superlocalmemory 3.4.19 → 3.4.22

package/src/superlocalmemory/server/routes/brain.py

@@ -0,0 +1,1234 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory v3.4.22 — LLD-04 §4.1 + §3
+
+"""``/api/v3/brain`` — unified Brain endpoint (LLD-04 v2).
+
+Merges the pre-3.4.22 Patterns / Learning / Behavioral dashboard tabs
+into one auth-gated, honestly-labelled JSON payload. Every metric has
+an ``is_real`` (numeric counters) or ``is_real_ml`` (ML-derived) flag
+and points at the real table it was computed from. No metric is
+fabricated — if a source does not exist we return zero and flag it.
+
+Security primitives consumed from LLD-07 §6:
+
+* ``verify_install_token`` — constant-time token check for auth gate.
+* ``redact_secrets``       — regex+entropy secret scrub.
+
+Deprecated shims (1 release grace) live below the main route. They
+return the historical-ish shape plus ``deprecated: true`` and
+``use_instead: /api/v3/brain``, and require the same auth.
+
+Design notes (LLD-04 §7 hard rules):
+
+* **U2** — every metric section carries ``is_real`` or ``is_real_ml``.
+* **U3** — ``learning.phase`` never returns ``3`` without an active
+  model row AND passing SHA check. Both conditions computed here.
+* **U4** — response shape must not contain the three pre-3.4.22
+  fabricated metrics (24h hit-rate, avg age on hit, skill-evolution
+  row counts). A static test greps this file to verify; therefore we
+  never spell those names anywhere in this module.
+* **U6** — install token required on ``/api/v3/brain`` and all
+  deprecated shim routes. See ``require_install_token`` below.
+* **U10** — feature count surfaced from ``features.FEATURE_DIM``;
+  stratum total from module constant ``_STRATA_TOTAL`` (48 = 4×3×4).
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+import sqlite3
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from typing import Any
+
+from fastapi import APIRouter, Depends, HTTPException, Request
+
+from superlocalmemory.core.security_primitives import (
+    redact_secrets,
+    verify_install_token,
+)
+from superlocalmemory.learning.database import LearningDatabase
+from superlocalmemory.learning.features import FEATURE_DIM
+
+logger = logging.getLogger("superlocalmemory.routes.brain")
+
+router = APIRouter(prefix="/api/v3", tags=["brain"])
+
+
+# ---------------------------------------------------------------------------
+# Constants — surfaced via the response so the UI never hard-codes.
+# ---------------------------------------------------------------------------
+
+# LLD-03 v2 stratum space = 4 query types × 3 entity bins × 4 time buckets.
+_STRATA_TOTAL: int = 48
+
+_VERSION: str = "3.4.22"
+
+# Banned metric names (LLD-04 U4). Kept as a tuple for grep visibility;
+# the source-level test asserts we don't accidentally reintroduce them.
+# NOTE: do NOT add the literal forbidden-key strings here — the U4 grep
+# guard runs over this file.
+
+# Memory directory (home-dir based). Always resolved at call time so that
+# tests can override via monkeypatch on ``_learning_db_path``.
+_MEMORY_DIR_DEFAULT = Path.home() / ".superlocalmemory"
+
+
+def _learning_db_path() -> Path:
+    """Return the path to the learning SQLite DB.
+
+    Separated as a function (not a module constant) so tests can
+    monkeypatch without touching Path.home. See
+    ``tests/test_api/test_brain_endpoint.py``.
+    """
+    return _MEMORY_DIR_DEFAULT / "learning.db"
+
+
+def _memory_db_path() -> Path:
+    return _MEMORY_DIR_DEFAULT / "memory.db"
+
+
+# ---------------------------------------------------------------------------
+# Auth dependency (LLD-04 §4.1, LLD-07 §6.6)
+# ---------------------------------------------------------------------------
+
+
+async def require_install_token(request: Request) -> None:
+    """FastAPI dependency: enforces install-token on Brain routes.
+
+    Accepts either of:
+
+        X-Install-Token: <token>
+        Authorization: Bearer <token>
+
+    Token comparison uses ``verify_install_token`` which wraps
+    ``hmac.compare_digest`` — constant time regardless of input.
+
+    Raises:
+        HTTPException(401) with ``WWW-Authenticate: Install-Token`` so
+        clients know how to retry. Never leaks whether the token file
+        exists, whether the header was missing, or which comparison
+        branch rejected the value.
+    """
+    header_token = request.headers.get("x-install-token")
+    presented: str | None = header_token
+    if not presented:
+        auth = request.headers.get("authorization", "")
+        if auth:
+            # Strip scheme; be tolerant of casing. Empty-after-strip → None.
+            scheme, _, value = auth.partition(" ")
+            if scheme.lower() == "bearer" and value.strip():
+                presented = value.strip()
+    if not presented or not verify_install_token(presented):
+        raise HTTPException(
+            status_code=401,
+            detail="install_token_required",
+            headers={"WWW-Authenticate": "Install-Token"},
+        )
+
+
+# ---------------------------------------------------------------------------
+# Secret-redacted preferences (LLD-04 §4.1, LLD-07 §6.3)
+# ---------------------------------------------------------------------------
+
+
+def redact_secrets_in_preferences(prefs: dict) -> dict:
+    """Deep-copy ``prefs``, scrub every string through ``redact_secrets``,
+    and surface the count of substitutions as ``redacted_count``.
+
+    The input is never mutated — we build a new structure and preserve
+    list/dict shapes exactly. Non-string scalars (ints, floats, bools,
+    None) pass through untouched. Unknown container types are left as-is
+    (best-effort forward compatibility).
+
+    Returns:
+        New dict with the same shape plus ``redacted_count`` at the top
+        level. If ``prefs`` is not a dict, returns
+        ``{"redacted_count": 0}`` (defensive).
+    """
+    if not isinstance(prefs, dict):
+        return {"redacted_count": 0}
+
+    counter = [0]  # mutable holder so the nested closure can increment.
+
+    def _scrub(value: Any) -> Any:
+        if isinstance(value, str):
+            scrubbed = redact_secrets(value)
+            if scrubbed != value:
+                counter[0] += 1
+            return scrubbed
+        if isinstance(value, dict):
+            return {k: _scrub(v) for k, v in value.items()}
+        if isinstance(value, list):
+            return [_scrub(v) for v in value]
+        if isinstance(value, tuple):
+            return tuple(_scrub(v) for v in value)
+        return value
+
+    out: dict = {k: _scrub(v) for k, v in prefs.items()}
+    out["redacted_count"] = counter[0]
+    out.setdefault("is_real", True)
+    out.setdefault("source", prefs.get("source", "_store_patterns"))
+    return out
+
+
+# ---------------------------------------------------------------------------
+# Section builders — each SELECTs from a real table. If a table is empty,
+# returns 0 with ``is_real: true`` (honest). Never invents a number.
+# ---------------------------------------------------------------------------
+
+
+def _load_raw_preferences(profile_id: str) -> dict:
+    """Load preference categories from the behavioral pattern store.
+
+    Returns a dict shaped like::
+
+        {"topics": [...], "entities": [...], "tech": [...], "source": ...}
+
+    with empty lists if the store / DB is not available. Overridden in
+    tests via monkeypatch to keep the Brain route decoupled from the
+    full behavioral pipeline.
+    """
+    out: dict[str, Any] = {
+        "topics": [], "entities": [], "tech": [],
+        "source": "_store_patterns",
+    }
+    db_path = _learning_db_path()
+    if not db_path.exists():
+        return out
+    try:
+        from superlocalmemory.learning.behavioral import BehavioralPatternStore
+        store = BehavioralPatternStore(str(db_path))
+        patterns = store.get_patterns(profile_id=profile_id) or []
+    except Exception as exc:  # pragma: no cover — defensive
+        logger.debug("load_raw_preferences: %s", exc)
+        return out
+
+    topics: list[dict[str, Any]] = []
+    entities: list[dict[str, Any]] = []
+    tech: list[dict[str, Any]] = []
+    for p in patterns:
+        ptype = p.get("pattern_type") or ""
+        meta = p.get("metadata") or {}
+        name = meta.get("value") or p.get("pattern_key") or ""
+        if not name:
+            continue
+        if ptype == "tech_preference":
+            tech.append({
+                "name": str(name),
+                "frequency": float(p.get("confidence", 0.0) or 0.0),
+            })
+        elif ptype in ("entity", "interest"):
+            entities.append({
+                "name": str(name),
+                "mention_count": int(p.get("evidence_count", 0) or 0),
+            })
+        elif ptype == "topic":
+            topics.append({
+                "name": str(name),
+                "strength": float(p.get("confidence", 0.0) or 0.0),
+            })
+    out["topics"] = topics
+    out["entities"] = entities
+    out["tech"] = tech
+    return out
+
+
+def _compute_preferences(profile_id: str) -> dict:
+    raw = _load_raw_preferences(profile_id)
+    return redact_secrets_in_preferences(raw)
+
+
+def _compute_learning_status(profile_id: str,
+                             lrn_db: LearningDatabase) -> dict:
+    """Section ``learning`` — LLD-02 §4.10 phase truth + counters."""
+    signals_total = _safe_count(lrn_db, "learning_signals", profile_id)
+    features_total = _safe_count(lrn_db, "learning_features", profile_id)
+    # Raw count of historic pre-v3.4.22 feedback rows (the source table)
+    legacy_feedback_rows = _safe_count(lrn_db, "learning_feedback", profile_id)
+    # Count of rows actually copied forward into learning_signals by
+    # legacy_migration.migrate_legacy_feedback (signal_type='legacy_feedback').
+    # The difference (raw - migrated) is what the dashboard's
+    # "Migrate legacy data" card surfaces as pending work.
+    legacy_migrated = _count_legacy_migrated(lrn_db, profile_id)
+    signals_last_hour = _count_signals_since(
+        lrn_db, profile_id,
+        datetime.now(timezone.utc) - timedelta(hours=1),
+    )
+
+    active = lrn_db.load_active_model(profile_id)
+    model_active = active is not None
+    model_version: str | None = None
+    model_trained_at: str | None = None
+    model_sha256_present = False
+    if model_active and isinstance(active, dict):
+        model_version = active.get("model_version")
+        model_trained_at = active.get("trained_at")
+        model_sha256_present = bool(active.get("bytes_sha256"))
+
+    phase, phase_label = _resolve_phase(signals_total, model_active,
+                                        model_sha256_present)
+
+    # S9-DASH-03: consult migration_log. If the legacy migration is
+    # marked complete but some rows remain uncopied, those rows are
+    # structurally un-migratable (malformed, duplicate, or missing
+    # required cols). Reporting them as "pending" is misleading and
+    # the dashboard card nags forever. After completion we report
+    # pending=0 so the card auto-hides.
+    migration_complete = _legacy_migration_sentinel_complete(lrn_db)
+    legacy_pending = max(0, legacy_feedback_rows - legacy_migrated)
+    if migration_complete:
+        legacy_pending = 0
+
+    return {
+        "phase": phase,
+        "phase_label": phase_label,
+        "signals_total": signals_total,
+        "signals_last_hour": signals_last_hour,
+        "features_total": features_total,
+        "feature_count_expected": FEATURE_DIM,
+        # Honest split (v3.4.22+): raw historic table count vs rows
+        # actually copied forward via the migrate-legacy flow.
+        "legacy_feedback_rows": legacy_feedback_rows,
+        "legacy_migrated_count": legacy_migrated,
+        "legacy_migration_pending": legacy_pending,
+        "legacy_migration_complete": migration_complete,
+        "model_active": model_active,
+        "model_version": model_version,
+        "model_trained_at": model_trained_at,
+        "model_sha256_present": model_sha256_present,
+        "is_real": True,
+    }
+
+
+def _resolve_phase(signals: int, model_active: bool,
+                   sha_present: bool) -> tuple[int, str]:
+    """LLD-02 §4.10 — phase truth, U3 enforcement.
+
+    Phase 3 requires BOTH ``model_active`` and a non-empty SHA value on
+    the active row. Missing either falls back to phase 2 (≥50 signals)
+    or phase 1 (cold).
+    """
+    if model_active and sha_present and signals >= 200:
+        return 3, "LightGBM ranker active"
+    if signals >= 50:
+        return 2, "Contextual bandit"
+    return 1, "Cold start (cross-encoder only)"
+
+
+def _compute_usage_stats(profile_id: str) -> dict:
+    """Section ``usage`` — real counters from ``learning_signals``.
+
+    All three values are honest aggregations of the last 24 h of signals.
+    ``is_real_ml`` stays False because these are counters, not ML output,
+    but the numbers themselves are pinned to real rows. LLD-04 U2 / §3.1.
+    """
+    lrn_db = _lazy_db_for(profile_id)
+    since = datetime.now(timezone.utc) - timedelta(hours=24)
+    recalls_24h = _count_signals_since(
+        lrn_db, profile_id, since,
+        signal_types=("recall_hit", "recall", "recall_miss",
+                      "candidate", "shown"),
+    )
+    return {
+        "recalls_last_24h": recalls_24h,
+        "top_query_types": _top_query_types(lrn_db, profile_id, since),
+        "top_time_buckets": _top_time_buckets(lrn_db, profile_id, since),
+        "source": "learning_signals_24h_aggregation",
+        "is_real_ml": False,
+        "disclaimer": "Statistical counters, not ML output.",
+    }
+
+
+def _top_query_types(
+    lrn_db: LearningDatabase,
+    profile_id: str,
+    since: datetime,
+    *,
+    limit: int = 5,
+) -> list[dict]:
+    """Return top-N query types from last-24h signals with percentages.
+
+    Query type lives in ``learning_features.features_json`` as one-hot
+    ``query_type_sh`` / ``query_type_mh`` / ``query_type_temp`` /
+    ``query_type_od`` (features.py FEATURE_NAMES). We aggregate by the
+    active one-hot slot. Missing table or zero rows → empty list.
+    """
+    # E.1 (v3.4.22 perf): push the query-type classification into SQL via
+    # json_extract so we don't drag 10k rows' worth of JSON through Python.
+    # SQLite's json1 extension treats json_extract as an ordinary scalar
+    # function, so the whole aggregation becomes a single COUNT(*)
+    # GROUP BY on a computed column. Falls back to the Python path only
+    # if json_extract is unavailable (very old SQLite builds).
+    counts: dict[str, int] = {
+        "single_hop": 0, "multi_hop": 0, "temporal": 0, "open_domain": 0,
+    }
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            agg_rows = conn.execute(
+                "SELECT "
+                "  CASE "
+                "    WHEN CAST(json_extract(f.features_json, '$.query_type_sh')   AS REAL) >= 0.5 THEN 'single_hop' "
+                "    WHEN CAST(json_extract(f.features_json, '$.query_type_mh')   AS REAL) >= 0.5 THEN 'multi_hop' "
+                "    WHEN CAST(json_extract(f.features_json, '$.query_type_temp') AS REAL) >= 0.5 THEN 'temporal' "
+                "    WHEN CAST(json_extract(f.features_json, '$.query_type_od')   AS REAL) >= 0.5 THEN 'open_domain' "
+                "    ELSE NULL END AS qt, "
+                "  COUNT(*) AS n "
+                "FROM learning_features f "
+                "JOIN learning_signals s ON s.id = f.signal_id "
+                "WHERE s.profile_id = ? AND s.created_at >= ? "
+                "  AND f.is_synthetic = 0 "
+                "GROUP BY qt",
+                (profile_id, since.isoformat()),
+            ).fetchall()
+            for row in agg_rows:
+                qt = row["qt"]
+                if qt in counts:
+                    counts[qt] = int(row["n"] or 0)
+        finally:
+            conn.close()
+    except sqlite3.Error:  # pragma: no cover — schema + json1 always present
+        return []
+    total = sum(counts.values())
+    if total == 0:
+        return []
+    ranked = sorted(counts.items(), key=lambda kv: kv[1], reverse=True)
+    return [
+        {"type": name, "pct": round(100.0 * n / total, 1)}
+        for name, n in ranked[:limit] if n > 0
+    ]
+
+
+def _top_time_buckets(
+    lrn_db: LearningDatabase,
+    profile_id: str,
+    since: datetime,
+    *,
+    limit: int = 3,
+) -> list[dict]:
+    """Return top-N hour buckets ("HH:00") from last-24h signals.
+
+    Parses ``created_at`` as ISO timestamp, buckets by hour-of-day, returns
+    top N with percentages. Honest empty when no rows.
+    """
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            rows = conn.execute(
+                "SELECT created_at FROM learning_signals "
+                "WHERE profile_id = ? AND created_at >= ? "
+                "LIMIT 10000",
+                (profile_id, since.isoformat()),
+            ).fetchall()
+        finally:
+            conn.close()
+    except sqlite3.Error:  # pragma: no cover
+        return []
+    buckets: dict[int, int] = {}
+    for r in rows:
+        ts = str(r["created_at"] or "")
+        if not ts:
+            continue
+        try:
+            # ISO: "YYYY-MM-DDTHH:MM:..." — take the "HH" slice.
+            hh = int(ts[11:13])
+        except (ValueError, IndexError):  # pragma: no cover
+            continue
+        buckets[hh] = buckets.get(hh, 0) + 1
+    total = sum(buckets.values())
+    if total == 0:
+        return []
+    ranked = sorted(buckets.items(), key=lambda kv: kv[1], reverse=True)
+    return [
+        {"bucket": f"{hh:02d}:00",
+         "pct": round(100.0 * n / total, 1)}
+        for hh, n in ranked[:limit] if n > 0
+    ]
+
+
+def _compute_bandit_snapshot(profile_id: str,
+                             lrn_db: LearningDatabase) -> dict:
+    """Section ``bandit`` — derive summary from ``ContextualBandit.snapshot``.
+
+    We intentionally compute the *derived* fields here (strata_active,
+    top_arm_global, unsettled_plays) rather than asking the bandit to
+    produce them — this keeps the stratum total constant surfaced from
+    one authoritative place (``_STRATA_TOTAL``) instead of duplicating.
+    """
+    strata_active = 0
+    top_arm_global: dict | None = None
+    unsettled_plays = 0
+    oldest_unsettled_seconds: int | None = None
+
+    try:
+        from superlocalmemory.learning.bandit import ContextualBandit
+        bandit = ContextualBandit(
+            db_path=lrn_db.path,  # shared learning DB
+            profile_id=profile_id,
+        )
+        snap = bandit.snapshot() or {}
+        strata_active = sum(1 for arms in snap.values() if arms)
+        # Global top arm by plays across all strata.
+        best: tuple[str, int] | None = None
+        for stratum_id, arms in snap.items():
+            for arm in arms:
+                plays = int(arm.get("plays", 0) or 0)
+                if best is None or plays > best[1]:
+                    best = (arm["arm_id"], plays)
+        if best is not None:
+            top_arm_global = {"arm_id": best[0], "plays": best[1]}
+        unsettled_plays, oldest_unsettled_seconds = _bandit_unsettled(
+            lrn_db.path, profile_id,
+        )
+    except Exception as exc:  # pragma: no cover — defensive
+        logger.debug("bandit snapshot: %s", exc)
+
+    return {
+        "strata_active": strata_active,
+        "strata_total": _STRATA_TOTAL,
+        "top_arm_global": top_arm_global,
+        "unsettled_plays": unsettled_plays,
+        "oldest_unsettled_seconds": oldest_unsettled_seconds,
+        "is_real": True,
+    }
+
+
+def _bandit_unsettled(db_path: str, profile_id: str) -> tuple[int, int | None]:
+    """Count unsettled bandit_plays rows + age (sec) of oldest one."""
+    try:
+        conn = sqlite3.connect(db_path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            row = conn.execute(
+                "SELECT COUNT(*) AS cnt, MIN(played_at) AS oldest "
+                "FROM bandit_plays "
+                "WHERE profile_id = ? AND settled_at IS NULL",
+                (profile_id,),
+            ).fetchone()
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return 0, None
+    if row is None:  # pragma: no cover — COUNT() always yields a row
+        return 0, None
+    cnt = int(row["cnt"] or 0)
+    oldest_iso = row["oldest"]
+    if not oldest_iso:
+        return cnt, None
+    try:
+        played_at = datetime.fromisoformat(oldest_iso)
+        if played_at.tzinfo is None:
+            played_at = played_at.replace(tzinfo=timezone.utc)
+        age = (datetime.now(timezone.utc) - played_at).total_seconds()
+        return cnt, int(max(age, 0))
+    except (ValueError, TypeError):
+        return cnt, None
+
+
+def _compute_cache_stats() -> dict:
+    """Section ``cache`` — DB file size + row count (if accessible)."""
+    db = _memory_db_path()
+    if not db.exists():
+        return {"db_size_bytes": 0, "entry_count": 0, "is_real": True}
+    size = db.stat().st_size
+    entry_count = 0
+    try:
+        conn = sqlite3.connect(str(db), timeout=5.0)
+        try:
+            row = conn.execute(
+                "SELECT COUNT(*) AS cnt FROM atomic_facts",
+            ).fetchone()
+            entry_count = int(row[0]) if row else 0
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        entry_count = 0
+    return {
+        "db_size_bytes": size,
+        "entry_count": entry_count,
+        "is_real": True,
+    }
+
+
+def _adapter_last_sync_ago(adapter_name: str) -> int | None:
+    """Seconds since adapter's most recent successful sync.
+
+    Reads from ``cross_platform_sync_log`` (LLD-07 M004) in ``memory.db``.
+    Returns ``None`` when the log is absent or has no successful row —
+    honest empty rather than a fabricated number.
+    """
+    try:
+        import sqlite3 as _sqlite3
+        from datetime import datetime as _dt, timezone as _tz
+        from pathlib import Path as _P
+
+        memory_db = _P.home() / ".superlocalmemory" / "memory.db"
+        if not memory_db.exists():
+            return None
+        conn = _sqlite3.connect(
+            f"file:{memory_db}?mode=ro", uri=True, timeout=1.0,
+        )
+        try:
+            cur = conn.execute(
+                "SELECT last_sync_at FROM cross_platform_sync_log "
+                "WHERE adapter_name=? AND success=1 "
+                "ORDER BY last_sync_at DESC LIMIT 1",
+                (adapter_name,),
+            )
+            row = cur.fetchone()
+        finally:
+            conn.close()
+        if row is None or row[0] is None:
+            return None
+        last = _dt.fromisoformat(str(row[0]).replace("Z", "+00:00"))
+        delta = (_dt.now(_tz.utc) - last).total_seconds()
+        return max(0, int(delta))
+    except Exception:  # pragma: no cover — defensive
+        return None
+
+
+def _compute_cross_platform() -> dict:
+    """Section ``cross_platform`` — live status per injection target.
+
+    Each adapter's ``is_active()`` is cheap (env + ``Path.is_dir()``) per
+    LLD-05 A10. Last-sync-at is read from ``cross_platform_sync_log`` in
+    ``memory.db`` (LLD-07 M004). On any adapter error, that adapter
+    reports ``active: false`` with ``reason: error:<ExcName>`` rather
+    than crashing the whole Brain endpoint (LLD-04 §2 — "honest, never
+    fake"). An unimportable adapter means the install is missing Wave 2C
+    components, which is legitimate for an older 3.4.20 → 3.4.22 upgrade
+    mid-migration.
+    """
+    out: dict = {}
+    # S8-SEC-06 fix: use the canonical factory instead of constructing
+    # adapters with no kwargs (all three require scope/base_dir/sync_log_db/
+    # recall_fn). ``build_default_adapters`` returns fully-wired instances
+    # matching what the background sync loop uses, so the panel reflects
+    # the same truth the daemon acts on.
+    try:
+        from superlocalmemory.cli.context_commands import (
+            build_default_adapters as _build_adapters,
+        )
+        _adapters = _build_adapters()
+    except Exception as exc:  # pragma: no cover — defensive
+        _adapters = []
+        logger.debug("brain: adapter factory failed: %s", exc)
+
+    # Summarise by adapter kind: cursor (project+global), antigravity
+    # (workspace+global), copilot. ``adapter.name`` disambiguates scope.
+    _seen: set[str] = set()
+    for a in _adapters:
+        name_attr = getattr(a, "name", "") or ""
+        cls = type(a).__name__
+        kind = None
+        if "Cursor" in cls:
+            kind = "cursor"
+        elif "Antigravity" in cls:
+            kind = "antigravity"
+        elif "Copilot" in cls:
+            kind = "copilot"
+        if kind is None:
+            # Unknown adapter class — skip to avoid polluting the JSON
+            # with an ``out[None]`` key. (S10-SEC-N-01 fix.)
+            continue
+        if kind in _seen:
+            # For cursor/antigravity, which have project+global, the first
+            # active one wins; this surface is a health indicator, not a
+            # per-scope breakdown. (Dev view can drill in later.)
+            continue
+        try:
+            is_active = bool(a.is_active())
+        except Exception as exc:  # pragma: no cover — defensive
+            out[kind] = {"active": False,
+                         "reason": f"error:{exc.__class__.__name__}"}
+            _seen.add(kind)
+            continue
+        out[kind] = {
+            "active": is_active,
+            "last_sync_seconds_ago": _adapter_last_sync_ago(name_attr or kind),
+        }
+        _seen.add(kind)
+    # Fill in missing slots so the shape stays stable.
+    for kind in ("cursor", "antigravity", "copilot"):
+        out.setdefault(kind, {"active": False, "reason": "adapter_unavailable"})
+    # Claude Code hook proxy: active once the install-token exists
+    # (required for prewarm per LLD-01 §4.4 R14).
+    try:
+        from superlocalmemory.core import security_primitives as _sp
+        out["claude_code"] = {
+            "active": _sp._install_token_path().exists(),
+            "hook": "UserPromptSubmit",
+        }
+    except Exception as exc:  # pragma: no cover
+        out["claude_code"] = {"active": False,
+                              "reason": f"error:{exc.__class__.__name__}"}
+    # MCP tool is registered by the unified daemon; if this route is
+    # reachable, the MCP server is up.
+    out["mcp"] = {"active": True, "tool": "mcp__slm"}
+    # CLI is trivially active on any install.
+    out["cli"] = {"active": True}
+    return out
+
+
+def _meta_now() -> dict:
+    return {
+        "generated_at": datetime.now(timezone.utc)
+        .replace(microsecond=0).isoformat()
+        .replace("+00:00", "Z"),
+        "honest_labels": True,
+        "version": _VERSION,
+    }
+
+
+# ---------------------------------------------------------------------------
+# SQL helpers
+# ---------------------------------------------------------------------------
+
+
+def _safe_count(lrn_db: LearningDatabase, table: str,
+                profile_id: str) -> int:
+    """``COUNT(*)`` from ``table`` where ``profile_id`` matches.
+
+    Returns 0 on any error (missing table, DB lock) — never raises
+    from the Brain endpoint.
+    """
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            row = conn.execute(
+                f"SELECT COUNT(*) AS cnt FROM {table} WHERE profile_id = ?",  # nosec - table name is hard-coded above
+                (profile_id,),
+            ).fetchone()
+            return int(row["cnt"]) if row else 0
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return 0
+
+
+def _legacy_migration_sentinel_complete(lrn_db: LearningDatabase) -> bool:
+    """Return True when ``migration_log`` marks the legacy feedback
+    migration as complete (S9-DASH-03).
+
+    Signals that the historic-data card should auto-hide even if a
+    small residual of un-migratable rows lingers in ``learning_feedback``.
+    """
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        try:
+            row = conn.execute(
+                "SELECT status FROM migration_log "
+                "WHERE name = 'LEG001_feedback_to_signals'",
+            ).fetchone()
+            if row is None:
+                return False
+            return str(row[0]).lower() == "complete"
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return False
+
+
+def _count_legacy_migrated(lrn_db: LearningDatabase, profile_id: str) -> int:
+    """Count rows copied forward by the legacy-feedback migration.
+
+    These rows live in ``learning_signals`` with ``signal_type='legacy_feedback'``
+    (see ``learning/legacy_migration.py``). The Brain endpoint exposes this
+    separately from the raw ``learning_feedback`` table count so the UI can
+    show a honest "pending migration" figure instead of the Stage-8 lie
+    that conflated the two.
+    """
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            row = conn.execute(
+                "SELECT COUNT(*) AS cnt FROM learning_signals "
+                "WHERE profile_id = ? AND signal_type = 'legacy_feedback'",
+                (profile_id,),
+            ).fetchone()
+            return int(row["cnt"]) if row else 0
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return 0
+
+
+def _count_signals_since(
+    lrn_db: LearningDatabase,
+    profile_id: str,
+    since: datetime,
+    *,
+    signal_types: tuple[str, ...] | None = None,
+) -> int:
+    """Count ``learning_signals`` rows since ``since`` for ``profile_id``."""
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            if signal_types:
+                placeholders = ",".join("?" * len(signal_types))
+                sql = (
+                    "SELECT COUNT(*) AS cnt FROM learning_signals "
+                    "WHERE profile_id = ? AND created_at >= ? "
+                    f"AND signal_type IN ({placeholders})"
+                )
+                params: tuple[Any, ...] = (
+                    profile_id, since.isoformat(), *signal_types,
+                )
+            else:
+                sql = (
+                    "SELECT COUNT(*) AS cnt FROM learning_signals "
+                    "WHERE profile_id = ? AND created_at >= ?"
+                )
+                params = (profile_id, since.isoformat())
+            row = conn.execute(sql, params).fetchone()
+            return int(row["cnt"]) if row else 0
+        finally:
+            conn.close()
+    except sqlite3.Error:  # pragma: no cover — table always present in tests
+        return 0
+
+
+def _lazy_db_for(profile_id: str) -> LearningDatabase:
+    return LearningDatabase(_learning_db_path())
+
+
+_EVOLUTION_MAX_DAYS = 90
+_EVOLUTION_DEFAULT_DAYS = 30
+
+
+def _compute_evolution_timeseries(
+    profile_id: str,
+    lrn_db: LearningDatabase,
+    *,
+    days: int = _EVOLUTION_DEFAULT_DAYS,
+) -> dict:
+    """Daily learning-signal counts for ``profile_id`` over the last ``days``.
+
+    Returns ``{"days": N, "points": [{"date": "YYYY-MM-DD", "signals": int,
+    "patterns_seen": int}, ...], "is_real": True, "source": "learning_signals"}``.
+
+    Points are left-aligned to midnight UTC and cover exactly ``days``
+    consecutive days including today. Missing days are zero-filled so the
+    chart renders a flat line instead of a gap.
+    """
+    try:
+        requested = int(days) if days is not None else _EVOLUTION_DEFAULT_DAYS
+    except (TypeError, ValueError):
+        requested = _EVOLUTION_DEFAULT_DAYS
+    days = max(1, min(requested, _EVOLUTION_MAX_DAYS))
+    today = datetime.now(timezone.utc).replace(
+        hour=0, minute=0, second=0, microsecond=0,
+    )
+    start = today - timedelta(days=days - 1)
+
+    # Daily counts from learning_signals.
+    counts: dict[str, int] = {}
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            rows = conn.execute(
+                "SELECT substr(created_at, 1, 10) AS d, COUNT(*) AS n "
+                "FROM learning_signals "
+                "WHERE profile_id = ? AND created_at >= ? "
+                "GROUP BY substr(created_at, 1, 10)",
+                (profile_id, start.isoformat()),
+            ).fetchall()
+            counts = {r["d"]: int(r["n"]) for r in rows if r["d"]}
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        counts = {}
+
+    points: list[dict[str, Any]] = []
+    for i in range(days):
+        d = start + timedelta(days=i)
+        key = d.date().isoformat()
+        points.append({
+            "date": key,
+            "signals": counts.get(key, 0),
+        })
+
+    total = sum(p["signals"] for p in points)
+    return {
+        "is_real": True,
+        "source": "learning_signals",
+        "days": days,
+        "total_signals": total,
+        "points": points,
+    }
+
+
+def _action_outcomes_count(lrn_db: LearningDatabase,
+                           profile_id: str) -> int:
+    """Row count in ``action_outcomes`` for ``profile_id``.
+
+    ``action_outcomes`` ships in v3.4.22 (M006). While absent, returns 0.
+    """
+    try:
+        conn = sqlite3.connect(lrn_db.path, timeout=5.0)
+        conn.row_factory = sqlite3.Row
+        try:
+            row = conn.execute(
+                "SELECT COUNT(*) AS cnt FROM action_outcomes "
+                "WHERE profile_id = ?",
+                (profile_id,),
+            ).fetchone()
+            return int(row["cnt"]) if row else 0
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return 0
+
+
+# ---------------------------------------------------------------------------
+# Routes
+# ---------------------------------------------------------------------------
+
+
+@router.get("/brain", dependencies=[Depends(require_install_token)])
+async def get_brain(profile_id: str = "default") -> dict:
+    """Unified Brain endpoint — LLD-04 §3.1.
+
+    Fan-out: each section is a synchronous SQLite reader. Running them
+    serially was a 7-round-trip chain on the hot path (PERF-v2-05).
+    We offload each to the default executor via ``asyncio.to_thread``
+    and ``asyncio.gather`` so the wall-clock is dominated by the slowest
+    single section, not the sum. Every section already swallows its own
+    errors and returns honest-empty on failure, so ``return_exceptions``
+    is unnecessary — but we still set ``return_exceptions=True`` as a
+    belt-and-suspenders guard so one broken reader can't 500 the whole
+    endpoint.
+    """
+    import asyncio
+
+    lrn_db = LearningDatabase(_learning_db_path())
+
+    (
+        preferences, learning, usage, bandit_snap, cache,
+        cross_platform, outcomes_rows, evolution,
+    ) = await asyncio.gather(
+        asyncio.to_thread(_compute_preferences, profile_id),
+        asyncio.to_thread(_compute_learning_status, profile_id, lrn_db),
+        asyncio.to_thread(_compute_usage_stats, profile_id),
+        asyncio.to_thread(_compute_bandit_snapshot, profile_id, lrn_db),
+        asyncio.to_thread(_compute_cache_stats),
+        asyncio.to_thread(_compute_cross_platform),
+        asyncio.to_thread(_action_outcomes_count, lrn_db, profile_id),
+        asyncio.to_thread(
+            _compute_evolution_timeseries, profile_id, lrn_db,
+            days=_EVOLUTION_DEFAULT_DAYS,
+        ),
+        return_exceptions=True,
+    )
+
+    # Replace any per-section exception with an honest-empty dict so the
+    # endpoint never propagates a half-rendered payload.
+    def _ok(value, fallback):
+        if isinstance(value, Exception):
+            return fallback
+        return value
+
+    return {
+        "profile_id": profile_id,
+        "preferences":      _ok(preferences, {"is_real": True,
+                                              "topics": [], "entities": [],
+                                              "tech": [], "redacted_count": 0,
+                                              "source": "_store_patterns"}),
+        "learning":         _ok(learning, {"is_real": True, "phase": 1,
+                                           "signals_total": 0}),
+        "usage":            _ok(usage, {"is_real_ml": False,
+                                        "recalls_last_24h": 0,
+                                        "top_query_types": [],
+                                        "top_time_buckets": []}),
+        "bandit":           _ok(bandit_snap, {"is_real": True,
+                                              "strata_active": 0,
+                                              "strata_total": 48}),
+        "cache":            _ok(cache, {"is_real": True,
+                                        "db_size_bytes": 0,
+                                        "entry_count": 0}),
+        "cross_platform":   _ok(cross_platform, {}),
+        "evolution_preview": _ok(evolution, {
+            "is_real": True, "source": "learning_signals",
+            "days": _EVOLUTION_DEFAULT_DAYS, "total_signals": 0, "points": [],
+        }),
+        "outcomes_preview": {
+            "action_outcomes_rows":
+                0 if isinstance(outcomes_rows, Exception) else outcomes_rows,
+            "ships_in": "3.4.22",
+        },
+        # S9-defer H-22: live tile data for the Reward / Shadow /
+        # Evolution-Cost dashboard tiles. Each block is a honest-empty
+        # default when the underlying table is missing (fresh install
+        # or older schema) so the UI can render "no data yet" without
+        # a 500. All three queries are cheap COUNT / AVG aggregates.
+        "reward_preview": _compute_reward_preview(profile_id),
+        "shadow_preview": _compute_shadow_preview(profile_id),
+        "evolution_cost_preview": _compute_evolution_cost_preview(profile_id),
+        "outcome_queue": _compute_outcome_queue_stats(profile_id),
+        "meta": _meta_now(),
+    }
+
+
+def _compute_outcome_queue_stats(profile_id: str) -> dict:
+    """S9-DASH-02: producer-side telemetry for the Brain panel.
+
+    Exposes the outcome-queue worker counters + a live count of
+    ``pending_outcomes`` so the operator can see the closed loop is
+    actually flowing: recall → enqueue → persist → finalize.
+    """
+    import sqlite3
+    from pathlib import Path
+    try:
+        from superlocalmemory.learning.outcome_queue import (
+            get_counters, queue_size,
+        )
+        counters = get_counters()
+        qsz = queue_size()
+    except Exception:
+        counters, qsz = {}, 0
+    home = Path.home() / ".superlocalmemory"
+    db = home / "memory.db"
+    pending_now = 0
+    if db.exists():
+        try:
+            conn = sqlite3.connect(str(db), timeout=1.0)
+            try:
+                row = conn.execute(
+                    "SELECT COUNT(*) FROM pending_outcomes "
+                    "WHERE profile_id = ? AND status = 'pending'",
+                    (profile_id,),
+                ).fetchone()
+                if row:
+                    pending_now = int(row[0] or 0)
+            finally:
+                conn.close()
+        except Exception:
+            pass
+    return {
+        "is_real": True,
+        "queue_depth": int(qsz),
+        "pending_outcomes_now": pending_now,
+        "counters": counters,
+        "source": "outcome_queue + pending_outcomes",
+    }
+
+
+# ---------------------------------------------------------------------------
+# S9-defer H-22 — dashboard-tile aggregates
+# ---------------------------------------------------------------------------
+
+
+def _compute_reward_preview(profile_id: str) -> dict:
+    """Reward-tile aggregate — count + mean reward over the last 24h."""
+    import sqlite3
+    from pathlib import Path
+    home = Path.home() / ".superlocalmemory"
+    db = home / "memory.db"
+    default = {
+        "is_real": False, "rows_24h": 0, "mean_reward_24h": 0.0,
+        "source": "action_outcomes",
+    }
+    if not db.exists():
+        return default
+    try:
+        conn = sqlite3.connect(str(db), timeout=1.0)
+        try:
+            row = conn.execute(
+                "SELECT COUNT(*) AS c, AVG(reward) AS m "
+                "FROM action_outcomes "
+                "WHERE profile_id = ? AND settled = 1 "
+                "  AND reward IS NOT NULL "
+                "  AND settled_at >= datetime('now', '-1 day')",
+                (profile_id,),
+            ).fetchone()
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return default
+    if row is None:
+        return default
+    return {
+        "is_real": True,
+        "rows_24h": int(row[0] or 0),
+        "mean_reward_24h": float(row[1] or 0.0),
+        "source": "action_outcomes",
+    }
+
+
+def _compute_shadow_preview(profile_id: str) -> dict:
+    """Shadow-tile aggregate — latest candidate + paired-obs count."""
+    import sqlite3
+    default = {
+        "is_real": False, "active_candidate_id": None,
+        "paired_observations": 0, "rollback_count_90d": 0,
+        "source": "learning_model_state+shadow_observations",
+    }
+    learning_db = _learning_db_path()
+    if not learning_db.exists():
+        return default
+    try:
+        conn = sqlite3.connect(str(learning_db), timeout=1.0)
+        try:
+            row = conn.execute(
+                "SELECT id FROM learning_model_state "
+                "WHERE profile_id = ? AND is_candidate = 1 "
+                "LIMIT 1",
+                (profile_id,),
+            ).fetchone()
+            cand_id = int(row[0]) if row and row[0] is not None else None
+            paired = 0
+            if cand_id is not None:
+                try:
+                    c = conn.execute(
+                        "SELECT COUNT(DISTINCT query_id) "
+                        "FROM shadow_observations "
+                        "WHERE candidate_id = ? "
+                        "  AND query_id IN ("
+                        "    SELECT query_id FROM shadow_observations "
+                        "    WHERE candidate_id = ? AND arm = 'active' "
+                        "    INTERSECT "
+                        "    SELECT query_id FROM shadow_observations "
+                        "    WHERE candidate_id = ? AND arm = 'candidate' "
+                        "  )",
+                        (cand_id, cand_id, cand_id),
+                    ).fetchone()
+                    paired = int(c[0]) if c and c[0] is not None else 0
+                except sqlite3.Error:
+                    paired = 0
+            # Rollback count in the last 90d. The rollback log is a
+            # separate table the learning subsystem writes; if it is
+            # absent we simply return 0.
+            try:
+                r2 = conn.execute(
+                    "SELECT COUNT(*) FROM model_state_history "
+                    "WHERE profile_id = ? AND action = 'rollback' "
+                    "  AND ts >= datetime('now', '-90 day')",
+                    (profile_id,),
+                ).fetchone()
+                rollbacks = int(r2[0]) if r2 and r2[0] is not None else 0
+            except sqlite3.Error:
+                rollbacks = 0
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return default
+    return {
+        "is_real": True,
+        "active_candidate_id": cand_id,
+        "paired_observations": paired,
+        "rollback_count_90d": rollbacks,
+        "source": "learning_model_state+shadow_observations",
+    }
+
+
+def _compute_evolution_cost_preview(profile_id: str) -> dict:
+    """Evolution-cost tile — last 7d spend + call count."""
+    import sqlite3
+    default = {
+        "is_real": False, "calls_7d": 0, "cost_usd_7d": 0.0,
+        "tokens_in_7d": 0, "tokens_out_7d": 0,
+        "source": "evolution_llm_cost_log",
+    }
+    learning_db = _learning_db_path()
+    if not learning_db.exists():
+        return default
+    try:
+        conn = sqlite3.connect(str(learning_db), timeout=1.0)
+        try:
+            row = conn.execute(
+                "SELECT COUNT(*) AS c, "
+                "       COALESCE(SUM(cost_usd), 0) AS cost, "
+                "       COALESCE(SUM(tokens_in), 0) AS tin, "
+                "       COALESCE(SUM(tokens_out), 0) AS tout "
+                "FROM evolution_llm_cost_log "
+                "WHERE profile_id = ? "
+                "  AND ts >= datetime('now', '-7 day')",
+                (profile_id,),
+            ).fetchone()
+        finally:
+            conn.close()
+    except sqlite3.Error:
+        return default
+    if row is None:
+        return default
+    return {
+        "is_real": True,
+        "calls_7d": int(row[0] or 0),
+        "cost_usd_7d": float(row[1] or 0.0),
+        "tokens_in_7d": int(row[2] or 0),
+        "tokens_out_7d": int(row[3] or 0),
+        "source": "evolution_llm_cost_log",
+    }
+
+
+@router.get("/brain/evolution-timeseries",
+            dependencies=[Depends(require_install_token)])
+async def get_brain_evolution_timeseries(
+    profile_id: str = "default",
+    days: int = _EVOLUTION_DEFAULT_DAYS,
+) -> dict:
+    """Daily learning-signal counts for ``profile_id`` over the last ``days``.
+
+    ``days`` is clamped to ``[1, 90]`` to keep the response bounded. Each
+    missing day is zero-filled so the chart renders a flat line, not a gap.
+    """
+    import asyncio
+
+    lrn_db = LearningDatabase(_learning_db_path())
+    result = await asyncio.to_thread(
+        _compute_evolution_timeseries, profile_id, lrn_db, days=days,
+    )
+    result["meta"] = _meta_now()
+    return result
+
+
+# ---------------------------------------------------------------------------
+# Deprecated shims — 1 release grace. All require the install token.
+# ---------------------------------------------------------------------------
+
+
+@router.get("/learning/stats",
+            dependencies=[Depends(require_install_token)])
+async def learning_stats_deprecated(profile_id: str = "default") -> dict:
+    lrn_db = LearningDatabase(_learning_db_path())
+    return {
+        "deprecated": True,
+        "use_instead": "/api/v3/brain",
+        "learning": _compute_learning_status(profile_id, lrn_db),
+    }
+
+
+@router.get("/patterns",
+            dependencies=[Depends(require_install_token)])
+async def patterns_deprecated(profile_id: str = "default") -> dict:
+    return {
+        "deprecated": True,
+        "use_instead": "/api/v3/brain",
+        "preferences": _compute_preferences(profile_id),
+    }
+
+
+@router.get("/behavioral",
+            dependencies=[Depends(require_install_token)])
+async def behavioral_deprecated(profile_id: str = "default") -> dict:
+    return {
+        "deprecated": True,
+        "use_instead": "/api/v3/brain",
+        "usage": _compute_usage_stats(profile_id),
+    }
+
+
+__all__ = (
+    "router",
+    "require_install_token",
+    "redact_secrets_in_preferences",
+)