superlocalmemory 3.8.6 → 3.8.8

package/CHANGELOG.md

@@ -5,6 +5,52 @@ All notable changes to SuperLocalMemory V3 will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [3.8.8] - 2026-07-27 — Live-database recall and vector integrity
+
+### Fixed
+- Foreground recall now retains priority while startup repair and canonical
+  remember enrichment are active. Background embedding work yields between
+  items and cannot cold-start the shared local model ahead of a user query.
+- Daemon readiness now verifies that the local embedding subprocess is
+  actually alive and has served a request instead of trusting a stale startup
+  flag.
+- Scene clustering reuses durable fact embeddings and ignores legacy scene
+  rows whose facts were already consolidated away. Mature databases no longer
+  issue thousands of redundant model requests or repeatedly recycle the
+  embedding worker after restart.
+- sqlite-vec row allocation is serialized across processes, and every metadata
+  pointer is validated against the vector payload's profile. Cross-profile
+  row-id collisions are repaired without exposing another profile's memory.
+- Vector writes roll back abandoned transactions, missing vectors are repaired
+  exactly, and semantic search expands past legacy orphan payloads.
+- Entity-cache startup reads are bounded and load only the columns used by the
+  entity channel.
+- Exact lexical evidence remains protected after learned ranking so a freshly
+  remembered exact marker cannot be displaced by a weaker semantic match.
+
+### Notes
+- Recall remains physically query-only; remember owns durable writes and
+  continues full canonical enrichment asynchronously after immediate queryable
+  admission.
+- This candidate must pass the installed wheel, installed npm tarball,
+  dashboard, MCP, CLI, and mature live-database gates before publication.
+
+## [3.8.7] - 2026-07-27 — Existing graph-store compatibility
+
+### Fixed
+- Upgrades now retain the PyCozo 0.3.0 native binding used to create existing
+  Cozo graph stores. Version 3.8.6 selected PyCozo 0.7.6, whose incompatible
+  on-disk format could stop the daemon at startup with
+  `Unknown storage version 1`.
+- A native Cozo panic is now contained at the optional projection boundary.
+  The daemon remains available on canonical SQLite and preserves the graph
+  files unchanged instead of losing access to memory.
+
+### Notes
+- Existing PyCozo 0.3.0 graphs require no rebuild or database migration.
+- All serialized-write, read-only recall, admission-journal, and dead-letter
+  fixes from 3.8.6 are unchanged.
+
 ## [3.8.6] - 2026-07-27 — Serialized writes and read-only recall
 
 ### Fixed

package/README.md

@@ -5,15 +5,15 @@
   </picture>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.8.6</h1>
+<h1 align="center">SuperLocalMemory V3.8.8</h1>
 <p align="center"><strong>Enterprise-grade, local-first memory for AI agents and teams.</strong><br/>
 <em>A persistent, auditable long-term brain for your agents that runs on your own infrastructure — with multi-workspace isolation, role-based access, and GDPR + EU AI Act governance controls built in.</em></p>
-<p align="center"><code>v3.8.6</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
+<p align="center"><code>v3.8.8</code> — one control plane: auditable retrieval · multi-scope memory (personal / shared / global) · Cache · Compress · trusted-peer Mesh · bounded loops — across CLI, MCP, dashboard, the <strong>Claude plugin</strong>, the <strong>Codex add-on</strong>, and documented IDE integrations.<br/>
 Proxy: <code>slm wrap claude</code> &nbsp;·&nbsp; MCP: add <code>slm_compress</code> to your config &nbsp;·&nbsp; Skill: zero-config</p>
 <p align="center"><strong>3 public research preprints</strong> (arXiv + Zenodo archives) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>
 
 <p align="center">
-  <a href="CHANGELOG.md"><img src="https://img.shields.io/badge/v3.8.6-Current_Release-2ea44f?style=for-the-badge&logo=checkmarx&logoColor=white" alt="v3.8.6 — Current Release"/></a>
+  <a href="CHANGELOG.md"><img src="https://img.shields.io/badge/v3.8.8-Current_Release-2ea44f?style=for-the-badge&logo=checkmarx&logoColor=white" alt="v3.8.8 — Current Release"/></a>
   <a href="https://arxiv.org/abs/2603.14588"><img src="https://img.shields.io/badge/arXiv-2603.14588-b31b1b?style=for-the-badge&logo=arxiv&logoColor=white" alt="arXiv Paper"/></a>
   <a href="#three-surfaces-proxy--mcp-tools--skill"><img src="https://img.shields.io/badge/Proxy_|_MCP_|_Skill-22c55e?style=for-the-badge" alt="Three Surfaces: Proxy, MCP Tools, Skill"/></a>
   <a href="https://pypi.org/project/superlocalmemory/"><img src="https://img.shields.io/pypi/v/superlocalmemory?style=for-the-badge&logo=pypi&logoColor=white" alt="PyPI"/></a>

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.8.6",
+  "version": "3.8.8",
   "description": "Local-first agent memory with MCP and an agent-native CLI. Documented clients include Claude Code, Cursor, and Windsurf.",
   "keywords": [
     "ai-memory",

package/plugin/.claude-plugin/plugin.json

@@ -15,5 +15,5 @@
   "mcpServers": "./.mcp.json",
   "name": "superlocalmemory",
   "repository": "https://github.com/qualixar/superlocalmemory",
-  "version": "3.8.6"
+  "version": "3.8.8"
 }

package/plugin/CLAUDE.md

@@ -1,4 +1,4 @@
-<!-- BEGIN SuperLocalMemory v3.8.6 -->
+<!-- BEGIN SuperLocalMemory v3.8.8 -->
 
 ## SuperLocalMemory (SLM) — Agent Rules
 
@@ -39,6 +39,6 @@ slm-recall · slm-remember · slm-session · slm-status · slm-cache · slm-comp
 ### Subagents
 slm-memory-advisor (memory decisions, session hygiene, scope/profile guidance) · slm-optimize-advisor (context compression + KV cache) · slm-governance-advisor (scope/roles/compliance/GDPR)
 
-<!-- END SuperLocalMemory v3.8.6 -->
+<!-- END SuperLocalMemory v3.8.8 -->
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/agents/slm-governance-advisor.md

@@ -77,4 +77,4 @@ slm-scope · slm-governance · slm-profile · slm-remember · slm-recall
 # What NOT to do
 Never session_init twice; never forget without dry-run preview; never store secrets; never bypass role checks; never claim an erasure succeeded without verifying via recall.
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/agents/slm-loop-runner.md

@@ -68,4 +68,4 @@ assessment. The gate is the authority.
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/agents/slm-memory-advisor.md

@@ -46,4 +46,4 @@ slm-recall · slm-remember · slm-session · slm-scope · slm-profile · slm-gov
 # What NOT to do
 Never session_init twice; never forget dry_run=False without reporting preview; never dump a whole file into remember; never invent a memory; never claim "saved" without success:true / clean CLI exit; never bypass scope or governance restrictions.
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/agents/slm-optimize-advisor.md

@@ -41,4 +41,4 @@ slm-compress · slm-cache · slm-status · slm-profile
 # What NOT to do
 Never compress code-for-edit/JSON-to-parse/<500 chars; never store secrets/ccr_ids; never let optimize failure block/alter the task; never claim a specific savings %; never carry ccr_ids across profile switches.
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/requirements.txt

@@ -1 +1 @@
-superlocalmemory==3.8.6
+superlocalmemory==3.8.8

package/plugin/skills/slm-cache/SKILL.md

@@ -145,4 +145,4 @@ These subcommands control daemon-level cache settings. They do not read or write
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/skills/slm-compress/SKILL.md

@@ -147,4 +147,4 @@ Content over 1 MB (1 000 000 bytes UTF-8) is processed but `reversible` is force
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/skills/slm-governance/SKILL.md

@@ -245,4 +245,4 @@ Before running any destructive operation (`forget`, `compact_memories`):
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-graph/SKILL.md

@@ -311,4 +311,4 @@ profile. See `slm-profile` for the full profile switching workflow.
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/skills/slm-loop/SKILL.md

@@ -96,4 +96,4 @@ paused, name the approval needed; when errored, quote the short detail.
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin/skills/slm-mesh/SKILL.md

@@ -279,4 +279,4 @@ mesh availability.
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-profile/SKILL.md

@@ -145,4 +145,4 @@ Name them differently in your MCP config (e.g. `superlocalmemory-personal` and
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-recall/SKILL.md

@@ -236,4 +236,4 @@ before recalling, then switch back. See `slm-profile` for workspace switching.
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-remember/SKILL.md

@@ -238,4 +238,4 @@ different workspace, use `switch_profile` first. See `slm-profile`.
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-scope/SKILL.md

@@ -173,4 +173,4 @@ to review the impact. See `slm-remember` for the full deletion discipline.
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-session/SKILL.md

@@ -227,4 +227,4 @@ explicitly and call `recall` with `include_global`/`include_shared` after
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin/skills/slm-status/SKILL.md

@@ -163,4 +163,4 @@ multi-profile setup. To switch the active profile, see `slm-profile`.
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin-src/rules/AGENTS.md

@@ -128,4 +128,4 @@ When the SLM MCP server is unavailable, use these CLI equivalents:
 - **slm-optimize-advisor** — context compression and KV cache
 - **slm-governance-advisor** — scope/role compliance, retention policies, GDPR
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin-src/skills/slm-cache/SKILL.md

@@ -145,4 +145,4 @@ These subcommands control daemon-level cache settings. They do not read or write
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin-src/skills/slm-compress/SKILL.md

@@ -147,4 +147,4 @@ Content over 1 MB (1 000 000 bytes UTF-8) is processed but `reversible` is force
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin-src/skills/slm-graph/SKILL.md

@@ -311,4 +311,4 @@ profile. See `slm-profile` for the full profile switching workflow.
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/plugin-src/skills/slm-recall/SKILL.md

@@ -236,4 +236,4 @@ before recalling, then switch back. See `slm-profile` for workspace switching.
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin-src/skills/slm-remember/SKILL.md

@@ -238,4 +238,4 @@ different workspace, use `switch_profile` first. See `slm-profile`.
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin-src/skills/slm-session/SKILL.md

@@ -227,4 +227,4 @@ explicitly and call `recall` with `include_global`/`include_shared` after
 
 ---
 
-*SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later*
+*SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later*

package/plugin-src/skills/slm-status/SKILL.md

@@ -163,4 +163,4 @@ multi-profile setup. To switch the active profile, see `slm-profile`.
 
 ---
 
-SuperLocalMemory v3.8.6 · Qualixar · AGPL-3.0-or-later
+SuperLocalMemory v3.8.8 · Qualixar · AGPL-3.0-or-later

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.8.6"
+version = "3.8.8"
 description = "Local-first agent memory with auditable hybrid retrieval"
 readme = "README.md"
 license = "AGPL-3.0-or-later"
@@ -82,7 +82,10 @@ dependencies = [
     # until `slm db scale prepare -> verify -> promote` proves parity with
     # canonical SQLite; dependency installation never migrates user data.
     "lancedb==0.30.2",
-    "pycozo[embedded]==0.7.6",
+    # Cozo's on-disk format is not forward-compatible across these native
+    # bindings. Existing SLM graph stores were created by 0.3.0; upgrading the
+    # wheel without an export/import migration makes daemon startup panic.
+    "pycozo[embedded]==0.3.0",
     # v3.6.10: LLMLingua-2 prose compression (aggressive mode, opt-in at runtime).
     # Hard dependency so `pip install superlocalmemory` ships compression-ready;
     # the ~560MB model downloads on first setup/warmup (fail-open). Verified the
@@ -121,13 +124,13 @@ lancedb = [
     "lancedb==0.30.2",
 ]
 cozo = [
-    "pycozo[embedded]==0.7.6",
+    "pycozo[embedded]==0.3.0",
 ]
 # Backwards-compatible explicit install target for automation and docs.  The
 # packages are also base dependencies from v3.7 onward.
 scale = [
     "lancedb==0.30.2",
-    "pycozo[embedded]==0.7.6",
+    "pycozo[embedded]==0.3.0",
 ]
 ingestion = [
     "keyring>=25.0.0",

package/src/superlocalmemory/__init__.py

@@ -32,7 +32,7 @@ if "OMP_NUM_THREADS" not in os.environ:
     os.environ["OMP_NUM_THREADS"] = "2"
 # ---------------------------------------------------------------------------
 
-__version__ = "3.8.6"
+__version__ = "3.8.8"
 
 _REQUIRED_VERSIONS = {
     "sentence_transformers": "5.3.0",

package/src/superlocalmemory/core/backend_orchestrator.py

@@ -438,7 +438,13 @@ class BackendOrchestrator:
             self._cozo = CozoDBGraphBackend(str(cozo_path / "graph"))
             self._update_status("cozo", "not_initialized")
             logger.info("CozoDB initialized at %s", cozo_path)
-        except Exception as exc:
+        except BaseException as exc:
+            # PyO3 exposes Rust panics as PanicException(BaseException), not
+            # Exception. An incompatible optional projection must never abort
+            # daemon startup or hide canonical SQLite memory. Re-raise genuine
+            # process-control exceptions; preserve the graph and degrade Cozo.
+            if not isinstance(exc, Exception) and type(exc).__name__ != "PanicException":
+                raise
             logger.warning("CozoDB init failed: %s", exc)
             self._cozo = None
 

package/src/superlocalmemory/core/embeddings.py

@@ -220,6 +220,7 @@ class EmbeddingService:
         self._owns_worker_lock = False
         self._request_count: int = 0
         self._http_client: object | None = None
+        self._remote_ready = False
 
         # Register for atexit cleanup (prevent orphaned workers)
         ref = weakref.ref(self, _live_embedding_services.discard)
@@ -246,6 +247,31 @@ class EmbeddingService:
             return bool(self._config.api_endpoint and self._config.api_key)
         return self._available
 
+    @property
+    def is_warm(self) -> bool:
+        """Return whether the configured backend has served a request.
+
+        ``is_available`` only means that the backend may be started.  A local
+        sentence-transformers cold start can take minutes on Apple Silicon, so
+        background enrichment must not mistake availability for readiness and
+        occupy the only worker ahead of an interactive recall.
+        """
+        config = getattr(self, "_config", None)
+        if config is not None and (
+            config.is_openai_compatible or config.is_cloud
+        ):
+            return bool(
+                getattr(self, "_remote_ready", False)
+                and self.is_available
+            )
+        proc = getattr(self, "_worker_proc", None)
+        if proc is None or getattr(self, "_request_count", 0) <= 0:
+            return False
+        try:
+            return proc.poll() is None
+        except Exception:
+            return False
+
     @property
     def dimension(self) -> int:
         return self._config.dimension
@@ -291,13 +317,27 @@ class EmbeddingService:
         """Embed a single text string. Returns list of floats or None."""
         if not text or not text.strip():
             raise ValueError("Cannot embed empty text")
+        from superlocalmemory.core.recall_gate import wait_for_foreground_idle
+        wait_for_foreground_idle()
         if self._config.is_openai_compatible:
-            vecs = self._openai_compatible_embed_batch([text])
-            vec = vecs[0]
-            self._validate_dimension(np.asarray(vec))
-            return vec
+            try:
+                vecs = self._openai_compatible_embed_batch([text])
+                vec = vecs[0]
+                self._validate_dimension(np.asarray(vec))
+                self._remote_ready = True
+                return vec
+            except Exception:
+                self._remote_ready = False
+                raise
         if self._config.is_cloud:
-            return self._cloud_embed_single(text)
+            try:
+                vec = self._cloud_embed_single(text)
+                self._validate_dimension(np.asarray(vec))
+                self._remote_ready = True
+                return vec
+            except Exception:
+                self._remote_ready = False
+                raise
         result = self._subprocess_embed([text])
         if result is None:
             return None
@@ -309,14 +349,34 @@ class EmbeddingService:
         """Embed a batch of texts."""
         if not texts:
             raise ValueError("Cannot embed empty batch")
+        from superlocalmemory.core.recall_gate import is_background_work
+        if is_background_work():
+            # A single large background batch can own the only local inference
+            # worker for tens of seconds. Slice it so a recall arriving after
+            # this call started gets priority before the next text.
+            return [self.embed(text) for text in texts]
         if self._config.is_openai_compatible:
-            results = self._openai_compatible_embed_batch(texts)
-            for vec in results:
-                if vec is not None:
-                    self._validate_dimension(np.asarray(vec))
-            return results
+            try:
+                results = self._openai_compatible_embed_batch(texts)
+                for vec in results:
+                    if vec is not None:
+                        self._validate_dimension(np.asarray(vec))
+                self._remote_ready = any(vec is not None for vec in results)
+                return results
+            except Exception:
+                self._remote_ready = False
+                raise
         if self._config.is_cloud:
-            return self._cloud_embed_batch(texts)
+            try:
+                results = self._cloud_embed_batch(texts)
+                for vec in results:
+                    if vec is not None:
+                        self._validate_dimension(np.asarray(vec))
+                self._remote_ready = any(vec is not None for vec in results)
+                return results
+            except Exception:
+                self._remote_ready = False
+                raise
         result = self._subprocess_embed(texts)
         if result is None:
             return [None] * len(texts)
@@ -353,6 +413,8 @@ class EmbeddingService:
         Includes a timeout (_SUBPROCESS_RESPONSE_TIMEOUT seconds) so the CLI
         never hangs indefinitely on cold model loads or network issues.
         """
+        from superlocalmemory.core.recall_gate import wait_for_foreground_idle
+        wait_for_foreground_idle()
         with self._lock:
             # Only an explicit terminal disable (``False``) short-circuits. A
             # ``None`` availability is the recall-health self-heal's "re-probe"

package/src/superlocalmemory/core/engine.py

@@ -637,7 +637,12 @@ class MemoryEngine:
             except (ValueError, TypeError):
                 _timeout_s = 0.5  # default 500 ms
             try:
-                _future = self._store_fast_embed_pool.submit(_embedder_ref.embed, fact_text)
+                def _best_effort_embed():
+                    from superlocalmemory.core.recall_gate import background_work
+                    with background_work():
+                        return _embedder_ref.embed(fact_text)
+
+                _future = self._store_fast_embed_pool.submit(_best_effort_embed)
                 try:
                     emb = _future.result(timeout=_timeout_s)
                     if emb:

package/src/superlocalmemory/core/ollama_embedder.py

@@ -90,6 +90,8 @@ class OllamaEmbedder:
         """
         if not text or not text.strip():
             raise ValueError("Cannot embed empty text")
+        from superlocalmemory.core.recall_gate import wait_for_foreground_idle
+        wait_for_foreground_idle()
 
         # V3.3.27: Check cache first
         cache_key = text.strip()
@@ -119,6 +121,9 @@ class OllamaEmbedder:
         """
         if not texts:
             raise ValueError("Cannot embed empty batch")
+        from superlocalmemory.core.recall_gate import is_background_work
+        if is_background_work():
+            return [self.embed(text) for text in texts]
 
         # V3.3.27: Split into cached and uncached
         results: list[list[float] | None] = [None] * len(texts)

package/src/superlocalmemory/core/recall_gate.py

@@ -14,23 +14,58 @@ recall.
 from __future__ import annotations
 
 import threading
+from contextlib import contextmanager
+from typing import Iterator
 
-_lock = threading.Lock()
+_condition = threading.Condition(threading.Lock())
 _active = 0
+_work_context = threading.local()
 
 
 def begin_recall() -> None:
     global _active
-    with _lock:
+    with _condition:
         _active += 1
 
 
 def end_recall() -> None:
     global _active
-    with _lock:
+    with _condition:
         _active = max(0, _active - 1)
+        if _active == 0:
+            _condition.notify_all()
 
 
 def in_flight() -> int:
-    with _lock:
+    with _condition:
         return _active
+
+
+@contextmanager
+def background_work() -> Iterator[None]:
+    """Mark best-effort work that must yield shared inference to recall.
+
+    The marker is thread-local because materialization, health probes, and
+    interactive handlers all share one resident engine and one embedder.
+    Nested callers restore the previous marker on exit.
+    """
+    previous = bool(getattr(_work_context, "background", False))
+    _work_context.background = True
+    try:
+        yield
+    finally:
+        _work_context.background = previous
+
+
+def is_background_work() -> bool:
+    """Return whether the current thread is running best-effort work."""
+    return bool(getattr(_work_context, "background", False))
+
+
+def wait_for_foreground_idle() -> None:
+    """Block background inference while an interactive recall is active."""
+    if not is_background_work():
+        return
+    with _condition:
+        while _active > 0:
+            _condition.wait(timeout=0.1)

package/src/superlocalmemory/core/recall_pipeline.py

@@ -100,6 +100,45 @@ def _apply_markers_to_response(response: RecallResponse) -> None:
         r.marker = _emit_marker(r.fact.fact_id)
 
 
+def _preserve_exact_lexical_evidence(
+    response: RecallResponse,
+    query: str,
+) -> None:
+    """Keep a deterministic exact BM25 hit ahead of learned refinements.
+
+    Adaptive and bandit ranking are valuable for ambiguous candidates, but
+    they must not demote a fact containing the caller's exact query behind
+    semantically similar noise. This guard runs after every learned layer and
+    changes only ordering; it does not introduce or bypass evidence.
+    """
+    normalized_query = " ".join(query.casefold().split())
+    if len(normalized_query) < 3 or len(response.results) < 2:
+        return
+    exact = [
+        result
+        for result in response.results
+        if (
+            float((result.channel_scores or {}).get("bm25", 0.0) or 0.0) > 0.0
+            and normalized_query
+            in " ".join(result.fact.content.casefold().split())
+        )
+    ]
+    if not exact:
+        return
+    strongest = max(
+        exact,
+        key=lambda result: float(
+            (result.channel_scores or {}).get("bm25", 0.0) or 0.0,
+        ),
+    )
+    if response.results[0] is strongest:
+        return
+    response.results = [
+        strongest,
+        *(result for result in response.results if result is not strongest),
+    ]
+
+
 # ---------------------------------------------------------------------------
 # Stage 8 SB-1 — feed shadow_router from recall-settled signals.
 #
@@ -854,6 +893,7 @@ def run_recall(
     except Exception as exc:
         logger.debug("Ranking pipeline skipped: %s", exc)
 
+    _preserve_exact_lexical_evidence(response, query)
     _mark("learning+ranking")
     # Deliberately no trust, Fisher, retention, lifecycle, popularity, or graph
     # mutation here.  Those state transitions require a separately authenticated