@mindrian_os/cli 2.0.0-beta.3 → 2.0.0-beta.33

package/scripts/rs-engine.py

@@ -0,0 +1,2073 @@
+#!/usr/bin/env python3
+"""
+rs-engine.py -- Reverse Salient Engine (Phase 89)
+==================================================
+
+Authoritative port of the Kwan 2023 reverse-salient algorithm. Runs on a
+single room (Mode A) and produces .rs-engine-results.json with pairs scored
+by abs(semantic - lsa) and classified by signed directionality.
+
+Algorithmic signature (RESEARCH.md Q1, ALGORITHM-SOURCE.md):
+  - Structural similarity: TF-IDF + TruncatedSVD(80) + topic-keyword
+    membership counting + L1 pairwise distance + invert-and-rescale.
+    This is NOT cosine-on-SVD. See lib/core/rs_math.py docstrings.
+  - Semantic similarity: sentence-transformer embeddings (MiniLM local by
+    default; multilingual-e5-large via Pinecone inference when available).
+  - Detection: iterative argmax on |semantic - lsa| with symmetric cleanup.
+  - Classification: signed (semantic - lsa) -> structural_transfer or
+    semantic_implementation.
+
+Embedding model is configurable per RESEARCH.md Q4:
+  - RS_EMBEDDING_MODEL unset (default) -> multilingual-e5-large via Pinecone
+    inference (Plan 89-03 wires the cold path).
+  - RS_EMBEDDING_MODEL=minilm                -> local all-MiniLM-L6-v2.
+  - RS_EMBEDDING_MODEL=bert-large-cased      -> explicit repro mode (heavy).
+
+Mode A (this plan, 89-01): internal single-room scan.
+Mode B/C extend this script in Plans 89-04 / 89-05 (multi-room + external
+corpus + hybrid). The mode dispatch is already wired here so those plans
+only add helpers.
+
+Three-surface usage:
+  CLI:     python scripts/rs-engine.py --mode internal --room ./room
+  Desktop: invoked by ReverseSalientAgent conversational trigger (Plan 89-07).
+  Cowork:  writes symlink into 00_Context/rs-engine-results.json when
+           COWORK=1 is set.
+
+License: BSL-1.1 (see LICENSE at repo root).
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import re
+import sqlite3
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Sequence, Tuple
+
+# Auto-install Python ML deps if missing (v1.10.9, plan 85-10, LAWRENCE-001).
+# rs-engine.py is the direct entry point for /mos:find-bottlenecks (invoked by
+# lib/agents/reverse-salient-agent.cjs); it never had this wiring even though
+# sibling whitespace scripts did (RCA phase-134-python-elimination-false-
+# complete -- Change 1). `requests` is included here (not just numpy) because
+# it is a transitive import in lib/core/rs_corpus.py, loaded a few lines below
+# this block -- and `requests` was the actual root cause of the original
+# Windows tester 2026-05-23 silent-failure class this file's own comments
+# reference (see runRsEngine's diagnostic-forwarding logic in
+# reverse-salient-agent.cjs).
+sys.path.insert(0, str(Path(__file__).resolve().parent / "lib"))
+from ensure_ml_deps import ensure
+ensure(["numpy", "requests"])
+
+# --- Guarded imports ---------------------------------------------------------
+
+try:
+    import numpy as np
+except ImportError:
+    print(
+        "rs-engine requires numpy. Run: pip install -r requirements-hsi.txt",
+        file=sys.stderr,
+    )
+    sys.exit(1)
+
+# Import shared math helpers. rs-engine.py is invoked from the repo root so
+# the "lib.core.rs_math" package path is canonical (plan verify block uses
+# this exact import).
+_THIS_FILE = Path(__file__).resolve()
+_REPO_ROOT = _THIS_FILE.parent.parent
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+from lib.core.rs_math import (  # noqa: E402
+    abs_diff_topk,
+    build_lsa_matrix,
+    classify_direction,
+)
+from lib.core.rs_corpus import (  # noqa: E402
+    fetch_corpus,
+    topic_slug,
+    semantic_gate,
+    SEMANTIC_FLOOR,
+)
+from lib.core.rs_rooms import (  # noqa: E402
+    load_multi_room_corpus,
+    summarize_corpus,
+)
+# Plan 89-05 wires Mode C (hybrid) via rs_hybrid. Import is optional at the
+# top so Mode A / Mode B remain importable when rs_hybrid is missing; the
+# hybrid dispatch in main() raises a clear error if the import failed.
+try:
+    from lib.core.rs_hybrid import (  # noqa: E402
+        build_unified_corpus as _rs_hybrid_build,
+        filter_cross_corpus_pairs as _rs_hybrid_filter,
+    )
+    _RS_HYBRID_AVAILABLE = True
+except Exception as _rs_hybrid_import_err:  # pragma: no cover -- defensive
+    _RS_HYBRID_AVAILABLE = False
+    _rs_hybrid_import_err_msg = str(_rs_hybrid_import_err)
+# Plan 89-03 wires the Pinecone rs-external cache for Mode B. The import is
+# optional so Mode A (internal) remains usable without Pinecone installed --
+# rs_cache has a guarded import of the pinecone SDK.
+try:
+    from lib.core.rs_cache import (  # noqa: E402
+        namespace_slug as _rs_cache_namespace_slug,
+        get_namespace_freshness as _rs_cache_freshness,
+        upsert_corpus as _rs_cache_upsert,
+        fetch_all_from_namespace as _rs_cache_fetch_all,
+        is_fresh as _rs_cache_is_fresh,
+        TTL_DAYS as _RS_CACHE_TTL_DAYS,
+    )
+    # Module reference (not just the named re-exports above) so the SEED-018
+    # semantic gate can call rs_cache._embed_via_bridge directly and build its
+    # topic vector from the SAME local encoder that produced the record
+    # vectors it gates -- see _gate_records_cached below (Plan 296-05).
+    from lib.core import rs_cache
+    _RS_CACHE_AVAILABLE = True
+except Exception as _rs_cache_import_err:  # pragma: no cover -- defensive
+    _RS_CACHE_AVAILABLE = False
+    _rs_cache_import_err_msg = str(_rs_cache_import_err)
+
+
+# --- Constants ---------------------------------------------------------------
+
+DEFAULT_EMBEDDING_MODEL = "multilingual-e5-large"
+# SKIP_FILES / SKIP_DIRS from the ONE shared source (Phase 200-01, SEED-018).
+# Was a local copy that the other walkers drifted from; now single-sourced.
+from lib.core.rs_corpus_exclude import SKIP_DIRS, SKIP_FILES  # noqa: E402
+CACHE_FILENAME = ".rs-engine-cache.json"
+RESULTS_FILENAME = ".rs-engine-results.json"
+DEFAULT_TOPK = 100
+DEFAULT_THRESHOLD = 0.30
+ENGINE_VERSION = "1"
+
+# Mode C (hybrid) post-filter overshoot. After abs-diff top-k over the
+# unified corpus, we discard any intra-corpus pair (room-room or
+# external-external) and keep only cross-corpus pairs. An overshoot of
+# 10x the requested topk gives enough headroom to reach topk on realistic
+# corpora where room artifacts (O(100)) are dwarfed by external docs
+# (O(2000)) and most of the strongest abs-diff pairs would be
+# external-external by volume alone. Smaller overshoots silently shrink
+# delivered pair count below topk; larger overshoots waste O(k log k)
+# sort work. Plan 89-05 Detailed Steps specify 10x.
+HYBRID_OVERSHOOT = 10
+
+
+# --- Artifact discovery ------------------------------------------------------
+
+def parse_frontmatter(content: str) -> Dict[str, str]:
+    fm_match = re.match(r"^---\n([\s\S]*?)\n---", content)
+    if not fm_match:
+        return {}
+    fields = {}
+    for line in fm_match.group(1).split("\n"):
+        if ":" in line:
+            key, _, val = line.partition(":")
+            fields[key.strip()] = val.strip().strip('"').strip("'")
+    return fields
+
+
+def extract_title(content: str, filepath: Path) -> str:
+    match = re.search(r"^# (.+)$", content, re.MULTILINE)
+    if match:
+        return match.group(1).strip()
+    return filepath.stem.replace("-", " ").title()
+
+
+def extract_body(content: str) -> str:
+    fm_match = re.match(r"^---\n[\s\S]*?\n---\n?", content)
+    if fm_match:
+        return content[fm_match.end():]
+    return content
+
+
+def discover_artifacts(room_dir: Path) -> List[Dict]:
+    """Walk room_dir for .md artifacts. Mirrors compute-hsi.py pattern.
+
+    The authoritative source's artifact loader assumed a flat SQLite
+    "artifacts" table. The real MindrianOS schema (lib/core/lazygraph-ops.cjs)
+    keeps artifacts on disk as .md files and tracks them in the nodes table
+    by reference. Plan 89-01 said to read room.db artifacts directly; that
+    table does not exist. We walk the filesystem instead (same pattern as
+    compute-hsi.py) so the port works against real rooms, and cache
+    embeddings in a JSON sidecar below. This is a documented Rule 3 fix.
+    """
+    artifacts: List[Dict] = []
+    room_path = room_dir.resolve()
+
+    for root, dirs, files in os.walk(room_path):
+        dirs[:] = [d for d in dirs if d not in SKIP_DIRS]
+        rel_root = Path(root).relative_to(room_path)
+        if str(rel_root) == ".":
+            # Skip room-root files -- not part of any section.
+            continue
+        section = str(rel_root).split(os.sep)[0]
+        for fname in sorted(files):
+            if not fname.endswith(".md"):
+                continue
+            if fname in SKIP_FILES:
+                continue
+            fpath = Path(root) / fname
+            try:
+                content = fpath.read_text(encoding="utf-8")
+            except (OSError, UnicodeDecodeError):
+                continue
+            body = extract_body(content).strip()
+            if len(body) < 50:
+                continue
+            artifact_id = str(rel_root / Path(fname).stem).replace(os.sep, "/")
+            artifacts.append({
+                "id": artifact_id,
+                "section": section,
+                "title": extract_title(content, fpath),
+                "path": str(fpath.relative_to(room_path)),
+                "text": body,
+            })
+    return artifacts
+
+
+# --- Embedding path ---------------------------------------------------------
+
+def _content_hash(text: str) -> str:
+    return hashlib.sha256(text.encode("utf-8")).hexdigest()[:16]
+
+
+def _load_embedding_cache(room_dir: Path) -> Dict[str, Dict]:
+    cache_path = room_dir / CACHE_FILENAME
+    if not cache_path.exists():
+        return {}
+    try:
+        raw = json.loads(cache_path.read_text(encoding="utf-8"))
+        if not isinstance(raw, dict) or raw.get("version") != ENGINE_VERSION:
+            return {}
+        embeddings = raw.get("embeddings", {})
+        return embeddings if isinstance(embeddings, dict) else {}
+    except (json.JSONDecodeError, OSError):
+        return {}
+
+
+def _save_embedding_cache(
+    room_dir: Path,
+    cache: Dict[str, Dict],
+    model_name: str,
+) -> None:
+    cache_path = room_dir / CACHE_FILENAME
+    payload = {
+        "version": ENGINE_VERSION,
+        "updated_at": datetime.now(timezone.utc).isoformat(),
+        "model": model_name,
+        "embeddings": cache,
+    }
+    try:
+        tmp_path = cache_path.with_suffix(cache_path.suffix + ".tmp")
+        tmp_path.write_text(json.dumps(payload, indent=2), encoding="utf-8")
+        tmp_path.replace(cache_path)
+    except OSError as e:
+        print(f"rs-engine: cache write failed: {e}", file=sys.stderr)
+
+
+def _embed_local_minilm(texts: Sequence[str]) -> Optional[np.ndarray]:
+    try:
+        from sentence_transformers import SentenceTransformer
+    except ImportError:
+        print(
+            "rs-engine: sentence-transformers not installed; "
+            "run `pip install -r requirements-hsi.txt`",
+            file=sys.stderr,
+        )
+        return None
+    model = SentenceTransformer("all-MiniLM-L6-v2")
+    vectors = model.encode(list(texts), show_progress_bar=False)
+    return np.asarray(vectors, dtype=np.float32)
+
+
+def _embed_via_pinecone_inference(texts: Sequence[str]) -> Optional[np.ndarray]:
+    """Cold path: call Pinecone inference API for multilingual-e5-large.
+
+    Plan-check Gap 1 (PLAN-CHECK.md) explicitly documents that this entry is
+    stubbed in Plan 89-01 and completed in Plans 89-03 / 89-05. Raising
+    NotImplementedError keeps the failure loud and localizable if 89-03 is
+    skipped, rather than silently falling back to MiniLM.
+
+    The plan authorizes this stub: PLAN-CHECK Gap 1 resolution says "89-01
+    must stub with NotImplementedError for the cold path, documented as
+    'completed in 89-03 / 89-05.'"
+    """
+    raise NotImplementedError(
+        "Pinecone inference embedding is wired in Plan 89-03 / 89-05. "
+        "For Plan 89-01 (Mode A internal), set RS_EMBEDDING_MODEL=minilm "
+        "or rely on cached embeddings in .rs-engine-cache.json."
+    )
+
+
+def compute_embeddings(
+    artifacts: Sequence[Dict],
+    room_dir: Path,
+) -> Tuple[Optional[np.ndarray], str]:
+    """Return (embedding_matrix, model_name). Uses cache where possible.
+
+    Precedence:
+      1. RS_EMBEDDING_MODEL=minilm OR default-but-Pinecone-unavailable
+         -> local MiniLM.
+      2. RS_EMBEDDING_MODEL=bert-large-cased -> repro path (not implemented
+         in 89-01; deferred to a future plan).
+      3. Default (no env var) AND Pinecone inference wired
+         -> multilingual-e5-large (stub raises; see docstring).
+    """
+    model_env = os.environ.get("RS_EMBEDDING_MODEL", "").strip().lower()
+
+    # Fall back to MiniLM when Pinecone stub cannot serve the cold path.
+    cold_model = "all-MiniLM-L6-v2" if model_env in ("", "minilm") else model_env
+    cache = _load_embedding_cache(room_dir)
+
+    # Resolve cache hits vs misses.
+    missing_indices: List[int] = []
+    dim: Optional[int] = None
+    vectors: List[Optional[np.ndarray]] = [None] * len(artifacts)
+    for idx, art in enumerate(artifacts):
+        digest = _content_hash(art["text"])
+        cache_entry = cache.get(art["id"])
+        if (
+            isinstance(cache_entry, dict)
+            and cache_entry.get("hash") == digest
+            and cache_entry.get("model") == cold_model
+            and isinstance(cache_entry.get("vector"), list)
+        ):
+            vec = np.asarray(cache_entry["vector"], dtype=np.float32)
+            vectors[idx] = vec
+            if dim is None:
+                dim = vec.shape[0]
+        else:
+            missing_indices.append(idx)
+
+    if missing_indices:
+        missing_texts = [artifacts[i]["text"] for i in missing_indices]
+        new_vectors: Optional[np.ndarray] = None
+        model_used = cold_model
+
+        if model_env == "minilm" or model_env == "" and cold_model == "all-MiniLM-L6-v2":
+            # MiniLM is the honest default until 89-03 wires Pinecone.
+            new_vectors = _embed_local_minilm(missing_texts)
+            model_used = "all-MiniLM-L6-v2"
+        elif model_env == DEFAULT_EMBEDDING_MODEL or model_env == "e5-large":
+            # Pinecone inference path. Plan 89-01 stub -> NotImplementedError.
+            new_vectors = _embed_via_pinecone_inference(missing_texts)
+            model_used = DEFAULT_EMBEDDING_MODEL
+        elif model_env == "bert-large-cased":
+            # Repro path -- not wired in 89-01; raise clear error.
+            raise NotImplementedError(
+                "RS_EMBEDDING_MODEL=bert-large-cased is not wired in Plan 89-01. "
+                "Use RS_EMBEDDING_MODEL=minilm for this plan."
+            )
+        else:
+            # Unknown model name -- default to MiniLM with a warning.
+            print(
+                f"rs-engine: unknown RS_EMBEDDING_MODEL={model_env!r}; "
+                "falling back to local MiniLM",
+                file=sys.stderr,
+            )
+            new_vectors = _embed_local_minilm(missing_texts)
+            model_used = "all-MiniLM-L6-v2"
+
+        if new_vectors is None:
+            return None, model_used
+
+        for offset, idx in enumerate(missing_indices):
+            vec = new_vectors[offset]
+            vectors[idx] = vec
+            if dim is None:
+                dim = vec.shape[0]
+            cache[artifacts[idx]["id"]] = {
+                "hash": _content_hash(artifacts[idx]["text"]),
+                "model": model_used,
+                "vector": vec.astype(np.float32).tolist(),
+            }
+        _save_embedding_cache(room_dir, cache, model_used)
+        cold_model = model_used
+
+    matrix = np.vstack([v for v in vectors if v is not None])
+    return matrix, cold_model
+
+
+def semantic_similarity_matrix(embeddings: np.ndarray) -> np.ndarray:
+    """Cosine similarity on row-normalized embedding vectors."""
+    from sklearn.metrics.pairwise import cosine_similarity
+    sim = cosine_similarity(embeddings)
+    return np.clip(sim, 0.0, 1.0).astype(np.float32)
+
+
+# --- SQLite edge writer ------------------------------------------------------
+
+# System provenance defaults for rs-engine's own node writes. Mirrors
+# lib/core/node-insert.cjs's SYSTEM_SOURCE_PATH / SYSTEM_CREATED_BY, using a
+# distinct source_path handle ('system:rs-engine' vs the JS module's
+# 'system:hsi-to-graph') so RS-written nodes stay distinguishable from
+# HSI-written ones per the RCA's Change 1 requirement.
+RS_ENGINE_SOURCE_PATH = "system:rs-engine"
+RS_ENGINE_CREATED_BY = "system"
+
+
+def _now_ms() -> int:
+    """Epoch milliseconds, matching the INTEGER created_at/last_seen_at
+    columns the Phase-109 migration writes (see phase-109-nodes-provenance.cjs,
+    which stores JS Date.now()-shaped epoch-ms integers)."""
+    return int(datetime.now(timezone.utc).timestamp() * 1000)
+
+
+def _nodes_table_is_migrated(conn: sqlite3.Connection) -> bool:
+    """Detect the Phase-109 provenance schema on the nodes table.
+
+    Mirrors lib/core/node-insert.cjs's isMigratedSchema: PRAGMA table_info(nodes)
+    and check for the source_path column. Defensive: any PRAGMA failure falls
+    back to the legacy 3-column shape (the wide insert would throw on a real
+    legacy db, so defaulting to legacy is the safe choice -- same rationale as
+    the JS helper).
+    """
+    try:
+        cols = conn.execute("PRAGMA table_info(nodes)").fetchall()
+        # PRAGMA table_info row shape: (cid, name, type, notnull, dflt_value, pk)
+        return any(row[1] == "source_path" for row in cols)
+    except sqlite3.Error:
+        return False
+
+
+def _upsert_node(
+    conn: sqlite3.Connection,
+    migrated: bool,
+    node_id: str,
+    node_type: str,
+    properties_json: str,
+) -> None:
+    """NOT-NULL-safe node upsert (RCA rs-engine-python-insert-not-null-and-detail-drop-regression).
+
+    Mirrors lib/core/node-insert.cjs's insertNode contract in Python. The bare
+    3-column insert this replaces crashed with
+    `NOT NULL constraint failed: nodes.source_path` on any room.db carrying
+    the Phase-109 wide nodes schema, since node-insert.cjs's Phase 140-01 fix
+    only covers its 4 JS call sites and this script is Python.
+
+    On the migrated (wide) schema, supplies the four NOT NULL provenance
+    columns with rs-engine's own system-write defaults. On the legacy
+    3-column schema, keeps the original bare insert unchanged.
+    """
+    if migrated:
+        now_ms = _now_ms()
+        conn.execute(
+            "INSERT INTO nodes "
+            "(id, type, properties, source_path, created_by, created_at, last_seen_at) "
+            "VALUES (?, ?, ?, ?, ?, ?, ?) "
+            "ON CONFLICT(id) DO UPDATE SET "
+            "properties = excluded.properties, last_seen_at = excluded.last_seen_at",
+            (
+                node_id,
+                node_type,
+                properties_json,
+                RS_ENGINE_SOURCE_PATH,
+                RS_ENGINE_CREATED_BY,
+                now_ms,
+                now_ms,
+            ),
+        )
+        return
+    conn.execute(
+        "INSERT INTO nodes (id, type, properties) VALUES (?, ?, ?) "
+        "ON CONFLICT(id) DO UPDATE SET properties = excluded.properties",
+        (node_id, node_type, properties_json),
+    )
+
+
+def write_reverse_salient_edges(
+    room_dir: Path,
+    pairs: Sequence[Dict],
+) -> int:
+    """Insert REVERSE_SALIENT edges into room.db carrying source='rs-engine'.
+
+    Plan 89-01 said to ALTER the REVERSE_SALIENT table to add a `source`
+    column; REVERSE_SALIENT is an edge TYPE not a table. The real schema
+    (lib/core/lazygraph-ops.cjs) uses a single `edges` table with a
+    flexible `properties` JSON column. Rule 3 adaptation: we namespace our
+    edges by writing `source: 'rs-engine'` inside properties, and we DO NOT
+    delete edges written by hsi-to-graph.cjs (which carry source='hsi'
+    implicitly or via future backfill). Coexistence is the RESEARCH.md Q1
+    decision (b) invariant.
+
+    Returns the number of edges actually inserted.
+    """
+    db_path = room_dir / ".mindrian" / "room.db"
+    if not db_path.exists():
+        # Room has no graph yet -- skip silently; JSON output is still the
+        # authoritative artifact for Plan 89-01.
+        return 0
+
+    conn = sqlite3.connect(str(db_path))
+    try:
+        conn.execute("PRAGMA journal_mode=WAL")
+        # Ensure the edges table exists (rooms initialized outside lazygraph).
+        conn.execute(
+            """
+            CREATE TABLE IF NOT EXISTS edges (
+              source TEXT NOT NULL,
+              target TEXT NOT NULL,
+              type TEXT NOT NULL,
+              properties TEXT DEFAULT '{}',
+              PRIMARY KEY (source, target, type)
+            )
+            """
+        )
+        conn.execute(
+            """
+            CREATE TABLE IF NOT EXISTS nodes (
+              id TEXT PRIMARY KEY,
+              type TEXT NOT NULL,
+              properties TEXT DEFAULT '{}'
+            )
+            """
+        )
+
+        # Cleanup: remove only rs-engine-sourced edges (json_extract safely
+        # returns NULL on missing key, so non-rs-engine edges are preserved).
+        conn.execute(
+            "DELETE FROM edges WHERE type = 'REVERSE_SALIENT' "
+            "AND json_extract(properties, '$.source') = 'rs-engine'"
+        )
+
+        upsert_edge = (
+            "INSERT INTO edges (source, target, type, properties) "
+            "VALUES (?, ?, ?, ?) ON CONFLICT(source, target, type) "
+            "DO UPDATE SET properties = excluded.properties"
+        )
+        # Detect schema ONCE per invocation (D-02a both-schema safety, mirrors
+        # node-insert.cjs's per-call PRAGMA check): a migrated db requires the
+        # wide insert, a legacy 3-col db (created by the CREATE TABLE IF NOT
+        # EXISTS above, on a room with no prior migration) rejects it.
+        migrated = _nodes_table_is_migrated(conn)
+
+        written = 0
+        for pair in pairs:
+            src = pair.get("source_artifact_id")
+            tgt = pair.get("target_artifact_id")
+            if not src or not tgt:
+                continue
+            # Upsert artifact nodes as a courtesy so downstream readers can
+            # resolve them without relying on compute-hsi having run first.
+            _upsert_node(
+                conn,
+                migrated,
+                src,
+                "Artifact",
+                json.dumps({
+                    "title": pair.get("source_title", ""),
+                    "section": pair.get("source_section", ""),
+                }),
+            )
+            _upsert_node(
+                conn,
+                migrated,
+                tgt,
+                "Artifact",
+                json.dumps({
+                    "title": pair.get("target_title", ""),
+                    "section": pair.get("target_section", ""),
+                }),
+            )
+            props = json.dumps({
+                "source": "rs-engine",
+                "lsa_score": pair["lsa_score"],
+                "semantic_score": pair["semantic_score"],
+                "signed_diff": pair["signed_diff"],
+                "abs_diff": pair["abs_diff"],
+                "direction": pair["direction"],
+                "innovation_type": pair["direction"],
+            })
+            conn.execute(upsert_edge, (src, tgt, "REVERSE_SALIENT", props))
+            written += 1
+        conn.commit()
+        return written
+    finally:
+        conn.close()
+
+
+# --- Mode A runner -----------------------------------------------------------
+
+def run_mode_internal(
+    room_dir: Path,
+    topk: int,
+    threshold: float,
+    no_thesis: bool,
+) -> Dict:
+    artifacts = discover_artifacts(room_dir)
+    if len(artifacts) < 2:
+        return {
+            "metadata": {
+                "mode": "internal",
+                "room_dir": str(room_dir),
+                "artifact_count": len(artifacts),
+                "timestamp": datetime.now(timezone.utc).isoformat(),
+                "engine_version": ENGINE_VERSION,
+            },
+            "pairs": [],
+        }
+
+    texts = [f"{a['title']}\n{a['text']}" for a in artifacts]
+
+    # Structural signal: authoritative topic-keyword-membership LSA.
+    lsa_matrix = build_lsa_matrix(texts)
+
+    # Semantic signal: embedding cosine.
+    embeddings, model_used = compute_embeddings(artifacts, room_dir)
+    if embeddings is None:
+        # Fallback: treat semantic as identity so abs_diff collapses to 1-lsa.
+        # This keeps the engine honest when no embedder is available.
+        print(
+            "rs-engine: no embedder available; semantic matrix set to identity",
+            file=sys.stderr,
+        )
+        n = len(artifacts)
+        sem_matrix = np.eye(n, dtype=np.float32)
+    else:
+        sem_matrix = semantic_similarity_matrix(embeddings)
+
+    top_pairs = abs_diff_topk(lsa_matrix, sem_matrix, k=topk)
+
+    pair_dicts: List[Dict] = []
+    for i, j, signed, absv in top_pairs:
+        if absv < threshold:
+            continue
+        a_i, a_j = artifacts[i], artifacts[j]
+        pair_dicts.append({
+            "source_artifact_id": a_i["id"],
+            "source_title": a_i["title"],
+            "source_section": a_i["section"],
+            "target_artifact_id": a_j["id"],
+            "target_title": a_j["title"],
+            "target_section": a_j["section"],
+            "lsa_score": round(float(lsa_matrix[i, j]), 4),
+            "semantic_score": round(float(sem_matrix[i, j]), 4),
+            "signed_diff": round(float(signed), 4),
+            "abs_diff": round(float(absv), 4),
+            "direction": classify_direction(signed),
+        })
+
+    edges_written = write_reverse_salient_edges(room_dir, pair_dicts)
+
+    result = {
+        "metadata": {
+            "mode": "internal",
+            "room_dir": str(room_dir),
+            "artifact_count": len(artifacts),
+            "topk_requested": topk,
+            "threshold": threshold,
+            "embedding_model": model_used,
+            "thesis_generated": False,  # Plan 89-01 always cost-free.
+            "no_thesis": bool(no_thesis),
+            "edges_written": edges_written,
+            "timestamp": datetime.now(timezone.utc).isoformat(),
+            "engine_version": ENGINE_VERSION,
+        },
+        "pairs": pair_dicts,
+    }
+    return result
+
+
+# --- Cross-room (Mode A multi-room) runner -----------------------------------
+
+# Overshoot multiplier for cross-room filtering. After abs-diff top-k, we
+# discard any pair whose two artifacts live in the same room. A modest 3x
+# overshoot keeps the post-filter yield close to the requested topk even when
+# a few of the strongest pairs happen to be intra-room. Larger overshoots
+# pay an O(k log k) cost but are wasted if yield is already >= topk.
+CROSS_ROOM_OVERSHOOT = 3
+CROSS_ROOM_WARN_SHARE = 0.05  # warn when any room contributes < 5% of corpus
+
+
+def run_mode_cross_room(
+    room_paths: Sequence[str],
+    topk: int,
+    threshold: float,
+    no_thesis: bool,
+) -> Dict:
+    """Mode A multi-room: scan N rooms and keep only cross-room pairs.
+
+    Output contract (per plan):
+      - metadata.mode == "cross-room"
+      - metadata.rooms == [room_id, ...] in input order
+      - every pair has source_room != target_room
+      - pairs carry source_room / target_room / source_artifact / target_artifact
+
+    The cross-room summary table is printed to stderr by the caller; this
+    function only computes the payload.
+    """
+    corpus = load_multi_room_corpus(room_paths)
+    room_ids = []
+    seen = set()
+    for art in corpus:
+        rid = art["room_id"]
+        if rid not in seen:
+            room_ids.append(rid)
+            seen.add(rid)
+
+    if len(corpus) < 10:
+        # The authoritative algorithm's LSA fit needs enough documents to
+        # produce stable topic keywords. Below 10, cross-room signal is
+        # unreliable. Plan's canonical threshold.
+        print(
+            f"[rs-engine] Only {len(corpus)} artifacts across "
+            f"{len(room_paths)} rooms. Need >= 10 for cross-room mode.",
+            file=sys.stderr,
+        )
+        return {
+            "metadata": {
+                "mode": "cross-room",
+                "rooms": room_ids,
+                "room_paths": [str(Path(rp).resolve()) for rp in room_paths],
+                "artifact_count": len(corpus),
+                "room_counts": summarize_corpus(corpus),
+                "topk_requested": topk,
+                "threshold": threshold,
+                "no_thesis": bool(no_thesis),
+                "thesis_generated": False,
+                "pair_matrix": {},
+                "timestamp": datetime.now(timezone.utc).isoformat(),
+                "engine_version": ENGINE_VERSION,
+            },
+            "pairs": [],
+        }
+
+    if len({a["room_id"] for a in corpus}) < 2:
+        # All usable artifacts came from a single room; no cross-room pair
+        # is possible by construction. Emit an empty-pairs result with a
+        # clear message rather than silently running the algorithm.
+        print(
+            "[rs-engine] All artifacts come from a single room; "
+            "cross-room filter yields zero pairs by construction.",
+            file=sys.stderr,
+        )
+        return {
+            "metadata": {
+                "mode": "cross-room",
+                "rooms": room_ids,
+                "room_paths": [str(Path(rp).resolve()) for rp in room_paths],
+                "artifact_count": len(corpus),
+                "room_counts": summarize_corpus(corpus),
+                "topk_requested": topk,
+                "threshold": threshold,
+                "no_thesis": bool(no_thesis),
+                "thesis_generated": False,
+                "pair_matrix": {},
+                "timestamp": datetime.now(timezone.utc).isoformat(),
+                "engine_version": ENGINE_VERSION,
+            },
+            "pairs": [],
+        }
+
+    # Share-skew warning (plan Risk 1). Any room contributing < 5% of the
+    # corpus can skew LSA toward the dominant rooms' vocabulary.
+    room_counts = summarize_corpus(corpus)
+    total = len(corpus)
+    for rid, count in room_counts.items():
+        share = count / total if total else 0.0
+        if share < CROSS_ROOM_WARN_SHARE:
+            print(
+                f"[rs-engine] Warning: room {rid!r} contributes "
+                f"{count}/{total} artifacts ({share:.1%}); LSA may skew "
+                f"away from it.",
+                file=sys.stderr,
+            )
+
+    # Build the multi-room corpus as if it were one flat artifact list. The
+    # engine's LSA fit + abs-diff top-k run byte-identical to Mode A; only
+    # the post-filter step is new.
+    texts = [f"{a['title']}\n{a['text']}" for a in corpus]
+
+    lsa_matrix = build_lsa_matrix(texts)
+
+    # Semantic signal: use the same path as Mode A single-room. We embed
+    # into a transient cache in the FIRST room's .mindrian-less root. The
+    # cache lives alongside the cross-room results file so multi-room
+    # embeddings cannot collide with a single-room Mode A cache on the same
+    # first-room path (different artifact-id shapes).
+    first_room_dir = Path(room_paths[0]).resolve()
+    cache_dir = first_room_dir / ".rs-engine-cross-room-cache"
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    # Remap to the compute_embeddings contract (it expects artifact["id"]).
+    # Use global_id so cache keys stay unique across rooms.
+    cache_artifacts = [
+        {"id": a["global_id"], "text": a["text"]}
+        for a in corpus
+    ]
+    embeddings, model_used = compute_embeddings(cache_artifacts, cache_dir)
+    if embeddings is None:
+        print(
+            "rs-engine: no embedder available; semantic matrix set to identity",
+            file=sys.stderr,
+        )
+        n = len(corpus)
+        sem_matrix = np.eye(n, dtype=np.float32)
+    else:
+        sem_matrix = semantic_similarity_matrix(embeddings)
+
+    # Overshoot so post-filter still yields topk when some pairs are intra-room.
+    overshoot_k = max(topk * CROSS_ROOM_OVERSHOOT, topk + 10)
+    top_pairs = abs_diff_topk(lsa_matrix, sem_matrix, k=overshoot_k)
+
+    # Cross-room filter + pair matrix accumulator.
+    pair_dicts: List[Dict] = []
+    pair_matrix: Dict[str, int] = {}
+    for i, j, signed, absv in top_pairs:
+        if absv < threshold:
+            continue
+        a_i, a_j = corpus[i], corpus[j]
+        if a_i["room_id"] == a_j["room_id"]:
+            continue  # intra-room pair -- discard.
+        src_room = a_i["room_id"]
+        tgt_room = a_j["room_id"]
+        # Canonical pair-matrix key: sort room_ids so (A, B) == (B, A) counts.
+        pm_key = " x ".join(sorted([src_room, tgt_room]))
+        pair_matrix[pm_key] = pair_matrix.get(pm_key, 0) + 1
+        pair_dicts.append({
+            "source_artifact_id": a_i["global_id"],
+            "source_artifact": a_i["artifact_id"],
+            "source_room": src_room,
+            "source_title": a_i["title"],
+            "source_section": a_i["section"],
+            "source_path": a_i["path"],
+            "target_artifact_id": a_j["global_id"],
+            "target_artifact": a_j["artifact_id"],
+            "target_room": tgt_room,
+            "target_title": a_j["title"],
+            "target_section": a_j["section"],
+            "target_path": a_j["path"],
+            "lsa_score": round(float(lsa_matrix[i, j]), 4),
+            "semantic_score": round(float(sem_matrix[i, j]), 4),
+            "signed_diff": round(float(signed), 4),
+            "abs_diff": round(float(absv), 4),
+            "direction": classify_direction(signed),
+        })
+        if len(pair_dicts) >= topk:
+            break
+
+    result = {
+        "metadata": {
+            "mode": "cross-room",
+            "rooms": room_ids,
+            "room_paths": [str(Path(rp).resolve()) for rp in room_paths],
+            "artifact_count": len(corpus),
+            "room_counts": room_counts,
+            "topk_requested": topk,
+            "threshold": threshold,
+            "embedding_model": model_used,
+            "thesis_generated": False,
+            "no_thesis": bool(no_thesis),
+            "edges_written": 0,  # Mode A multi-room does NOT write room.db edges.
+            "pair_matrix": pair_matrix,
+            "timestamp": datetime.now(timezone.utc).isoformat(),
+            "engine_version": ENGINE_VERSION,
+        },
+        "pairs": pair_dicts,
+    }
+    return result
+
+
+def _print_cross_room_summary(result: Dict) -> None:
+    """Print the plan-canonical Room-A x Room-B | Bridges table to stderr."""
+    matrix = result.get("metadata", {}).get("pair_matrix") or {}
+    if not matrix:
+        print("[rs-engine] No cross-room bridges found.", file=sys.stderr)
+        return
+    # Sort pair keys by bridge count desc, then alphabetically for stability.
+    rows = sorted(matrix.items(), key=lambda kv: (-kv[1], kv[0]))
+    max_key_len = max(len(k) for k, _ in rows)
+    header_key = "Room pair".ljust(max_key_len)
+    print(f"\n{header_key} | Bridges", file=sys.stderr)
+    print(f"{'-' * max_key_len}-+--------", file=sys.stderr)
+    for key, count in rows:
+        print(f"{key.ljust(max_key_len)} | {count}", file=sys.stderr)
+    print("", file=sys.stderr)
+
+
+# --- Mode B runner -----------------------------------------------------------
+
+def _write_corpus_jsonl(corpus_dir: Path, docs: Sequence[Dict]) -> Path:
+    """Persist the raw fetched corpus for provenance.
+
+    One JSON document per line keeps the file streaming-friendly and easy to
+    diff. Plan 89-03 will key the Pinecone upsert on (source, external_id)
+    read directly from this file.
+    """
+    corpus_dir.mkdir(parents=True, exist_ok=True)
+    path = corpus_dir / "_corpus.jsonl"
+    with path.open("w", encoding="utf-8") as fh:
+        for doc in docs:
+            fh.write(json.dumps(doc, ensure_ascii=False) + "\n")
+    return path
+
+
+def _corpus_docs_to_artifacts(docs: Sequence[Dict]) -> List[Dict]:
+    """Shape fetched docs into the artifact dict the rest of the engine
+    expects (id, section, title, path, text)."""
+    artifacts: List[Dict] = []
+    for doc in docs:
+        external_id = doc.get("external_id") or ""
+        source = doc.get("source") or "external"
+        # Build an artifact id that is stable across runs and unique within
+        # the corpus. External ids can contain URL characters; stripping to
+        # an id-friendly slug keeps downstream SQLite/Pinecone keys predictable.
+        eid_slug = re.sub(r"[^a-zA-Z0-9]+", "-", external_id).strip("-")
+        if not eid_slug:
+            eid_slug = re.sub(r"[^a-zA-Z0-9]+", "-", doc.get("title") or "untitled").strip("-")
+        artifact_id = f"{source}/{eid_slug}"
+        text = (doc.get("abstract") or "").strip()
+        if not text:
+            continue
+        artifacts.append({
+            "id": artifact_id,
+            "section": source,
+            "title": (doc.get("title") or "").strip() or external_id,
+            "path": doc.get("url") or external_id,
+            "text": text,
+            "year": doc.get("year"),
+            "authors": doc.get("authors", []),
+            "doi": doc.get("doi"),
+        })
+    return artifacts
+
+
+def _records_to_artifacts(records: Sequence[Dict]) -> List[Dict]:
+    """Shape rs_cache records (id + values + metadata) into artifact dicts.
+
+    Plan 296-05 warm path: records come straight from the per-room local
+    signal cache (lib.core.rs_cache) and carry the abstract in
+    metadata['abstract'], the public DOI in metadata['doi'], and the human
+    title in metadata['title']. The local encoder's vectors live in
+    record['values']; we attach them so the caller can build the semantic
+    matrix directly without re-embedding.
+    """
+    artifacts: List[Dict] = []
+    for rec in records:
+        rid = rec.get("id") or ""
+        meta = rec.get("metadata", {}) or {}
+        abstract = (meta.get("abstract") or "").strip()
+        if not abstract:
+            # A namespace record without a usable abstract cannot feed LSA;
+            # silently drop, matching the 89-02 empty-abstract filter.
+            continue
+        source = (meta.get("source") or "external").strip() or "external"
+        title = (meta.get("title") or "").strip() or rid
+        # Reconstruct a stable artifact id matching Mode B's cold path shape
+        # (source/slug). This keeps pair ids comparable between warm and
+        # cold runs over the same topic.
+        eid_slug = re.sub(r"[^a-zA-Z0-9]+", "-", rid).strip("-")
+        if not eid_slug:
+            eid_slug = re.sub(r"[^a-zA-Z0-9]+", "-", title or "untitled").strip("-")
+        artifact_id = f"{source}/{eid_slug}"
+        artifacts.append({
+            "id": artifact_id,
+            "section": source,
+            "title": title,
+            "path": f"https://doi.org/{meta['doi']}" if meta.get("doi") else rid,
+            "text": abstract,
+            "year": meta.get("year") or None,
+            "authors": [],
+            "doi": meta.get("doi") or None,
+            "_cached_values": rec.get("values") or [],
+        })
+    return artifacts
+
+
+def _signal_cache_available(room_dir: Optional[Any] = None) -> Tuple[bool, str]:
+    """Return (ok, reason). Reason is empty on success, explanatory on skip.
+
+    Renamed from the prior remote-key gate (Plan 296-05). The former hosted
+    API-key presence check and the RS_EMBEDDING_MODEL=minilm opt-out are both
+    DROPPED here: neither means anything against a per-room local sidecar
+    cache, and leaving the key check in place would make the local path
+    unreachable for the overwhelming majority of users, who have no key. In
+    their place: the room must resolve, since the local cache is scoped to a
+    room by construction (lib.core.rs_cache, closes SEED-029 finding F8).
+    """
+    if not _RS_CACHE_AVAILABLE:
+        return False, f"rs_cache import failed: {globals().get('_rs_cache_import_err_msg', 'unknown')}"
+    if not room_dir:
+        return False, "no room scope for the signal cache"
+    return True, ""
+
+
+def _build_sem_matrix_from_records(artifacts: Sequence[Dict]) -> Optional[np.ndarray]:
+    """Build the pairwise cosine similarity matrix from cached embeddings.
+
+    Returns None if any artifact is missing a vector (triggers fallback to
+    local embedding path). All vectors must share dimensionality.
+    """
+    vectors: List[List[float]] = []
+    for a in artifacts:
+        v = a.get("_cached_values") or []
+        if not v:
+            return None
+        vectors.append(list(v))
+    if not vectors:
+        return None
+    dim = len(vectors[0])
+    if any(len(v) != dim for v in vectors):
+        return None
+    matrix = np.asarray(vectors, dtype=np.float32)
+    return semantic_similarity_matrix(matrix)
+
+
+# --- SEED-018 H2 semantic-floor gate: LIVE production wiring ----------------
+#
+# The gate itself lives in lib/core/rs_corpus.semantic_gate (implemented + unit
+# tested in Plan 200-01). Until this wiring it had NO production caller -- it was
+# dormant. The helpers below call it on the two live Mode B corpus paths:
+#
+#   1. Local signal-cache (warm/cold) path: gate the raw `records` (each carrying
+#      its local-encoder vector in record['values']) against a topic vector from
+#      the SAME encoder (rs_cache._embed_via_bridge), BEFORE _records_to_artifacts.
+#      Plan 296-05: rerouted off Pinecone inference -- after the repoint, records
+#      carry local 384-dim vectors, so a topic vector from a different encoder
+#      would either produce a dimension mismatch or a meaningless cosine, and in
+#      both cases the gate would stop filtering without raising.
+#   2. Local fallback path (Plan 89-02): gate the fetched `docs` using the reused
+#      local MiniLM encoder (_embed_local_minilm), batched so the model loads once.
+#
+# Safe-by-default (this is the core scoring input):
+#   - No-op when the topic vector cannot be obtained (encode / inference absent).
+#   - Reuses SEMANTIC_FLOOR (default 0.15, tunable via RS_SEMANTIC_FLOOR).
+#   - STARVATION GUARD: never drops the usable count below the differential
+#     minimum (the engine early-outs under 2 artifacts). Worst case a no-op,
+#     never a regression.
+#   - Reuses the EXISTING encoders only (Pinecone inference + MiniLM). No new
+#     encoder, no new npm/pip package, no new egress class.
+
+# The engine early-outs when fewer than this many artifacts survive (see the
+# `len(artifacts) < 2` guard in run_mode_external). The gate must never push the
+# corpus below it.
+_DIFFERENTIAL_MIN_ARTIFACTS = 2
+
+
+def _semantic_cosine(a: Sequence[float], b: Sequence[float]) -> float:
+    """Minimal cosine similarity for the starvation-guard ranking. Returns 0.0
+    on shape mismatch or a zero-norm vector."""
+    if not a or not b or len(a) != len(b):
+        return 0.0
+    dot = 0.0
+    na = 0.0
+    nb = 0.0
+    for x, y in zip(a, b):
+        dot += x * y
+        na += x * x
+        nb += y * y
+    if na == 0.0 or nb == 0.0:
+        return 0.0
+    return dot / ((na ** 0.5) * (nb ** 0.5))
+
+
+def _gate_with_starvation_guard(
+    topic_vec: Optional[Sequence[float]],
+    candidates: List[Dict],
+    encode_fn,
+    floor: Optional[float] = None,
+    min_keep: int = _DIFFERENTIAL_MIN_ARTIFACTS,
+    label: str = "corpus",
+) -> List[Dict]:
+    """Apply the H2 semantic-floor gate with a mandatory starvation guard.
+
+    Delegates the actual floor filtering to rs_corpus.semantic_gate (the single
+    gate implementation) so both live paths share one tested code path. When the
+    floor would push the survivor count below the differential minimum on an
+    otherwise-sufficient corpus, keep the top-N candidates by similarity instead
+    of starving score(), and log the skip to stderr.
+    """
+    if not topic_vec or not callable(encode_fn):
+        # No usable topic vector / encoder -> safe no-op pass-through.
+        return list(candidates)
+
+    original_n = len(candidates)
+    # Production caller of the SEED-018 H2 gate (rs_corpus.semantic_gate).
+    kept = semantic_gate(topic_vec, candidates, encode_fn, floor)
+
+    if len(kept) >= min_keep or original_n < min_keep:
+        return kept
+
+    # Starvation guard: the floor would collapse the corpus below the
+    # differential minimum. Preserve the top-N most-similar candidates so the
+    # differential still runs. Worst case the gate is a no-op, never a
+    # regression that starves the scorer.
+    ranked = sorted(
+        candidates,
+        key=lambda c: _semantic_cosine(topic_vec, encode_fn(c)),
+        reverse=True,
+    )
+    keep_n = max(min_keep, len(kept))
+    guarded = ranked[:keep_n]
+    print(
+        f"[rs-engine] semantic gate skipped: would starve corpus "
+        f"(kept {len(guarded)})",
+        file=sys.stderr,
+    )
+    return guarded
+
+
+def _embed_topic_via_signal_cache(topic: str) -> Optional[List[float]]:
+    """Embed `topic` through rs_cache's local encoder (Plan 296-05).
+
+    Reuses the SAME local encoder (rs_cache._embed_via_bridge, which wraps
+    scripts/rs-vector-bridge.cjs -> lib/core/eureka/embedding-spine.cjs) that
+    produced the record vectors this topic vector is compared against. This
+    is the load-bearing part of the reroute: after the repoint, records carry
+    local 384-dim vectors, so a topic vector from ANY other encoder either
+    produces a dimension mismatch or a meaningless cosine -- and in both
+    cases the gate stops filtering off-topic documents without raising.
+
+    Returns a plain float list, or None on ANY failure (rs_cache unavailable,
+    bridge unreachable, unexpected shape) so the caller degrades to a safe
+    no-op. Adds no new encoder and no new egress class -- the bridge is the
+    same local encoder the rest of this repoint already uses.
+    """
+    if not _RS_CACHE_AVAILABLE:
+        return None
+    try:
+        vectors, provenance = rs_cache._embed_via_bridge([topic])
+    except Exception as e:  # pragma: no cover -- defensive
+        print(
+            f"[rs-engine] semantic gate: topic embed unavailable ({e}); "
+            f"gate no-op for this run.",
+            file=sys.stderr,
+        )
+        return None
+    if not vectors or provenance is None:
+        return None
+    return [float(x) for x in vectors[0]]
+
+
+def _gate_records_cached(topic: str, records: List[Dict]) -> List[Dict]:
+    """Gate local signal-cache `records` by the semantic floor before
+    _records_to_artifacts.
+
+    Renamed from the prior remote-cache gate (Plan 296-05). Each record carries
+    its local-encoder vector in record['values']; the topic vector now comes
+    from the SAME encoder via _embed_topic_via_signal_cache, not a different
+    one. No-op (returns records unchanged) when the topic vector is
+    unavailable.
+    """
+    if not records:
+        return records
+    topic_vec = _embed_topic_via_signal_cache(topic)
+    if not topic_vec:
+        return records
+    return _gate_with_starvation_guard(
+        topic_vec,
+        records,
+        lambda rec: rec.get("values") or [],
+        label="signal-cache-records",
+    )
+
+
+def _gate_docs_local(topic: str, docs: List[Dict]) -> List[Dict]:
+    """Gate local-fallback `docs` by the semantic floor using the reused local
+    MiniLM encoder. Batched so SentenceTransformer loads once. Clean no-op when
+    sentence-transformers is absent (encode returns None)."""
+    if not docs:
+        return docs
+    topic_arr = _embed_local_minilm([topic])
+    if topic_arr is None:
+        # sentence-transformers not installed -> safe no-op (Tri-Polar degrade).
+        return docs
+    texts = [
+        f"{d.get('title', '')}\n{d.get('abstract', '')}".strip() for d in docs
+    ]
+    cand_arr = _embed_local_minilm(texts)
+    if cand_arr is None or len(cand_arr) != len(docs):
+        return docs
+    topic_vec = [float(x) for x in topic_arr[0]]
+    vec_by_id = {id(d): [float(x) for x in cand_arr[i]] for i, d in enumerate(docs)}
+    return _gate_with_starvation_guard(
+        topic_vec,
+        docs,
+        lambda d: vec_by_id.get(id(d), []),
+        label="local-fallback",
+    )
+
+
+def run_mode_external(
+    room_dir: Path,
+    topic: str,
+    topk: int,
+    threshold: float,
+    no_thesis: bool,
+) -> Tuple[Dict, Path]:
+    """Mode B: fetch external corpus (cached in the per-room local signal
+    cache), run LSA + abs_diff.
+
+    Plan 296-05 wiring: lib.core.rs_cache's per-room local sidecar caches
+    embedded corpora per topic, scoped to room_dir. On each call:
+      1. Compute namespace for topic (rs_cache.namespace_slug).
+      2. Check freshness, scoped to this room. If < 30 days, skip fetch and
+         use the cached vectors directly (warm path).
+      3. Else, fetch_corpus -> upsert_corpus -> fetch_all_from_namespace
+         (cold path; embeds through the local encoder bridge, then re-reads
+         so the vectors used for scoring came from the cache, not a
+         from-memory list).
+      4. Fall back to the Plan 89-02 local path (fetch_corpus + local
+         MiniLM embedding) when the signal cache is unavailable (rs_cache
+         import failed, or the room does not resolve).
+
+    Returns (result dict, corpus_dir) -- CLI writes results JSON inside
+    corpus_dir at {room}/research/{topic-slug}/.rs-engine-results.json.
+    """
+    slug = topic_slug(topic)
+    if not slug:
+        raise ValueError("topic slug resolved to empty string")
+    corpus_dir = room_dir / "research" / slug
+    corpus_dir.mkdir(parents=True, exist_ok=True)
+
+    # Overshoot target_n so dedup + threshold filtering can still hit topk.
+    target_n = max(topk * 20, topk * 2)
+
+    # Freshness + cache decision.
+    signal_ok, signal_skip_reason = _signal_cache_available(room_dir=room_dir)
+    cache_mode = "bypass"
+    cache_age_days: Optional[float] = None
+    cache_namespace: Optional[str] = None
+    records: List[Dict] = []
+    docs: List[Dict] = []
+
+    if signal_ok:
+        ns = _rs_cache_namespace_slug(topic)
+        cache_namespace = ns
+        try:
+            cache_age_days = _rs_cache_freshness(ns, room_dir=room_dir)
+        except Exception as e:
+            print(f"rs-engine: signal cache freshness check failed ({e})", file=sys.stderr)
+            cache_age_days = None
+
+        if _rs_cache_is_fresh(cache_age_days, ttl_days=_RS_CACHE_TTL_DAYS):
+            # WARM path: namespace is within TTL. Skip fetch.
+            cache_mode = "warm"
+            print(
+                f"[rs-engine] Cache hit (age={cache_age_days:.1f} days, "
+                f"ttl={_RS_CACHE_TTL_DAYS}). Using local signal cache {ns}.",
+                file=sys.stderr,
+            )
+            try:
+                records = _rs_cache_fetch_all(ns, limit=target_n, room_dir=room_dir)
+            except Exception as e:
+                print(
+                    f"rs-engine: warm-path fetch failed ({e}); "
+                    f"falling back to cold path.",
+                    file=sys.stderr,
+                )
+                records = []
+                cache_mode = "bypass"
+
+        if cache_mode != "warm":
+            # COLD path: fetch from external APIs, upsert to the local
+            # signal cache, then re-fetch so the vectors used for scoring
+            # come from the cache's own round-trip (not a from-memory list).
+            age_label = "unset" if cache_age_days is None else f"{cache_age_days:.1f} days"
+            print(
+                f"[rs-engine] Cache miss (age={age_label}). "
+                f"Fetching corpus for topic {topic!r}...",
+                file=sys.stderr,
+            )
+            docs = fetch_corpus(topic, target_n=target_n)
+            if len(docs) >= 2:
+                try:
+                    _rs_cache_upsert(topic, docs, room_dir=room_dir)
+                    cache_mode = "cold"
+                    # Re-read from the local signal cache so the semantic
+                    # matrix uses the cached (round-tripped) local vectors.
+                    records = _rs_cache_fetch_all(ns, limit=target_n, room_dir=room_dir)
+                except Exception as e:
+                    print(
+                        f"rs-engine: local signal cache upsert/fetch failed ({e}); "
+                        f"falling back to local embedding path.",
+                        file=sys.stderr,
+                    )
+                    records = []
+                    cache_mode = "bypass"
+    else:
+        # Signal cache path not available -- preserve Plan 89-02 behavior.
+        print(
+            f"[rs-engine] Signal cache skipped ({signal_skip_reason}); "
+            f"using Plan 89-02 local embedding path.",
+            file=sys.stderr,
+        )
+
+    # Provenance sidecar: keep _corpus.jsonl in both paths.
+    # Warm path has no fresh fetch; docs stays empty but we still want the
+    # sidecar to exist (stale from prior cold run). If absent, skip quietly
+    # so the warm path does not accidentally wipe prior provenance.
+    if docs:
+        corpus_path = _write_corpus_jsonl(corpus_dir, docs)
+    else:
+        corpus_path = corpus_dir / "_corpus.jsonl"
+        if not corpus_path.exists():
+            # Warm path with no prior sidecar (first time the namespace is
+            # warm on this machine). Create an empty file so downstream
+            # readers do not trip on a missing path.
+            corpus_path.write_text("", encoding="utf-8")
+
+    # Build artifacts + texts depending on path taken.
+    if records:
+        # SEED-018 H2 LIVE-WIRE (local signal-cache path): drop off-topic records
+        # using the local-encoder vectors they already carry vs a topic vector
+        # from the SAME encoder, BEFORE they become artifacts and reach the
+        # differential. Starvation-guarded + no-op when the topic vector is
+        # unavailable.
+        records = _gate_records_cached(topic, records)
+        artifacts = _records_to_artifacts(records)
+    else:
+        # Fallback to Plan 89-02 behavior.
+        if not docs:
+            docs = fetch_corpus(topic, target_n=target_n)
+            corpus_path = _write_corpus_jsonl(corpus_dir, docs)
+        # SEED-018 H2 LIVE-WIRE (local fallback): gate off-topic docs with the
+        # reused local MiniLM encoder before they reach the differential.
+        docs = _gate_docs_local(topic, docs)
+        artifacts = _corpus_docs_to_artifacts(docs)
+
+    # Early out if the usable-artifact count collapses.
+    if len(artifacts) < 2:
+        return (
+            {
+                "metadata": {
+                    "mode": "external",
+                    "topic": topic,
+                    "topic_slug": slug,
+                    "room_dir": str(room_dir),
+                    "corpus_dir": str(corpus_dir),
+                    "corpus_path": str(corpus_path),
+                    "corpus_size": len(docs) if docs else len(records),
+                    "artifact_count": len(artifacts),
+                    "tier_counts": _count_by_source(docs) if docs else {},
+                    "cache_mode": cache_mode,
+                    "cache_age_days": round(cache_age_days, 2) if cache_age_days is not None else None,
+                    "cache_namespace": cache_namespace,
+                    "cache_ttl_days": _RS_CACHE_TTL_DAYS if signal_ok else None,
+                    "topk_requested": topk,
+                    "threshold": threshold,
+                    "no_thesis": bool(no_thesis),
+                    "thesis_generated": False,
+                    "timestamp": datetime.now(timezone.utc).isoformat(),
+                    "engine_version": ENGINE_VERSION,
+                },
+                "pairs": [],
+            },
+            corpus_dir,
+        )
+
+    texts = [f"{a['title']}\n{a['text']}" for a in artifacts]
+
+    # Structural signal (authoritative topic-keyword-membership LSA).
+    lsa_matrix = build_lsa_matrix(texts)
+
+    # Semantic signal. Warm/cold path: vectors come straight from Pinecone.
+    # Bypass path: fall back to Plan 89-02's local embedder.
+    sem_matrix: Optional[np.ndarray] = None
+    model_used: str
+    if cache_mode in ("warm", "cold") and records:
+        sem_matrix = _build_sem_matrix_from_records(artifacts)
+        if sem_matrix is not None:
+            model_used = "multilingual-e5-large"
+        else:
+            # Vectors missing -- rare but defensible. Drop to local path.
+            print(
+                "rs-engine: cached records missing values; "
+                "falling back to local embedding for semantic matrix.",
+                file=sys.stderr,
+            )
+            embeddings, model_used = compute_embeddings(artifacts, corpus_dir)
+            sem_matrix = (
+                semantic_similarity_matrix(embeddings)
+                if embeddings is not None
+                else np.eye(len(artifacts), dtype=np.float32)
+            )
+    else:
+        embeddings, model_used = compute_embeddings(artifacts, corpus_dir)
+        if embeddings is None:
+            print(
+                "rs-engine: no embedder available for Mode B; "
+                "semantic matrix set to identity.",
+                file=sys.stderr,
+            )
+            sem_matrix = np.eye(len(artifacts), dtype=np.float32)
+        else:
+            sem_matrix = semantic_similarity_matrix(embeddings)
+
+    top_pairs = abs_diff_topk(lsa_matrix, sem_matrix, k=topk)
+
+    pair_dicts: List[Dict] = []
+    for i, j, signed, absv in top_pairs:
+        if absv < threshold:
+            continue
+        a_i, a_j = artifacts[i], artifacts[j]
+        pair_dicts.append({
+            "source_artifact_id": a_i["id"],
+            "source_title": a_i["title"],
+            "source_section": a_i["section"],
+            "source_url": a_i.get("path", ""),
+            "source_doi": a_i.get("doi"),
+            "target_artifact_id": a_j["id"],
+            "target_title": a_j["title"],
+            "target_section": a_j["section"],
+            "target_url": a_j.get("path", ""),
+            "target_doi": a_j.get("doi"),
+            "lsa_score": round(float(lsa_matrix[i, j]), 4),
+            "semantic_score": round(float(sem_matrix[i, j]), 4),
+            "signed_diff": round(float(signed), 4),
+            "abs_diff": round(float(absv), 4),
+            "direction": classify_direction(signed),
+        })
+
+    result = {
+        "metadata": {
+            "mode": "external",
+            "topic": topic,
+            "topic_slug": slug,
+            "room_dir": str(room_dir),
+            "corpus_dir": str(corpus_dir),
+            "corpus_path": str(corpus_path),
+            "corpus_size": len(docs) if docs else len(records),
+            "artifact_count": len(artifacts),
+            "tier_counts": _count_by_source(docs) if docs else {},
+            "cache_mode": cache_mode,
+            "cache_age_days": round(cache_age_days, 2) if cache_age_days is not None else None,
+            "cache_namespace": cache_namespace,
+            "cache_ttl_days": _RS_CACHE_TTL_DAYS if signal_ok else None,
+            "topk_requested": topk,
+            "threshold": threshold,
+            "embedding_model": model_used,
+            "thesis_generated": False,
+            "no_thesis": bool(no_thesis),
+            "edges_written": 0,  # Mode B does NOT write room.db edges.
+            "timestamp": datetime.now(timezone.utc).isoformat(),
+            "engine_version": ENGINE_VERSION,
+        },
+        "pairs": pair_dicts,
+    }
+    return result, corpus_dir
+
+
+def _count_by_source(docs: Sequence[Dict]) -> Dict[str, int]:
+    counts: Dict[str, int] = {}
+    for d in docs:
+        src = d.get("source") or "unknown"
+        counts[src] = counts.get(src, 0) + 1
+    return counts
+
+
+# --- Mode C (hybrid) runner --------------------------------------------------
+
+# Design note on semantic similarity for the unified corpus (Plan 89-05,
+# updated Plan 296-05): the external side used to carry cached 1024-dim
+# multilingual-e5-large vectors from Pinecone, which would have been a
+# dimensional bug (plan Risk 1) if mixed directly with the room side's
+# 384-dim local MiniLM vectors. Plan 296-05's repoint of lib.core.rs_cache
+# onto the local signal cache CLOSES that hazard by construction: the
+# external side now also carries local 384-dim vectors from the same
+# encoder as the room side. The safe path below is UNCHANGED regardless --
+# embed the entire unified corpus in a single model space (local MiniLM via
+# compute_embeddings()) so the pairwise cosine is well-defined -- because
+# doing so also lets a run with a cold, empty, or unavailable signal cache
+# degrade the same way it always has, rather than depending on the cache
+# having already run.
+
+
+def run_mode_hybrid(
+    room_dir: Path,
+    topic: str,
+    topk: int,
+    threshold: float,
+    no_thesis: bool,
+    external_target: int = 2000,
+) -> Tuple[Dict, Path]:
+    """Mode C: unified room + external corpus with cross-corpus filter.
+
+    The killer feature of Phase 89: rediscovers external-domain concepts
+    that have a structural match in the user's room but never entered the
+    room's vocabulary, and vice versa. Consumed downstream by the Plan
+    89-06 bridge-writer for Obsidian rendering.
+
+    Returns (result dict, corpus_dir) -- CLI writes results JSON inside
+    corpus_dir at {room}/research/{topic-slug}/.rs-engine-results.json.
+
+    Cross-corpus contract:
+      - metadata.mode == "hybrid"
+      - metadata.hybrid == True (explicit flag per plan must_haves[2])
+      - every pair has origin mismatch: one room side, one external side
+      - pair_dicts carry both room_artifact_id and external_doc fields
+      - direction is the standard Mode A / Mode B signed classification
+
+    Edge cases:
+      - rs_hybrid import failed: raise NotImplementedError early so the
+        CLI surfaces a clear "cannot run Mode C" message.
+      - Empty room corpus: return empty pairs with metadata preserved so
+        downstream readers do not trip on missing keys.
+      - Empty external corpus (rare -- cold fetch returned nothing):
+        same empty-pairs result. Caller sees metadata.external_count=0.
+      - Unified corpus < 10 artifacts: skip LSA (same threshold the
+        cross-room path uses). LSA topic keywords are unstable below
+        this size.
+    """
+    if not _RS_HYBRID_AVAILABLE:
+        raise NotImplementedError(
+            "rs-engine: --mode hybrid requires lib.core.rs_hybrid. "
+            f"Import failed: {globals().get('_rs_hybrid_import_err_msg', 'unknown')}"
+        )
+
+    # Signal-cache availability check (mirrors run_mode_external's gate,
+    # Plan 296-05). Not blocking: build_unified_corpus's own
+    # _load_external_records degrades to an empty external side
+    # (cache_mode="bypass") on its own when the signal cache cannot be
+    # reached, via the room_dir it now threads through. This call exists so
+    # the reason is narrated up front instead of only discoverable later in
+    # hybrid_meta.
+    signal_ok, signal_skip_reason = _signal_cache_available(room_dir=str(room_dir))
+    if not signal_ok:
+        print(
+            f"[rs-engine hybrid] Signal cache skipped ({signal_skip_reason}); "
+            f"external side of the unified corpus will be empty this run.",
+            file=sys.stderr,
+        )
+
+    # Build the unified corpus + origin mask via rs_hybrid.
+    corpus, origin_mask, hybrid_meta = _rs_hybrid_build(
+        str(room_dir),
+        topic,
+        external_target=external_target,
+    )
+    slug = hybrid_meta.get("topic_slug") or topic_slug(topic)
+    if not slug:
+        raise ValueError("topic slug resolved to empty string")
+    corpus_dir = room_dir / "research" / slug
+    corpus_dir.mkdir(parents=True, exist_ok=True)
+
+    room_count = int(hybrid_meta.get("room_count", 0))
+    external_count = int(hybrid_meta.get("external_count", 0))
+
+    def _empty_result(reason: str) -> Tuple[Dict, Path]:
+        print(f"[rs-engine hybrid] {reason}", file=sys.stderr)
+        return (
+            {
+                "metadata": {
+                    "mode": "hybrid",
+                    "hybrid": True,
+                    "topic": topic,
+                    "topic_slug": slug,
+                    "room_dir": str(room_dir),
+                    "corpus_dir": str(corpus_dir),
+                    "room_id": hybrid_meta.get("room_id"),
+                    "room_count": room_count,
+                    "external_count": external_count,
+                    "artifact_count": room_count + external_count,
+                    "cache_mode": hybrid_meta.get("cache_mode"),
+                    "cache_age_days": hybrid_meta.get("cache_age_days"),
+                    "cache_namespace": hybrid_meta.get("cache_namespace"),
+                    "cache_ttl_days": hybrid_meta.get("cache_ttl_days"),
+                    "topk_requested": topk,
+                    "threshold": threshold,
+                    "no_thesis": bool(no_thesis),
+                    "thesis_generated": False,
+                    "edges_written": 0,
+                    "timestamp": datetime.now(timezone.utc).isoformat(),
+                    "engine_version": ENGINE_VERSION,
+                    "skip_reason": reason,
+                },
+                "pairs": [],
+            },
+            corpus_dir,
+        )
+
+    # Edge: no room artifacts -- cross-corpus filter yields zero pairs by
+    # construction; save the LSA compute.
+    if room_count == 0:
+        return _empty_result(
+            "Room has no indexable .md artifacts; cross-corpus filter "
+            "yields zero pairs by construction."
+        )
+    # Edge: no external corpus.
+    if external_count == 0:
+        return _empty_result(
+            f"External corpus for topic {topic!r} is empty (cache_mode="
+            f"{hybrid_meta.get('cache_mode')}); cross-corpus filter "
+            f"yields zero pairs by construction."
+        )
+    # Edge: combined corpus too small for stable LSA topic-keyword fit.
+    if len(corpus) < 10:
+        return _empty_result(
+            f"Unified corpus has only {len(corpus)} artifacts; need >= 10 "
+            f"for stable LSA fit."
+        )
+
+    # Structural signal over the unified vocabulary.
+    texts = [f"{a.get('title','')}\n{a['text']}" for a in corpus]
+    lsa_matrix = build_lsa_matrix(texts)
+
+    # Semantic signal: embed the full unified corpus in one model space so
+    # the sem_matrix is dimensionally homogeneous. Room and external share
+    # the SAME model by construction (Plan Risk 1 mitigation). We reuse
+    # compute_embeddings (MiniLM-or-cache) and scope the cache to the
+    # research/{slug}/ directory to avoid polluting Mode A's room-root
+    # .rs-engine-cache.json -- different corpus shape, different keys.
+    cache_artifacts = [
+        {"id": c["global_id"], "text": t}
+        for c, t in zip(corpus, texts)
+    ]
+    embeddings, model_used = compute_embeddings(cache_artifacts, corpus_dir)
+    if embeddings is None:
+        print(
+            "rs-engine hybrid: no embedder available; "
+            "semantic matrix set to identity.",
+            file=sys.stderr,
+        )
+        n = len(corpus)
+        sem_matrix = np.eye(n, dtype=np.float32)
+    else:
+        sem_matrix = semantic_similarity_matrix(embeddings)
+
+    # Overshoot top-k, then post-filter to cross-corpus only.
+    overshoot_k = max(topk * HYBRID_OVERSHOOT, topk + 10)
+    raw_pairs = abs_diff_topk(lsa_matrix, sem_matrix, k=overshoot_k)
+
+    cross_pairs = _rs_hybrid_filter(raw_pairs, origin_mask, topk)
+
+    # Apply threshold + build full pair dicts with room_artifact / external_doc.
+    pair_dicts: List[Dict] = []
+    for cp in cross_pairs:
+        absv = cp["abs_diff"]
+        if absv < threshold:
+            continue
+        i, j = cp["i"], cp["j"]
+        room_idx = cp["room_side"]
+        ext_idx = cp["external_side"]
+        room_art = corpus[room_idx]
+        ext_doc = corpus[ext_idx]
+        signed = cp["signed_diff"]
+
+        # Preserve the standard Mode A / Mode B fields (source_*/target_*)
+        # so the bridge-writer resolvePairIdentity picks them up without
+        # per-mode branches. Canonical orientation: source = room side,
+        # target = external side. This maps the signed-direction
+        # interpretation from plan must_haves[3]:
+        #   structural_transfer (signed > 0):
+        #       "your work has the vocabulary, world has the concept"
+        #   semantic_implementation (signed < 0):
+        #       "your concept exists in the world's vocabulary"
+        pair_dicts.append({
+            "source_artifact_id": room_art["global_id"],
+            "source_artifact": room_art.get("artifact_id") or room_art["global_id"],
+            "source_room": room_art.get("room_id"),
+            "source_section": room_art.get("section"),
+            "source_title": room_art.get("title", ""),
+            "source_path": room_art.get("path", ""),
+            "target_artifact_id": ext_doc["global_id"],
+            "target_external_id": ext_doc.get("external_id") or ext_doc["global_id"],
+            "target_doi": ext_doc.get("doi"),
+            "target_source": ext_doc.get("source"),
+            "target_title": ext_doc.get("title", ""),
+            "target_url": ext_doc.get("path", ""),
+            "target_year": ext_doc.get("year"),
+            # Rich cross-corpus annotations consumed by Plan 89-06 bridge
+            # writer and Plan 89-07 release dashboard.
+            "room_artifact": {
+                "global_id": room_art["global_id"],
+                "artifact_id": room_art.get("artifact_id"),
+                "room_id": room_art.get("room_id"),
+                "section": room_art.get("section"),
+                "title": room_art.get("title", ""),
+                "path": room_art.get("path", ""),
+            },
+            "external_doc": {
+                "global_id": ext_doc["global_id"],
+                "external_id": ext_doc.get("external_id"),
+                "source": ext_doc.get("source"),
+                "title": ext_doc.get("title", ""),
+                "doi": ext_doc.get("doi"),
+                "year": ext_doc.get("year"),
+                "url": ext_doc.get("path", ""),
+            },
+            "lsa_score": round(float(lsa_matrix[i, j]), 4),
+            "semantic_score": round(float(sem_matrix[i, j]), 4),
+            "signed_diff": round(float(signed), 4),
+            "abs_diff": round(float(absv), 4),
+            "direction": classify_direction(signed),
+        })
+
+    result = {
+        "metadata": {
+            "mode": "hybrid",
+            "hybrid": True,
+            "topic": topic,
+            "topic_slug": slug,
+            "room_dir": str(room_dir),
+            "corpus_dir": str(corpus_dir),
+            "room_id": hybrid_meta.get("room_id"),
+            "room_count": room_count,
+            "external_count": external_count,
+            "artifact_count": len(corpus),
+            "cache_mode": hybrid_meta.get("cache_mode"),
+            "cache_age_days": hybrid_meta.get("cache_age_days"),
+            "cache_namespace": hybrid_meta.get("cache_namespace"),
+            "cache_ttl_days": hybrid_meta.get("cache_ttl_days"),
+            "topk_requested": topk,
+            "threshold": threshold,
+            "embedding_model": model_used,
+            "no_thesis": bool(no_thesis),
+            "thesis_generated": False,
+            # Mode C does NOT write REVERSE_SALIENT edges into room.db.
+            # Plan 89-06 bridge-writer produces the user-visible artifact
+            # surface; Plan 89-07 wires the optional room.db edge with
+            # properties.source='rs-engine-hybrid' alongside the agent.
+            "edges_written": 0,
+            "hybrid_overshoot": HYBRID_OVERSHOOT,
+            "external_target": external_target,
+            "timestamp": datetime.now(timezone.utc).isoformat(),
+            "engine_version": ENGINE_VERSION,
+        },
+        "pairs": pair_dicts,
+    }
+    return result, corpus_dir
+
+
+# --- Cowork tri-polar surface ------------------------------------------------
+
+def _write_cowork_symlink(room_dir: Path, results_path: Path) -> None:
+    """When COWORK=1, expose results under 00_Context/ for team visibility."""
+    if os.environ.get("COWORK") != "1":
+        return
+    cowork_dir = room_dir / "00_Context"
+    if not cowork_dir.is_dir():
+        return
+    symlink_path = cowork_dir / "rs-engine-results.json"
+    try:
+        if symlink_path.is_symlink() or symlink_path.exists():
+            symlink_path.unlink()
+        os.symlink(results_path, symlink_path)
+    except OSError as e:
+        print(f"rs-engine: cowork symlink skipped: {e}", file=sys.stderr)
+
+
+# --- CLI ---------------------------------------------------------------------
+
+def _write_results(room_dir: Path, result: Dict, output: Optional[str]) -> Path:
+    path = Path(output).resolve() if output else (room_dir / RESULTS_FILENAME)
+    path.write_text(json.dumps(result, indent=2), encoding="utf-8")
+    return path
+
+
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(
+        prog="rs-engine",
+        description=(
+            "Reverse Salient Engine (Phase 89). Faithfully ports the Kwan 2023 "
+            "topic-keyword-membership LSA + signed abs-diff detection "
+            "algorithm. Mode A runs on a single room; Mode B fetches an "
+            "external research corpus (OpenAlex, arXiv, Tavily); Mode C "
+            "(hybrid) unifies the room with an external corpus and keeps only "
+            "cross-corpus pairs -- the killer feature of Phase 89."
+        ),
+        epilog=(
+            "Three-surface usage:\n"
+            "  CLI     python scripts/rs-engine.py --mode internal --room ./room\n"
+            "          python scripts/rs-engine.py --rooms ~/rooms/a ~/rooms/b ~/rooms/c --topk 50\n"
+            "          python scripts/rs-engine.py --mode external --topic 'quantum biology' --room ./room\n"
+            "          python scripts/rs-engine.py --mode hybrid --topic 'quantum biology' --room ./room --topk 30\n"
+            "  Desktop invoked by ReverseSalientAgent conversational trigger\n"
+            "          (Plan 89-07).\n"
+            "  Cowork  set COWORK=1 to symlink results under 00_Context/.\n"
+        ),
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--mode",
+        choices=["internal", "external", "hybrid"],
+        default="internal",
+        help=(
+            "internal=Mode A (single room, Plan 89-01); "
+            "external=Mode B (research corpus, Plan 89-02); "
+            "hybrid=Mode C (room x external unified matrix, Plan 89-05)."
+        ),
+    )
+    parser.add_argument("--room", required=False, help="Path to room directory (Mode A / B / C).")
+    parser.add_argument(
+        "--rooms",
+        nargs="+",
+        required=False,
+        help=(
+            "Two or more room paths for cross-room Mode A (Plan 89-04). "
+            "Rediscovers bridges between user projects. Mutually exclusive "
+            "with --room. Results land in the FIRST room's .rs-engine-results.json "
+            "with metadata.mode='cross-room'."
+        ),
+    )
+    parser.add_argument(
+        "--topic",
+        required=False,
+        help=(
+            "Free-text research topic (required for --mode external). "
+            "Normalized to a kebab-case slug via lib.core.rs_corpus.topic_slug "
+            "for consistent {room}/research/{topic-slug}/ paths."
+        ),
+    )
+    parser.add_argument(
+        "--topk",
+        type=int,
+        default=DEFAULT_TOPK,
+        help=f"Maximum number of pairs to keep (default {DEFAULT_TOPK}).",
+    )
+    parser.add_argument(
+        "--threshold",
+        type=float,
+        default=DEFAULT_THRESHOLD,
+        help=(
+            f"Minimum abs(semantic - lsa) to keep a pair (default {DEFAULT_THRESHOLD}). "
+            "0.0 disables filtering."
+        ),
+    )
+    parser.add_argument(
+        "--no-thesis",
+        action="store_true",
+        help=(
+            "Skip LLM innovation-thesis generation (RESEARCH Q6). Plan 89-01 always "
+            "runs cost-free; this flag is accepted for compatibility with later plans."
+        ),
+    )
+    parser.add_argument(
+        "--external-target",
+        type=int,
+        default=2000,
+        help=(
+            "Max external docs to include in --mode hybrid's unified corpus "
+            "(default 2000; clamped to rs_hybrid.MAX_EXTERNAL_TARGET). "
+            "Ignored for Mode A / Mode B."
+        ),
+    )
+    parser.add_argument(
+        "--output",
+        default=None,
+        help=f"Output JSON path (default: {{room}}/{RESULTS_FILENAME}).",
+    )
+    return parser
+
+
+def main(argv: Optional[Sequence[str]] = None) -> int:
+    parser = build_parser()
+    args = parser.parse_args(argv)
+
+    # Cross-room Mode A multi-room (Plan 89-04). --rooms is mutually
+    # exclusive with --room. When --rooms is set, --mode is ignored
+    # (cross-room is implicitly "internal multi-room").
+    if args.rooms:
+        if args.room:
+            print(
+                "rs-engine: --rooms and --room are mutually exclusive. "
+                "Pick one.",
+                file=sys.stderr,
+            )
+            return 2
+        if len(args.rooms) < 2:
+            print(
+                "rs-engine: --rooms requires at least two room paths for "
+                "cross-room mode.",
+                file=sys.stderr,
+            )
+            return 2
+        if args.mode == "external":
+            print(
+                "rs-engine: --mode external is incompatible with --rooms. "
+                "Use --room for external mode.",
+                file=sys.stderr,
+            )
+            return 2
+        if args.mode == "hybrid":
+            print(
+                "rs-engine: --mode hybrid is incompatible with --rooms. "
+                "Hybrid unifies ONE room with an external corpus; use --room "
+                "for hybrid mode.",
+                file=sys.stderr,
+            )
+            return 2
+
+        topk = max(1, args.topk)
+        threshold = max(0.0, args.threshold)
+
+        try:
+            result = run_mode_cross_room(
+                args.rooms,
+                topk=topk,
+                threshold=threshold,
+                no_thesis=args.no_thesis,
+            )
+        except NotImplementedError as e:
+            print(f"rs-engine: {e}", file=sys.stderr)
+            return 3
+        except ValueError as e:
+            print(f"rs-engine: {e}", file=sys.stderr)
+            return 2
+        except Exception as e:  # pragma: no cover -- defensive envelope
+            print(f"rs-engine: error: {e}", file=sys.stderr)
+            return 1
+
+        # Output lands in the FIRST room's .rs-engine-results.json by
+        # default (per plan). --output overrides if supplied.
+        first_room_dir = Path(args.rooms[0]).resolve()
+        if not first_room_dir.is_dir():
+            print(
+                f"rs-engine: first room {first_room_dir} is not a directory; "
+                f"cannot write results.",
+                file=sys.stderr,
+            )
+            return 1
+        output_path = _write_results(first_room_dir, result, args.output)
+
+        _write_cowork_symlink(first_room_dir, output_path)
+        _print_cross_room_summary(result)
+
+        meta = result["metadata"]
+        print(
+            f"rs-engine: {len(result['pairs'])} cross-room pairs from "
+            f"{meta['artifact_count']} artifacts across "
+            f"{len(meta['rooms'])} rooms written to {output_path} "
+            f"(model={meta.get('embedding_model', 'n/a')})",
+            file=sys.stderr,
+        )
+        return 0
+
+    if not args.room:
+        print(
+            f"rs-engine: --room is required for --mode {args.mode} "
+            f"(or pass --rooms for cross-room mode)",
+            file=sys.stderr,
+        )
+        return 2
+
+    room_dir = Path(args.room).resolve()
+    # Allow the room directory to be auto-created for Mode B when it does
+    # not exist yet. Mode A still requires an existing room. This matches
+    # the plan's end-to-end smoke which targets /tmp/test-room. Hybrid
+    # Mode C requires an existing room (its whole point is room x external).
+    if args.mode == "external" and not room_dir.exists():
+        try:
+            room_dir.mkdir(parents=True, exist_ok=True)
+        except OSError as e:
+            print(f"rs-engine: could not create {room_dir}: {e}", file=sys.stderr)
+            return 1
+    if not room_dir.is_dir():
+        print(f"rs-engine: {room_dir} is not a directory", file=sys.stderr)
+        return 1
+
+    topk = max(1, args.topk)
+    threshold = max(0.0, args.threshold)
+
+    try:
+        if args.mode == "internal":
+            result = run_mode_internal(
+                room_dir,
+                topk=topk,
+                threshold=threshold,
+                no_thesis=args.no_thesis,
+            )
+            output_path = _write_results(room_dir, result, args.output)
+        elif args.mode == "hybrid":
+            if not args.topic:
+                print(
+                    "rs-engine: --topic is required for --mode hybrid",
+                    file=sys.stderr,
+                )
+                return 2
+            result, corpus_dir = run_mode_hybrid(
+                room_dir,
+                topic=args.topic,
+                topk=topk,
+                threshold=threshold,
+                no_thesis=args.no_thesis,
+                external_target=args.external_target,
+            )
+            default_path = corpus_dir / RESULTS_FILENAME
+            if args.output:
+                output_path = Path(args.output).resolve()
+                output_path.parent.mkdir(parents=True, exist_ok=True)
+                output_path.write_text(
+                    json.dumps(result, indent=2), encoding="utf-8",
+                )
+            else:
+                default_path.parent.mkdir(parents=True, exist_ok=True)
+                default_path.write_text(
+                    json.dumps(result, indent=2), encoding="utf-8",
+                )
+                output_path = default_path
+        else:  # args.mode == "external"
+            if not args.topic:
+                print(
+                    "rs-engine: --topic is required for --mode external",
+                    file=sys.stderr,
+                )
+                return 2
+            result, corpus_dir = run_mode_external(
+                room_dir,
+                topic=args.topic,
+                topk=topk,
+                threshold=threshold,
+                no_thesis=args.no_thesis,
+            )
+            # Default external output: {room}/research/{topic-slug}/.rs-engine-results.json
+            default_path = corpus_dir / RESULTS_FILENAME
+            if args.output:
+                output_path = Path(args.output).resolve()
+                output_path.parent.mkdir(parents=True, exist_ok=True)
+                output_path.write_text(
+                    json.dumps(result, indent=2), encoding="utf-8",
+                )
+            else:
+                default_path.parent.mkdir(parents=True, exist_ok=True)
+                default_path.write_text(
+                    json.dumps(result, indent=2), encoding="utf-8",
+                )
+                output_path = default_path
+    except NotImplementedError as e:
+        print(f"rs-engine: {e}", file=sys.stderr)
+        return 3
+    except ValueError as e:
+        print(f"rs-engine: {e}", file=sys.stderr)
+        return 2
+    except Exception as e:  # pragma: no cover - defensive envelope
+        print(f"rs-engine: error: {e}", file=sys.stderr)
+        return 1
+
+    _write_cowork_symlink(room_dir, output_path)
+
+    meta = result["metadata"]
+    if args.mode == "internal":
+        print(
+            f"rs-engine: {len(result['pairs'])}/{meta['artifact_count']} pairs written to "
+            f"{output_path} (model={meta['embedding_model']}, edges={meta['edges_written']})",
+            file=sys.stderr,
+        )
+    elif args.mode == "hybrid":
+        print(
+            f"rs-engine: {len(result['pairs'])} cross-corpus pairs from "
+            f"{meta.get('room_count', 0)} room + {meta.get('external_count', 0)} "
+            f"external = {meta.get('artifact_count', 0)} artifacts "
+            f"(cache={meta.get('cache_mode', 'n/a')}) written to {output_path} "
+            f"(model={meta.get('embedding_model', 'n/a')})",
+            file=sys.stderr,
+        )
+    else:
+        tiers = meta.get("tier_counts", {})
+        tier_str = ", ".join(f"{k}={v}" for k, v in tiers.items()) or "empty"
+        print(
+            f"rs-engine: {len(result['pairs'])} pairs from "
+            f"{meta.get('artifact_count', 0)} artifacts (corpus={meta.get('corpus_size', 0)}, "
+            f"{tier_str}) written to {output_path} "
+            f"(model={meta.get('embedding_model', 'n/a')})",
+            file=sys.stderr,
+        )
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())