-words embedding model, lexical overlap implies semantic similarity,
+ * so a document that matches a query's words but not its meaning barely
+ * exists. The arithmetic, however, is exact.
+ */
+export declare function rrfWeight(rank0: number): number;
+/**
+ * Fuse ranked id lists by reciprocal rank. Pure, for testing and reuse.
+ *
+ * The property that matters: a top-ranked find from ONE ranker must be able
+ * to outrank a document both rankers place mid-list. That is the case the
+ * semantic layer exists to serve — a strict paraphrase is found by the
+ * semantic ranker alone — and it is exactly what k=60 at depth 50 broke.
+ */
+export declare function fuseByRank(lists: string[][]): {
+    id: string;
+    score: number;
+}[];
+/**
+ * BM25 and semantic recall, fused by reciprocal rank.
+ *
+ * RRF over raw score blending is deliberate: BM25 scores are unbounded and
+ * corpus-dependent while cosines sit in [-1, 1], so any weighted sum would be
+ * a hidden tuning knob that drifts with corpus size. Ranks have neither
+ * problem, and lexical survival is measurable (a keyword hit must not be
+ * pushed out of the fused list by semantic noise).
+ *
+ * Degrades silently to BM25 alone when no vectors are stored or the model is
+ * not installed (A2) — the caller cannot tell the difference except by score.
+ */
+export declare function searchEventsHybrid(query: string, opts?: SearchOptions): SearchHit[];
 //# sourceMappingURL=search.d.ts.map

package/dist/transcripts/search.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"search.d.ts","sourceRoot":"","sources":["../../src/transcripts/search.ts"],"names":[],"mappings":"AAiBA,MAAM,MAAM,SAAS,GAAG;IACtB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;IAClB,eAAe,EAAE,MAAM,CAAC;IACxB,GAAG,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;CACf,CAAC;AAEF,MAAM,MAAM,aAAa,GAAG;IAC1B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB,CAAC;AAiBF;;;;GAIG;AACH,wBAAgB,YAAY,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,GAAG,CAAC,MAAM,CAAC,GAAG,MAAM,CAiCrE;AAED,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,GAAE,aAAkB,GAAG,SAAS,EAAE,CAiEjF"}
+{"version":3,"file":"search.d.ts","sourceRoot":"","sources":["../../src/transcripts/search.ts"],"names":[],"mappings":"AAoBA,MAAM,MAAM,SAAS,GAAG;IACtB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;IAClB,eAAe,EAAE,MAAM,CAAC;IACxB,GAAG,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;CACf,CAAC;AAEF,MAAM,MAAM,aAAa,GAAG;IAC1B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB,CAAC;AAiBF;;;;GAIG;AACH,wBAAgB,YAAY,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,GAAG,CAAC,MAAM,CAAC,GAAG,MAAM,CAiCrE;AAED,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,GAAE,aAAkB,GAAG,SAAS,EAAE,CA2EjF;AAgDD;;;;;;;;GAQG;AACH,wBAAgB,SAAS,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAE/C;AAED;;;;;;;GAOG;AACH,wBAAgB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,EAAE,GAAG;IAAE,EAAE,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,EAAE,CAQ7E;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,GAAE,aAAkB,GAAG,SAAS,EAAE,CA2FvF"}

package/dist/transcripts/search.js

@@ -12,6 +12,9 @@
 import { getTranscriptsDb } from './store.js';
 import { normalizeProjectDir } from '../memory/local.js';
 import { tokenize } from './tokenize.js';
+import { STRIDE_CHARS, WINDOW_CHARS } from './chunk.js';
+import { searchVectors } from './vectors.js';
+import { getDocCache, scopeMask } from './cache.js';
 const K1 = 1.2;
 const B = 0.75;
 // Terms present in more than half the corpus carry ~no signal but force
@@ -74,7 +77,8 @@ export function searchEvents(query, opts = {}) {
     const terms = [...new Set(tokenize(query))];
     if (terms.length === 0)
         return [];
-    const stats = db.prepare(`SELECT COUNT(*) AS n, COALESCE(AVG(len), 0) AS avgdl FROM search_docs`).get();
+    const docs = getDocCache();
+    const stats = { n: docs.count, avgdl: docs.avgdl };
     if (stats.n === 0)
         return [];
     // Resolve terms against the dictionary in one round trip.
@@ -88,47 +92,241 @@ export function searchEvents(query, opts = {}) {
     let kept = withDf.filter(t => t.df / stats.n <= DF_CAP_RATIO);
     if (kept.length === 0)
         kept = withDf; // every term is that common — keep them all
-    const where = [`p.term_id IN (${kept.map(() => '?').join(', ')})`];
-    const params = kept.map(t => t.id);
-    if (opts.projectDir) {
-        where.push('d.project_dir = ?');
-        params.push(normalizeProjectDir(opts.projectDir));
-    }
-    if (opts.sourceSessionId) {
-        where.push('d.source_session_id = ?');
-        params.push(opts.sourceSessionId);
-    }
-    const rows = db.prepare(`SELECT p.term_id, p.tf, d.event_id, d.archive_id, d.source_session_id, d.project_dir, d.seq, d.kind, d.len
-       FROM search_postings p JOIN search_docs d ON d.id = p.doc_id
-      WHERE ${where.join(' AND ')}`).all(...params);
+    // Postings only — three narrow columns, no join. Joining search_docs per
+    // POSTING to fetch columns needed for the final handful of rows cost 7.1 ms
+    // per common term at 50 sessions; this costs 2.6 ms. Scope and length come
+    // from the cached metadata instead, and full rows are read for the top-K.
+    const rows = db.prepare(`SELECT term_id, doc_id, tf FROM search_postings WHERE term_id IN (${kept.map(() => '?').join(', ')})`).all(...kept.map(t => t.id));
     if (rows.length === 0)
         return [];
+    const mask = scopeMask(docs, {
+        projectDir: opts.projectDir ? normalizeProjectDir(opts.projectDir) : undefined,
+        sourceSessionId: opts.sourceSessionId,
+    });
     const dfById = new Map(kept.map(t => [t.id, t.df]));
     const acc = new Map();
     for (const r of rows) {
+        if (mask && mask[r.doc_id] !== 1)
+            continue;
         const df = dfById.get(r.term_id) ?? 1;
-        const prev = acc.get(r.event_id);
-        const inc = bm25(r.tf, df, stats.n, r.len, stats.avgdl);
-        if (prev)
-            prev.score += inc;
-        else
-            acc.set(r.event_id, { doc: r, score: inc });
+        const inc = bm25(r.tf, df, stats.n, docs.len[r.doc_id] ?? 0, stats.avgdl);
+        acc.set(r.doc_id, (acc.get(r.doc_id) ?? 0) + inc);
     }
+    if (acc.size === 0)
+        return [];
     const limit = Math.min(Math.max(opts.limit ?? 10, 1), 50);
-    const top = [...acc.values()].sort((a, b) => b.score - a.score).slice(0, limit);
-    // Snippets: fetch text for the returned rows only.
-    const ids = top.map(t => t.doc.event_id);
-    const textRows = db.prepare(`SELECT id, text FROM events WHERE id IN (${ids.map(() => '?').join(', ')})`).all(...ids);
-    const textById = new Map(textRows.map(r => [r.id, r.text ?? '']));
+    const top = [...acc.entries()].sort((a, b) => b[1] - a[1]).slice(0, limit);
+    // Metadata and text for the returned rows only.
+    const docIds = top.map(([id]) => id);
+    const marks2 = docIds.map(() => '?').join(', ');
+    const metaRows = db.prepare(`SELECT d.id, d.event_id, d.archive_id, d.source_session_id, d.seq, d.kind, e.text
+       FROM search_docs d JOIN events e ON e.id = d.event_id
+      WHERE d.id IN (${marks2})`).all(...docIds);
+    const metaById = new Map(metaRows.map(r => [r.id, r]));
     const termSet = new Set(kept.map(t => t.term));
-    return top.map(({ doc, score }) => ({
-        eventId: doc.event_id,
-        archiveId: doc.archive_id,
-        sourceSessionId: doc.source_session_id,
-        seq: doc.seq,
-        kind: doc.kind,
-        snippet: buildSnippet(textById.get(doc.event_id) ?? '', termSet),
-        score,
-    }));
+    const out = [];
+    for (const [docId, score] of top) {
+        const m = metaById.get(docId);
+        if (!m)
+            continue;
+        out.push({
+            eventId: m.event_id,
+            archiveId: m.archive_id,
+            sourceSessionId: m.source_session_id,
+            seq: m.seq,
+            kind: m.kind,
+            snippet: buildSnippet(m.text ?? '', termSet),
+            score,
+        });
+    }
+    return out;
+}
+// --- Hybrid retrieval -------------------------------------------------------
+// How deep each ranker is consulted before fusing. Deeper than the returned
+// limit so a result ranked mid-list by one ranker can still be lifted by the
+// other — the entire mechanism by which a paraphrase beats keyword overlap.
+const CANDIDATE_DEPTH = 50;
+/**
+ * Reciprocal-rank fusion constant.
+ *
+ * NOT 60. That value comes from the RRF paper's TREC runs over ~1000-document
+ * lists; at our candidate depth it is actively harmful. With k=60 and d=50 the
+ * whole ranking compresses into a 1.8x spread (rank 1 = 0.0164, rank 50 =
+ * 0.0091), which destroys rank information and lets agreement dominate
+ * quality: two mid-list agreements (0.011 + 0.011) outscore one top-ranked
+ * single-ranker find (0.0164).
+ *
+ * That is fatal for the case this layer exists to serve. A strict paraphrase
+ * shares no vocabulary with its target, so the target is found by the semantic
+ * ranker ALONE and earns exactly one contribution — while any keyword-adjacent
+ * event earns a comparable one for free. The semantic discovery loses to noise.
+ *
+ * Requirement: a rank-1 hit from a single ranker must be able to outrank a pair
+ * of mid-list agreements. 1/(k+1) > 2/(k+d/2) solves to k < d/2 - 2, i.e.
+ * k < 23 at d=50. 10 sits comfortably inside that bound and keeps the top of
+ * the list sharply separated (rank 1 = 0.091, rank 10 = 0.050, rank 50 = 0.017).
+ */
+const RRF_K = 10;
+/**
+ * BM25 contributes only candidates it is actually confident about: anything
+ * below this fraction of its own top score is tail noise that would otherwise
+ * occupy fusion slots and crowd out semantic finds.
+ *
+ * Applied to BM25 only, deliberately. The two rankers have different score
+ * geometries: BM25 decays fast and spans orders of magnitude, so relative
+ * score is meaningful. Cosines sit in a narrow band — measured on the real
+ * corpus, off-topic queries reach 0.40 while on-topic start at 0.45 — so a
+ * relative floor there would be noise itself. Depth is the semantic side's
+ * only honest control.
+ *
+ * Only the tail is affected, so a strong keyword hit can never be dropped and
+ * lexical survival is preserved by construction.
+ */
+const BM25_RELATIVE_FLOOR = 0.25;
+/**
+ * One ranker's contribution for a 0-based rank.
+ *
+ * Exported so the property the fix turns on can be asserted directly rather
+ * than through retrieval, where it is not reliably constructible: with a
+ * bag-of-words embedding model, lexical overlap implies semantic similarity,
+ * so a document that matches a query's words but not its meaning barely
+ * exists. The arithmetic, however, is exact.
+ */
+export function rrfWeight(rank0) {
+    return 1 / (RRF_K + rank0 + 1);
+}
+/**
+ * Fuse ranked id lists by reciprocal rank. Pure, for testing and reuse.
+ *
+ * The property that matters: a top-ranked find from ONE ranker must be able
+ * to outrank a document both rankers place mid-list. That is the case the
+ * semantic layer exists to serve — a strict paraphrase is found by the
+ * semantic ranker alone — and it is exactly what k=60 at depth 50 broke.
+ */
+export function fuseByRank(lists) {
+    const acc = new Map();
+    for (const list of lists) {
+        list.forEach((id, i) => acc.set(id, (acc.get(id) ?? 0) + rrfWeight(i)));
+    }
+    return [...acc.entries()]
+        .map(([id, score]) => ({ id, score }))
+        .sort((a, b) => b.score - a.score);
+}
+/**
+ * BM25 and semantic recall, fused by reciprocal rank.
+ *
+ * RRF over raw score blending is deliberate: BM25 scores are unbounded and
+ * corpus-dependent while cosines sit in [-1, 1], so any weighted sum would be
+ * a hidden tuning knob that drifts with corpus size. Ranks have neither
+ * problem, and lexical survival is measurable (a keyword hit must not be
+ * pushed out of the fused list by semantic noise).
+ *
+ * Degrades silently to BM25 alone when no vectors are stored or the model is
+ * not installed (A2) — the caller cannot tell the difference except by score.
+ */
+export function searchEventsHybrid(query, opts = {}) {
+    const limit = Math.min(Math.max(opts.limit ?? 10, 1), 50);
+    const lexicalAll = searchEvents(query, { ...opts, limit: CANDIDATE_DEPTH });
+    const lexFloor = (lexicalAll[0]?.score ?? 0) * BM25_RELATIVE_FLOOR;
+    const lexical = lexicalAll.filter(h => h.score >= lexFloor);
+    // Cosine ranking has no notion of "no match" — it always returns its nearest
+    // neighbours, however far away they are. A similarity floor cannot fix that
+    // here: measured on the real corpus, off-topic queries reach 0.40 while true
+    // strict paraphrases go as low as 0.21, so the distributions overlap and any
+    // threshold that suppresses noise also cuts the recall this layer exists for.
+    //
+    // What DOES separate them is whether the query has any footing in the corpus
+    // at all. A paraphrase shares no words with its target but still speaks the
+    // corpus's language; `zzz-nonexistent-term` shares nothing with anything. So
+    // the semantic layer is allowed to generalize, never to extrapolate from a
+    // query the corpus has no word of — and "nothing matched" stays expressible.
+    // Unfiltered on the fallback paths: the relative floor exists to stop tail
+    // noise occupying FUSION slots. With no fusion happening it would just be
+    // hiding results the lexical ranker legitimately found.
+    if (!queryHasCorpusTerms(query))
+        return lexicalAll.slice(0, limit);
+    let semantic = [];
+    try {
+        semantic = searchVectors(query, {
+            projectDir: opts.projectDir ? normalizeProjectDir(opts.projectDir) : undefined,
+            sourceSessionId: opts.sourceSessionId,
+            limit: CANDIDATE_DEPTH,
+        });
+    }
+    catch {
+        semantic = []; // a broken model must not take the lexical path down with it
+    }
+    if (semantic.length === 0)
+        return lexicalAll.slice(0, limit);
+    const db = getTranscriptsDb();
+    // Resolve the semantic hits' doc ids to event metadata in one round trip.
+    const docIds = semantic.map(s => s.docId);
+    const docRows = db.prepare(`SELECT id, event_id, archive_id, source_session_id, seq, kind
+       FROM search_docs WHERE id IN (${docIds.map(() => '?').join(', ')})`).all(...docIds);
+    const docById = new Map(docRows.map(r => [r.id, r]));
+    const fused = new Map();
+    lexical.forEach((hit, i) => {
+        fused.set(hit.eventId, { hit, rrf: rrfWeight(i) });
+    });
+    semantic.forEach((s, i) => {
+        const doc = docById.get(s.docId);
+        if (!doc)
+            return;
+        const prev = fused.get(doc.event_id);
+        const inc = rrfWeight(i);
+        if (prev) {
+            prev.rrf += inc;
+            prev.chunkIndex = s.chunkIndex;
+        }
+        else
+            fused.set(doc.event_id, { doc, chunkIndex: s.chunkIndex, rrf: inc });
+    });
+    const ranked = [...fused.entries()]
+        .sort((a, b) => b[1].rrf - a[1].rrf)
+        .slice(0, limit);
+    // Text is needed only for entries that arrived semantically — lexical hits
+    // already carry a snippet built from their matched terms.
+    const needText = ranked.filter(([, e]) => !e.hit).map(([eventId]) => eventId);
+    const textById = new Map();
+    if (needText.length > 0) {
+        const rows = db.prepare(`SELECT id, text FROM events WHERE id IN (${needText.map(() => '?').join(', ')})`).all(...needText);
+        for (const r of rows)
+            textById.set(r.id, r.text ?? '');
+    }
+    const terms = new Set(tokenize(query));
+    return ranked.map(([eventId, e]) => {
+        if (e.hit)
+            return { ...e.hit, score: e.rrf };
+        const doc = e.doc;
+        const text = textById.get(eventId) ?? '';
+        return {
+            eventId,
+            archiveId: doc.archive_id,
+            sourceSessionId: doc.source_session_id,
+            seq: doc.seq,
+            kind: doc.kind,
+            // Aim the excerpt at the window that actually matched, so a long event
+            // does not open on an unrelated first paragraph (C6's opportunity).
+            snippet: buildSnippet(windowAround(text, e.chunkIndex ?? 0), terms),
+            score: e.rrf,
+        };
+    });
+}
+/** True when at least one query token appears in the corpus dictionary. */
+function queryHasCorpusTerms(query) {
+    const terms = [...new Set(tokenize(query))];
+    if (terms.length === 0)
+        return false;
+    const db = getTranscriptsDb();
+    const hit = db.prepare(`SELECT 1 AS present FROM search_terms WHERE term IN (${terms.map(() => '?').join(', ')}) LIMIT 1`).get(...terms);
+    return hit !== undefined;
+}
+/** The slice of `text` the winning chunk covered, widened by one stride each
+ *  way so the snippet has context to work with. */
+function windowAround(text, chunkIndex) {
+    if (text.length <= WINDOW_CHARS)
+        return text;
+    const start = Math.max(0, chunkIndex * STRIDE_CHARS - STRIDE_CHARS);
+    return text.slice(start, start + WINDOW_CHARS + STRIDE_CHARS * 2);
 }
 //# sourceMappingURL=search.js.map