superlocalmemory 3.4.19 → 3.4.22

package/src/superlocalmemory/learning/consolidation_worker.py

@@ -2,15 +2,24 @@
 # Licensed under AGPL-3.0-or-later - see LICENSE file
 # Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
 
-"""Sleep-Time Consolidation Worker — background memory maintenance.
+"""Sleep-Time Consolidation Worker — **back-compat shim**.
 
-Runs periodically (every 6 hours or on-demand) to:
-1. Decay confidence on unused facts (floor 0.1)
-2. Deduplicate near-identical facts
-3. Auto-retrain the adaptive ranker when signal threshold is met
-4. Report consolidation stats
+As of v3.4.22 (F4.A Stage-8 H-01 fix), the 1344-LOC god-module was
+split into five cohesive files:
 
-Inspired by: Letta's sleep-time compute, neuroscience memory consolidation.
+  - ``consolidation_cycle.py``     — :class:`ConsolidationWorker`.
+  - ``pattern_miner.py``           — :func:`generate_patterns`.
+  - ``ranker_retrain_legacy.py``   — deprecated cold-start path.
+  - ``ranker_retrain_online.py``   — LLD-10 candidate seams.
+  - ``ranker_common.py``           — training-matrix + NDCG helpers.
+
+This shim exists so that the 3830 live tests + dashboard + MCP tools +
+managed server routes keep importing from
+``superlocalmemory.learning.consolidation_worker`` with zero churn.
+
+Tests that ``monkeypatch.setattr(cw_mod, "_train_booster", fake)`` work
+unchanged because :func:`_run_shadow_cycle` is defined here and resolves
+its helper names through *this* module's globals.
 
 Part of Qualixar | Author: Varun Pratap Bhardwaj
 """
@@ -19,521 +28,180 @@ from __future__ import annotations
 
 import logging
 import sqlite3
-from datetime import datetime, timezone
-from pathlib import Path
 
 logger = logging.getLogger(__name__)
 
 
-class ConsolidationWorker:
-    """Background memory maintenance worker.
-
-    Call `run()` periodically or via dashboard button.
-    All operations are safe — they improve quality without losing data.
+# ---------------------------------------------------------------------------
+# Re-exports — anything tests or production code reached for on the old
+# module stays reachable here under the same name.
+# ---------------------------------------------------------------------------
+
+from superlocalmemory.learning.consolidation_cycle import (  # noqa: E402
+    ConsolidationWorker,
+)
+from superlocalmemory.learning.ranker_common import (  # noqa: E402
+    _build_training_matrix,
+    _compute_eval_metrics,
+    _shadow_test_improved,
+)
+from superlocalmemory.learning.ranker_retrain_legacy import (  # noqa: E402
+    _retrain_ranker_impl,
+    _LEGACY_RETRAIN_DEPRECATED,
+)
+from superlocalmemory.learning.ranker_retrain_online import (  # noqa: E402
+    RETRAIN_HOURS_THRESHOLD,
+    RETRAIN_HYPERPARAM_CAPS,
+    RETRAIN_MODEL_SIZE_BYTES_CAP,
+    RETRAIN_NEW_OUTCOMES_THRESHOLD,
+    RETRAIN_WALL_TIME_BUDGET_SEC,
+    RetrainWallTimeExceeded,
+    _check_rollback,
+    _feature_names,
+    _fetch_training_rows,
+    _measure_serialized_size,
+    _persist_candidate,
+    _promote_candidate,
+    _train_booster,
+)
+
+
+__all__ = (
+    "ConsolidationWorker",
+    "RetrainWallTimeExceeded",
+    "RETRAIN_HYPERPARAM_CAPS",
+    "RETRAIN_WALL_TIME_BUDGET_SEC",
+    "RETRAIN_MODEL_SIZE_BYTES_CAP",
+    "RETRAIN_NEW_OUTCOMES_THRESHOLD",
+    "RETRAIN_HOURS_THRESHOLD",
+    "_retrain_ranker_impl",
+    "_LEGACY_RETRAIN_DEPRECATED",
+    "_build_training_matrix",
+    "_compute_eval_metrics",
+    "_shadow_test_improved",
+    "_feature_names",
+    "_fetch_training_rows",
+    "_measure_serialized_size",
+    "_train_booster",
+    "_persist_candidate",
+    "_promote_candidate",
+    "_check_rollback",
+    "_run_shadow_cycle",
+)
+
+
+# ---------------------------------------------------------------------------
+# Orchestrator — defined HERE so tests patching ``cw_mod._train_booster``
+# actually intercept the helper call.
+# ---------------------------------------------------------------------------
+
+
+def _run_shadow_cycle(
+    *,
+    memory_db_path: str,
+    learning_db_path: str,
+    profile_id: str,
+) -> dict:
+    """Top-level online retrain cycle — runs inside the consolidation
+    worker.
+
+    Orchestrates: fetch rows → train → size-check → persist candidate
+    (NOT auto-promote). Promotion happens separately once the live
+    shadow-router accumulates enough observations (see
+    :mod:`superlocalmemory.core.shadow_router`).
+
+    Helper functions are looked up via this module's namespace so test
+    monkey-patches on ``consolidation_worker`` take effect without any
+    test churn.
+
+    Returns a dict with keys:
+      * ``aborted``: reason string if aborted (``'insufficient_data'``,
+        ``'model_too_large'``, ``'wall_time_exceeded'``, ``'train_error'``).
+      * ``candidate_persisted``: True if a candidate row was written.
+      * ``promoted``: False (always — promotion is a separate step).
+      * ``metrics``: training metrics dict on success.
     """
-
-    def __init__(self, memory_db: str | Path, learning_db: str | Path) -> None:
-        self._memory_db = str(memory_db)
-        self._learning_db = str(learning_db)
-
-    def run(self, profile_id: str, dry_run: bool = False) -> dict:
-        """Run full consolidation cycle. Returns stats."""
-        stats = {
-            "decayed": 0,
-            "deduped": 0,
-            "retrained": False,
-            "signal_count": 0,
-            "ranker_phase": 1,
-            "timestamp": datetime.now(timezone.utc).isoformat(),
-        }
-
-        # 1. Confidence decay on unused facts
-        try:
-            from superlocalmemory.learning.signals import LearningSignals
-            decayed = LearningSignals.decay_confidence(
-                self._memory_db, profile_id, rate=0.001,
-            )
-            stats["decayed"] = decayed
-            if not dry_run:
-                logger.info("Confidence decay: %d facts affected", decayed)
-        except Exception as exc:
-            logger.debug("Decay failed: %s", exc)
-
-        # 2. Deduplication (mark near-identical facts)
-        try:
-            deduped = self._deduplicate(profile_id, dry_run)
-            stats["deduped"] = deduped
-        except Exception as exc:
-            logger.debug("Dedup failed: %s", exc)
-
-        # 3. Generate behavioral patterns from memories
-        try:
-            patterns = self._generate_patterns(profile_id, dry_run)
-            stats["patterns_generated"] = patterns
-        except Exception as exc:
-            logger.debug("Pattern generation failed: %s", exc)
-
-        # 4. Recompute graph intelligence (v3.4.2: wired into learning pipeline)
-        try:
-            from superlocalmemory.core.graph_analyzer import GraphAnalyzer
-            conn_ga = sqlite3.connect(self._memory_db, timeout=10)
-            conn_ga.execute("PRAGMA busy_timeout=5000")
-            conn_ga.row_factory = sqlite3.Row
-
-            class _DBProxy:
-                """Minimal DB proxy for GraphAnalyzer compatibility."""
-                def __init__(self, connection: sqlite3.Connection) -> None:
-                    self._conn = connection
-                def execute(self, sql: str, params: tuple = ()) -> list:
-                    cursor = self._conn.execute(sql, params)
-                    if sql.strip().upper().startswith(("INSERT", "UPDATE", "DELETE", "ALTER", "CREATE")):
-                        self._conn.commit()
-                        return []
-                    return cursor.fetchall()
-
-            ga = GraphAnalyzer(_DBProxy(conn_ga))
-            if not dry_run:
-                ga_result = ga.compute_and_store(profile_id)
-                stats["graph_nodes"] = ga_result.get("node_count", 0)
-                stats["graph_communities"] = ga_result.get("community_count", 0)
-                logger.info(
-                    "Graph analysis: %d nodes, %d communities",
-                    stats["graph_nodes"], stats["graph_communities"],
-                )
-            conn_ga.close()
-        except Exception as exc:
-            logger.debug("Graph analysis failed: %s", exc)
-
-        # 5. Check if ranker should retrain
-        try:
-            from superlocalmemory.learning.feedback import FeedbackCollector
-            collector = FeedbackCollector(Path(self._learning_db))
-            signal_count = collector.get_feedback_count(profile_id)
-            stats["signal_count"] = signal_count
-            stats["ranker_phase"] = 1 if signal_count < 50 else (2 if signal_count < 200 else 3)
-
-            # Auto-retrain at threshold crossings
-            if signal_count >= 200 and not dry_run:
-                retrained = self._retrain_ranker(profile_id, signal_count)
-                stats["retrained"] = retrained
-        except Exception as exc:
-            logger.debug("Retrain check failed: %s", exc)
-
-        # 6. Entity compilation (v3.4.3: compiled truth per entity)
-        if not dry_run:
-            try:
-                from superlocalmemory.learning.entity_compiler import EntityCompiler
-                from superlocalmemory.core.config import SLMConfig
-                config = SLMConfig.load()
-                compiler = EntityCompiler(self._memory_db, config)
-                ec_result = compiler.compile_all(profile_id)
-                stats["entities_compiled"] = ec_result.get("compiled", 0)
-                if ec_result["compiled"] > 0:
-                    logger.info("Entity compilation: %d entities compiled",
-                                ec_result["compiled"])
-            except Exception as exc:
-                logger.debug("Entity compilation failed: %s", exc)
-
-        return stats
-
-    def _deduplicate(self, profile_id: str, dry_run: bool) -> int:
-        """Find and mark near-duplicate facts.
-
-        Uses content similarity (exact prefix match for now).
-        Does NOT delete — marks with lower confidence.
-        """
-        try:
-            conn = sqlite3.connect(self._memory_db, timeout=10)
-            conn.execute("PRAGMA busy_timeout=5000")
-            conn.row_factory = sqlite3.Row
-
-            rows = conn.execute(
-                "SELECT fact_id, content FROM atomic_facts "
-                "WHERE profile_id = ? ORDER BY created_at",
-                (profile_id,),
-            ).fetchall()
-
-            seen_prefixes: dict[str, str] = {}
-            duplicates = []
-
-            for r in rows:
-                d = dict(r)
-                prefix = d["content"][:100].strip().lower()
-                if prefix in seen_prefixes:
-                    duplicates.append(d["fact_id"])
-                else:
-                    seen_prefixes[prefix] = d["fact_id"]
-
-            if duplicates and not dry_run:
-                for fid in duplicates:
-                    conn.execute(
-                        "UPDATE atomic_facts SET confidence = MAX(0.1, confidence * 0.5) "
-                        "WHERE fact_id = ?",
-                        (fid,),
-                    )
-                conn.commit()
-
-            conn.close()
-            return len(duplicates)
-        except Exception:
-            return 0
-
-    def _generate_patterns(self, profile_id: str, dry_run: bool) -> int:
-        """Mine behavioral patterns from ALL memory sources.
-
-        v3.4.1: Expanded from 3 to 7 pattern types. No 500-fact cap.
-        Analyzes: facts, signals, co-retrieval edges, channel credits,
-        entities, sessions, graph communities.
-        """
-        try:
-            from superlocalmemory.learning.behavioral import BehavioralPatternStore
-            import re
-            from collections import Counter, defaultdict
-
-            conn = sqlite3.connect(self._memory_db, timeout=10)
-            conn.execute("PRAGMA busy_timeout=5000")
-            conn.row_factory = sqlite3.Row
-
-            # v3.4.1: No cap — analyze ALL facts
-            facts = conn.execute(
-                "SELECT fact_id, content, fact_type, created_at, session_id, "
-                "confidence, canonical_entities_json "
-                "FROM atomic_facts "
-                "WHERE profile_id = ? AND lifecycle = 'active' "
-                "ORDER BY created_at DESC",
-                (profile_id,),
-            ).fetchall()
-
-            if len(facts) < 5:
-                conn.close()
-                return 0
-
-            store = BehavioralPatternStore(self._learning_db)
-            generated = 0
-
-            # ── 1. Tech Preferences (expanded keyword list) ───────────
-            tech_keywords = {
-                "python": "Python", "javascript": "JavaScript",
-                "typescript": "TypeScript", "react": "React",
-                "vue": "Vue", "angular": "Angular",
-                "postgresql": "PostgreSQL", "mysql": "MySQL",
-                "sqlite": "SQLite", "docker": "Docker",
-                "kubernetes": "Kubernetes", "aws": "AWS",
-                "azure": "Azure", "gcp": "GCP",
-                "node": "Node.js", "fastapi": "FastAPI",
-                "django": "Django", "flask": "Flask",
-                "rust": "Rust", "go": "Go", "java": "Java",
-                "git": "Git", "npm": "npm", "pip": "pip",
-                "langchain": "LangChain", "ollama": "Ollama",
-                "pytorch": "PyTorch", "claude": "Claude",
-                "openai": "OpenAI", "anthropic": "Anthropic",
-                "redis": "Redis", "mongodb": "MongoDB",
-                "graphql": "GraphQL", "nextjs": "Next.js",
-                "terraform": "Terraform", "nginx": "Nginx",
-                "linux": "Linux", "macos": "macOS",
-                "vscode": "VS Code", "neovim": "Neovim",
-            }
-
-            tech_counts: Counter = Counter()
-            for f in facts:
-                content = dict(f)["content"].lower()
-                for keyword, label in tech_keywords.items():
-                    if keyword in content:
-                        tech_counts[label] += 1
-
-            for tech, count in tech_counts.most_common(20):
-                if count >= 2 and not dry_run:
-                    confidence = min(1.0, count / max(len(facts) * 0.1, 10))
-                    store.record_pattern(
-                        profile_id=profile_id,
-                        pattern_type="tech_preference",
-                        data={"topic": tech, "pattern_key": tech,
-                              "value": tech, "key": "tech",
-                              "evidence": count},
-                        success_rate=confidence,
-                        confidence=confidence,
-                    )
-                    generated += 1
-
-            # ── 2. Topic Interests (word frequency) ───────────────────
-            stopwords = frozenset({
-                "the", "is", "a", "an", "in", "on", "at", "to", "for",
-                "of", "and", "or", "not", "with", "that", "this", "was",
-                "are", "be", "has", "had", "have", "from", "by", "it",
-                "its", "as", "but", "were", "been", "being", "would",
-                "could", "should", "will", "may", "might", "can", "do",
-                "does", "did", "about", "into", "over", "after", "before",
-                "then", "than", "also", "just", "like", "more", "some",
-                "only", "other", "such", "each", "every", "both", "most",
-            })
-            word_counts: Counter = Counter()
-            for f in facts:
-                words = re.findall(r'\b[a-zA-Z]{4,}\b', dict(f)["content"].lower())
-                for w in words:
-                    if w not in stopwords:
-                        word_counts[w] += 1
-
-            for topic, count in word_counts.most_common(15):
-                if count >= 3 and not dry_run:
-                    confidence = min(1.0, count / max(len(facts) * 0.05, 15))
-                    store.record_pattern(
-                        profile_id=profile_id,
-                        pattern_type="interest",
-                        data={"topic": topic, "pattern_key": topic,
-                              "count": count, "evidence": count},
-                        success_rate=confidence,
-                        confidence=confidence,
-                    )
-                    generated += 1
-
-            # ── 3. Temporal Activity Patterns ─────────────────────────
-            hour_counts: Counter = Counter()
-            for f in facts:
-                created = dict(f).get("created_at", "")
-                try:
-                    if "T" in created:
-                        hour = int(created.split("T")[1][:2])
-                    elif " " in created:
-                        hour = int(created.split(" ")[1][:2])
-                    else:
-                        continue
-                    period = ("morning" if 6 <= hour < 12 else
-                              "afternoon" if 12 <= hour < 18 else
-                              "evening" if 18 <= hour < 22 else "night")
-                    hour_counts[period] += 1
-                except (ValueError, IndexError):
-                    pass
-
-            total_hours = sum(hour_counts.values())
-            for period, count in hour_counts.most_common():
-                if count >= 2 and total_hours > 0 and not dry_run:
-                    pct = round(count / total_hours * 100)
-                    store.record_pattern(
-                        profile_id=profile_id,
-                        pattern_type="temporal",
-                        data={"topic": period, "pattern_key": period,
-                              "value": f"{period} ({pct}%)",
-                              "evidence": count, "key": period,
-                              "distribution": dict(hour_counts)},
-                        success_rate=pct / 100,
-                        confidence=min(1.0, count / max(total_hours * 0.1, 5)),
-                    )
-                    generated += 1
-
-            # ── 4. Entity Preferences (v3.4.1 NEW) ───────────────────
-            import json as _json
-            entity_counts: Counter = Counter()
-            for f in facts:
-                raw = dict(f).get("canonical_entities_json", "")
-                if raw:
-                    try:
-                        for ent in _json.loads(raw):
-                            entity_counts[ent] += 1
-                    except (ValueError, TypeError):
-                        pass
-
-            # v3.4.7: Resolve entity IDs to readable canonical names
-            entity_names: dict = {}
-            try:
-                eid_list = list(entity_counts.keys())
-                if eid_list:
-                    placeholders = ",".join("?" * len(eid_list))
-                    name_rows = conn.execute(
-                        f"SELECT entity_id, canonical_name FROM canonical_entities "
-                        f"WHERE entity_id IN ({placeholders})",
-                        eid_list,
-                    ).fetchall()
-                    entity_names = {dict(r)["entity_id"]: dict(r)["canonical_name"] for r in name_rows}
-            except Exception:
-                pass
-
-            for entity, count in entity_counts.most_common(15):
-                if count >= 3 and not dry_run:
-                    readable = entity_names.get(entity, entity)
-                    confidence = min(1.0, count / max(len(facts) * 0.05, 10))
-                    store.record_pattern(
-                        profile_id=profile_id,
-                        pattern_type="entity_preferences",
-                        data={"topic": readable, "pattern_key": f"entity:{readable}",
-                              "value": readable, "evidence": count,
-                              "source": "entity_frequency"},
-                        success_rate=confidence,
-                        confidence=confidence,
-                    )
-                    generated += 1
-
-            # ── 5. Session Activity Patterns (v3.4.1 NEW) ────────────
-            session_counts: Counter = Counter()
-            for f in facts:
-                sid = dict(f).get("session_id", "")
-                if sid:
-                    session_counts[sid] += 1
-
-            if session_counts:
-                avg_facts_per_session = sum(session_counts.values()) / len(session_counts)
-                heavy_sessions = [s for s, c in session_counts.items() if c > avg_facts_per_session * 2]
-                if heavy_sessions and not dry_run:
-                    store.record_pattern(
-                        profile_id=profile_id,
-                        pattern_type="session_activity",
-                        data={"pattern_key": "heavy_session_usage",
-                              "value": f"{len(heavy_sessions)} intensive sessions",
-                              "evidence": len(heavy_sessions),
-                              "avg_facts": round(avg_facts_per_session, 1),
-                              "total_sessions": len(session_counts)},
-                        success_rate=0.8,
-                        confidence=min(1.0, len(heavy_sessions) / 5),
-                    )
-                    generated += 1
-
-            # ── 6. Fact Type Distribution (v3.4.1 NEW) ────────────────
-            type_counts: Counter = Counter()
-            for f in facts:
-                ft = dict(f).get("fact_type", "semantic")
-                type_counts[ft] += 1
-
-            total_ft = sum(type_counts.values())
-            if total_ft > 0 and not dry_run:
-                dominant_type = type_counts.most_common(1)[0]
-                pct = round(dominant_type[1] / total_ft * 100)
-                store.record_pattern(
-                    profile_id=profile_id,
-                    pattern_type="fact_type_distribution",
-                    data={"pattern_key": "memory_style",
-                          "value": f"{dominant_type[0]} dominant ({pct}%)",
-                          "evidence": dominant_type[1],
-                          "distribution": dict(type_counts)},
-                    success_rate=pct / 100,
-                    confidence=min(1.0, dominant_type[1] / 20),
-                )
-                generated += 1
-
-            # ── 7. Channel Performance (v3.4.1 NEW — from signals) ────
-            try:
-                learn_conn = sqlite3.connect(self._learning_db, timeout=10)
-                learn_conn.row_factory = sqlite3.Row
-
-                # Retrieval usage patterns from learning_feedback
-                channel_rows = learn_conn.execute(
-                    "SELECT channel, COUNT(*) AS cnt, "
-                    "AVG(signal_value) AS avg_signal "
-                    "FROM learning_feedback "
-                    "WHERE profile_id = ? "
-                    "GROUP BY channel ORDER BY cnt DESC",
-                    (profile_id,),
-                ).fetchall()
-
-                for row in channel_rows:
-                    d = dict(row)
-                    ch = d.get("channel", "unknown")
-                    cnt = d.get("cnt", 0)
-                    avg_sig = round(float(d.get("avg_signal", 0) or 0), 3)
-                    if cnt >= 5 and not dry_run:
-                        store.record_pattern(
-                            profile_id=profile_id,
-                            pattern_type="channel_performance",
-                            data={"pattern_key": f"channel:{ch}",
-                                  "value": f"{ch} ({cnt} hits, {avg_sig} avg)",
-                                  "evidence": cnt,
-                                  "avg_signal": avg_sig},
-                            success_rate=avg_sig,
-                            confidence=min(1.0, cnt / 50),
-                        )
-                        generated += 1
-
-                # Co-retrieval cluster patterns
-                try:
-                    coret_rows = learn_conn.execute(
-                        "SELECT fact_a, fact_b, co_access_count "
-                        "FROM co_retrieval_edges "
-                        "WHERE profile_id = ? AND co_access_count >= 3 "
-                        "ORDER BY co_access_count DESC LIMIT 20",
-                        (profile_id,),
-                    ).fetchall()
-                    if coret_rows and not dry_run:
-                        store.record_pattern(
-                            profile_id=profile_id,
-                            pattern_type="co_retrieval_clusters",
-                            data={"pattern_key": "co_retrieval_clusters",
-                                  "value": f"{len(coret_rows)} strong fact pairs",
-                                  "evidence": len(coret_rows),
-                                  "top_pair_count": dict(coret_rows[0]).get("co_access_count", 0) if coret_rows else 0},
-                            success_rate=0.7,
-                            confidence=min(1.0, len(coret_rows) / 10),
-                        )
-                        generated += 1
-                except Exception:
-                    pass
-
-                learn_conn.close()
-            except Exception as exc:
-                logger.debug("Signal pattern mining failed: %s", exc)
-
-            # ── 8. Community Membership (v3.4.1 NEW — from graph) ─────
-            try:
-                comm_rows = conn.execute(
-                    "SELECT community_id, COUNT(*) AS cnt "
-                    "FROM fact_importance "
-                    "WHERE profile_id = ? AND community_id IS NOT NULL "
-                    "GROUP BY community_id ORDER BY cnt DESC",
-                    (profile_id,),
-                ).fetchall()
-                if comm_rows and not dry_run:
-                    total_comm = sum(dict(r)["cnt"] for r in comm_rows)
-                    store.record_pattern(
-                        profile_id=profile_id,
-                        pattern_type="knowledge_structure",
-                        data={"pattern_key": "knowledge_structure",
-                              "value": f"{len(comm_rows)} topic communities, {total_comm} classified facts",
-                              "evidence": total_comm,
-                              "community_count": len(comm_rows)},
-                        success_rate=0.8,
-                        confidence=min(1.0, len(comm_rows) / 5),
-                    )
-                    generated += 1
-            except Exception:
-                pass
-
-            conn.close()
-
-            logger.info(
-                "Pattern mining: %d patterns generated for profile %s "
-                "from %d facts",
-                generated, profile_id, len(facts),
-            )
-            return generated
-        except Exception as exc:
-            logger.warning("Pattern generation error: %s", exc)
-            return 0
-
-    def _retrain_ranker(self, profile_id: str, signal_count: int) -> bool:
-        """Retrain the adaptive ranker from accumulated feedback."""
-        try:
-            from superlocalmemory.learning.feedback import FeedbackCollector
-            from superlocalmemory.learning.ranker import AdaptiveRanker
-
-            collector = FeedbackCollector(Path(self._learning_db))
-            feedback = collector.get_feedback(profile_id, limit=500)
-
-            if len(feedback) < 200:
-                return False
-
-            # Build training data from feedback
-            training_data = []
-            for f in feedback:
-                label = f.get("signal_value", 0.5)
-                training_data.append({
-                    "features": {"signal_value": label},
-                    "label": label,
-                })
-
-            ranker = AdaptiveRanker(signal_count=signal_count)
-            trained = ranker.train(training_data)
-
-            if trained:
-                logger.info("Ranker retrained with %d examples (Phase 3)", len(training_data))
-
-            return trained
-        except Exception as exc:
-            logger.debug("Retrain failed: %s", exc)
-            return False
+    out: dict = {
+        "aborted": None, "candidate_persisted": False,
+        "promoted": False, "metrics": None,
+    }
+
+    try:
+        rows, _qids = _fetch_training_rows(learning_db_path, profile_id)
+    except Exception as exc:
+        logger.debug("fetch_training_rows failed: %s", exc)
+        out["aborted"] = "fetch_error"
+        return out
+
+    if len(rows) < 20:
+        out["aborted"] = "insufficient_data"
+        return out
+
+    # Load prior active for in-sample shadow.
+    try:
+        from superlocalmemory.learning.database import LearningDatabase
+        db = LearningDatabase(learning_db_path)
+        prior_row = db.load_active_model(profile_id)
+    except Exception:
+        prior_row = None
+
+    feature_names = _feature_names()
+
+    try:
+        booster, metrics = _train_booster(
+            learning_db_path, profile_id,
+            training_rows=rows, feature_names=feature_names,
+            prior_row=prior_row,
+        )
+    except RetrainWallTimeExceeded as exc:
+        out["aborted"] = "wall_time_exceeded"
+        out["metrics"] = {"wall_time_sec": exc.elapsed_sec}
+        return out
+    except Exception as exc:
+        logger.debug("train_booster failed: %s", exc)
+        out["aborted"] = "train_error"
+        return out
+
+    # Model-size guardrail (LLD-10 §3.2 post-train check).
+    size_bytes = _measure_serialized_size(booster)
+    if size_bytes > RETRAIN_MODEL_SIZE_BYTES_CAP:
+        logger.warning(
+            "retrain: candidate %.2f MB exceeds %.1f MB cap — rejecting",
+            size_bytes / 1e6, RETRAIN_MODEL_SIZE_BYTES_CAP / 1e6,
+        )
+        out["aborted"] = "model_too_large"
+        out["metrics"] = metrics
+        return out
+
+    # In-sample shadow gate — cheap filter before spending live recalls.
+    if prior_row is not None:
+        if not _shadow_test_improved(
+            prior_row, booster, rows, feature_names,
+        ):
+            out["aborted"] = "insample_shadow_fail"
+            out["metrics"] = metrics
+            return out
+
+    try:
+        state_bytes = booster.model_to_string().encode("utf-8")
+    except Exception as exc:  # pragma: no cover — defensive
+        logger.debug("model serialise failed: %s", exc)
+        out["aborted"] = "serialise_error"
+        return out
+
+    try:
+        _persist_candidate(
+            learning_db_path, profile_id=profile_id,
+            state_bytes=state_bytes, feature_names=feature_names,
+            trained_on_count=len(rows), metrics=metrics,
+            shadow_results={"in_sample_pass": prior_row is not None},
+        )
+    except sqlite3.Error as exc:
+        logger.warning("persist_candidate failed: %s", exc)
+        out["aborted"] = "persist_error"
+        return out
+
+    out["candidate_persisted"] = True
+    out["promoted"] = False  # Promotion is a separate live-shadow step.
+    out["metrics"] = metrics
+    return out