perf(doc-trainer): skip cross-linking/emotion + inline type in train mode (−42%)

src/surreal_memory/engine/pipeline_steps.py

@@ -1126,6 +1126,10 @@ async def execute(
         config: BrainConfig,
     ) -> PipelineContext:
         assert ctx.anchor_neuron is not None
+        # Bulk doc-training: skip emotion extraction (find_neurons type=STATE
+        # scans type partition; doc knowledge has no emotional valence).
+        if getattr(ctx, "skip_conflicts", False):
+            return ctx
         result = self.sentiment_extractor.extract(ctx.content, language=ctx.language)
 
         if result.valence == Valence.NEUTRAL or not result.emotion_tags:
@@ -1291,7 +1295,7 @@ async def execute(
         storage: NeuralStorage,
         config: BrainConfig,
     ) -> PipelineContext:
-        if ctx.skip_conflicts:
+        if getattr(ctx, "skip_conflicts", False):
             return ctx
 
         assert ctx.anchor_neuron is not None
@@ -1436,6 +1440,14 @@ async def execute(
         config: BrainConfig,
     ) -> PipelineContext:
         # Collect entity and concept neurons from this encode
+        # Bulk doc-training: skip cross-linking to existing neurons. Each
+        # find_neurons(type, content_exact) call scans the full type partition
+        # (40k+ rows on a large brain) because SDB 3.2.0's planner picks
+        # idx_neuron_type and filters content post-index — ~1 s/call, 3+/chunk.
+        # ENRICH consolidation (post-train) adds cross-links anyway.
+        if getattr(ctx, "skip_conflicts", False):
+            return ctx
+
         linkable = [
             n
             for n in ctx.entity_neurons
@@ -1532,6 +1544,9 @@ async def execute(
     ) -> PipelineContext:
         if ctx.anchor_neuron is None or not ctx.entity_neurons:
             return ctx
+        # Bulk doc-training: skip (same rationale as SemanticLinkingStep).
+        if getattr(ctx, "skip_conflicts", False):
+            return ctx
 
         anchor_id = ctx.anchor_neuron.id
         new_ids = {n.id for n in ctx.neurons_created}

src/surreal_memory/storage/surrealdb/store.py

@@ -697,8 +697,14 @@ async def find_neurons(
         params: dict[str, Any] = {}
 
         if type is not None:
-            conditions.append("type = $ntype")
-            params["ntype"] = type.value
+            # Inline type as a literal too — same planner gotcha as brain_id: a
+            # parameterized ``$ntype`` makes the planner pick idx_neuron_brain
+            # (single-col) and filter type over all brain rows, instead of the
+            # composite idx_neuron_type (brain_id, type). Measured 1147 ms/q vs
+            # single-digit ms/q with the literal on a 92k-neuron brain. type.value
+            # comes from the NeuronType enum (uppercase identifier), so inlining
+            # is injection-safe.
+            conditions.append(f"type = '{type.value}'")
 
         if content_exact is not None:
             conditions.append("content = $content_exact")
@@ -742,25 +748,22 @@ async def find_neurons_exact_batch(
     ) -> dict[str, Neuron]:
         """Batched exact-content lookup — one round-trip for N contents.
 
-        Overrides the base N+1 default (sequential ``find_neurons`` per content).
-        Uses ``content IN $contents`` with the brain_id inlined as a literal so
-        the planner uses the brain_id/content indexes (a parameterized brain_id
-        full-scans). Returns the first match per content string.
+        Uses ``content IN $contents`` with brain_id inlined. On a 92k-neuron
+        brain this is ~1.3 s/call but still 4-5x faster than the base N+1
+        default (each individual find_neurons with type scans the type
+        partition at ~750 ms; for N=8 candidates that's 6 s vs 1.3 s).
         """
         if not contents:
             return {}
         brain_id = self._get_brain_id()
         conds = [f"brain_id = {_brain_literal(brain_id)}", "content IN $contents"]
         params: dict[str, Any] = {"contents": list(dict.fromkeys(contents))}
         if type is not None:
-            conds.append("type = $ntype")
-            params["ntype"] = type.value
+            conds.append(f"type = '{type.value}'")
         if ephemeral is not None:
             conds.append("ephemeral = $ephemeral")
             params["ephemeral"] = ephemeral
         where = " AND ".join(conds)
-        # OMIT embedding_vec — callers only need identity/type/content, and the
-        # 1024-3072-float vector is ~4-8 KB/row across a result set of N matches.
         rows = await self._query(f"SELECT * OMIT embedding_vec FROM neuron WHERE {where}", **params)
         out: dict[str, Neuron] = {}
         for r in rows: