self-review 2026-07-22: persist canonical snapshot on write routes; newest-first tie-break in node listings

Durability: the 2026-07-21 fix stopped read routes writing the canonical
snapshot but left no save on ANY write path — every mutation lived in RAM
until a manual POST /api/save. Observed live: two restarts reverted the
store to a 17h-old snapshot, destroying same-day writes. persist_canonical()
now runs after node/edge create, knowledge capture, forget, strengthen, and
load-merge. ISE telemetry excluded deliberately (48h-pruned, loss-tolerant,
~2/min; snapshotting 28MB per heartbeat is waste).

Listing order: scan routes sort by salience with store-order ties, so
equal-salience telemetry (all ISEs are 0.3) returned OLDEST first — a
limited /api/nodes query silently returned a stale window, and a 41h-old
heartbeat series read as a live outage during this review. Ties now break
newest-first by created_at.

.gitea/workflows/ci-dev.yaml

@@ -214,18 +214,9 @@ jobs:
         env:
           GCP_SA_KEY: ${{ secrets.GCP_SA_KEY }}
         run: |
-          # Fail loudly: previously this step had no `set -e`, so an auth or
-          # upload failure was swallowed (step exited 0 on the trailing echo)
-          # and the SDK silently never published. Surface failures now.
-          set -euo pipefail
-          if [ -z "${GCP_SA_KEY:-}" ]; then
-            echo "FATAL: GCP_SA_KEY secret is empty — cannot authenticate to publish" >&2
-            exit 1
-          fi
           echo "${GCP_SA_KEY}" > /tmp/gcp-key.json
           gcloud auth activate-service-account --key-file=/tmp/gcp-key.json
           gcloud config set project neuron-785695
-          echo "Publishing as active account: $(gcloud config get-value account 2>/dev/null)"
 
           VERSION="${GITHUB_SHA:0:8}"
 
@@ -277,12 +268,6 @@ jobs:
         # Patches ci-base:dev in-place: pulls the existing image (which has all
         # system deps — Node, Go, gcloud, Docker CLI, etc.) and overlays the freshly
         # built El SDK on top. Keeps the full ci-base rebuild fast and incremental.
-        #
-        # continue-on-error: this is a CI-cache optimization, NOT the release
-        # artifact. It runs Docker (pull/build/push ~600MB) on the host-mode GCE
-        # runner where DinD/Docker availability is fragile. A failure here must
-        # never block or redden the job — the SDK publish above is the deliverable.
-        continue-on-error: true
         if: github.event_name == 'push'
         env:
           GCP_SA_KEY: ${{ secrets.GCP_SA_KEY }}

.gitea/workflows/ci-stage.yaml

@@ -212,21 +212,12 @@ jobs:
         env:
           GCP_SA_KEY: ${{ secrets.GCP_SA_KEY }}
         run: |
-          # Fail loudly: previously this step had no `set -e`, so an auth or
-          # upload failure was swallowed (step exited 0 on the trailing echo)
-          # and the SDK silently never published. Surface failures now.
-          set -euo pipefail
-          if [ -z "${GCP_SA_KEY:-}" ]; then
-            echo "FATAL: GCP_SA_KEY secret is empty — cannot authenticate to publish" >&2
-            exit 1
-          fi
           echo "${GCP_SA_KEY}" > /tmp/gcp-key.json
           apt-get install -y -qq apt-transport-https ca-certificates curl
           echo "deb [trusted=yes] https://packages.cloud.google.com/apt cloud-sdk main" > /etc/apt/sources.list.d/google-cloud-sdk.list
           apt-get update -qq && apt-get install -y google-cloud-cli
           gcloud auth activate-service-account --key-file=/tmp/gcp-key.json
           gcloud config set project neuron-785695
-          echo "Publishing as active account: $(gcloud config get-value account 2>/dev/null)"
 
           VERSION="${GITHUB_SHA:0:8}"
 
@@ -262,12 +253,6 @@ jobs:
         # Patches ci-base:stage in-place: pulls the existing image (which has all
         # system deps — Node, Go, gcloud, Docker CLI, etc.) and overlays the freshly
         # built El SDK on top. Keeps the full ci-base rebuild fast and incremental.
-        #
-        # continue-on-error: this is a CI-cache optimization, NOT the release
-        # artifact. It runs Docker (pull/build/push ~600MB) on the host-mode GCE
-        # runner where DinD/Docker availability is fragile. A failure here must
-        # never block or redden the job — the SDK publish above is the deliverable.
-        continue-on-error: true
         if: github.event_name == 'push'
         env:
           GCP_SA_KEY: ${{ secrets.GCP_SA_KEY }}

.gitea/workflows/sdk-release.yaml

@@ -288,21 +288,12 @@ jobs:
         env:
           GCP_SA_KEY: ${{ secrets.GCP_SA_KEY }}
         run: |
-          # Fail loudly: previously this step had no `set -e`, so an auth or
-          # upload failure was swallowed (step exited 0 on the trailing echo)
-          # and the SDK silently never published. Surface failures now.
-          set -euo pipefail
-          if [ -z "${GCP_SA_KEY:-}" ]; then
-            echo "FATAL: GCP_SA_KEY secret is empty — cannot authenticate to publish" >&2
-            exit 1
-          fi
           echo "${GCP_SA_KEY}" > /tmp/gcp-key.json
           apt-get install -y -qq apt-transport-https ca-certificates curl
           echo "deb [trusted=yes] https://packages.cloud.google.com/apt cloud-sdk main" > /etc/apt/sources.list.d/google-cloud-sdk.list
           apt-get update -qq && apt-get install -y google-cloud-cli
           gcloud auth activate-service-account --key-file=/tmp/gcp-key.json
           gcloud config set project neuron-785695
-          echo "Publishing as active account: $(gcloud config get-value account 2>/dev/null)"
 
           VERSION="${GITHUB_SHA:0:8}"
 
@@ -354,12 +345,6 @@ jobs:
         # Patches ci-base:latest in-place: pulls the existing image (which has all
         # system deps — Node, Go, gcloud, Docker CLI, etc.) and overlays the freshly
         # built El SDK on top. Keeps the full ci-base rebuild fast and incremental.
-        #
-        # continue-on-error: this is a CI-cache optimization, NOT the release
-        # artifact. It runs Docker (pull/build/push ~600MB) on the host-mode GCE
-        # runner where DinD/Docker availability is fragile. A failure here must
-        # never block or redden the job — the SDK publish above is the deliverable.
-        continue-on-error: true
         if: github.event_name == 'push'
         env:
           GCP_SA_KEY: ${{ secrets.GCP_SA_KEY }}

engram/.gitea/workflows/ci-dev.yaml

@@ -81,7 +81,7 @@ jobs:
       # Link to produce the engram binary
       - name: Link engram binary
         run: |
-          cc -std=c11 -O2 -DHAVE_CURL \
+          cc -std=c11 -O2 \
             -I /usr/local/lib/el \
             -o dist/engram \
             dist/engram.c \

engram/.gitea/workflows/ci-stage.yaml

@@ -88,7 +88,7 @@ jobs:
       # Link to produce the engram binary
       - name: Link engram binary
         run: |
-          cc -std=c11 -O2 -DHAVE_CURL \
+          cc -std=c11 -O2 \
             -I /usr/local/lib/el \
             -o dist/engram \
             dist/engram.c \

engram/.gitea/workflows/engram-release.yaml

@@ -62,7 +62,7 @@ jobs:
       # Link to produce the engram binary
       - name: Link engram binary
         run: |
-          cc -std=c11 -O2 -DHAVE_CURL \
+          cc -std=c11 -O2 \
             -I /usr/local/lib/el \
             -o dist/engram \
             dist/engram.c \

engram/dist/engram.c

@@ -10,6 +10,7 @@ el_val_t query_param(el_val_t path, el_val_t key);
 el_val_t query_int(el_val_t path, el_val_t key, el_val_t default_val);
 el_val_t extract_id(el_val_t path, el_val_t prefix);
 el_val_t route_stats(el_val_t method, el_val_t path, el_val_t body);
+el_val_t persist_canonical(void);
 el_val_t route_create_node(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_get_node(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_scan_nodes(el_val_t method, el_val_t path, el_val_t body);
@@ -20,18 +21,23 @@ el_val_t route_create_edge(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_neighbors(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_strengthen(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_forget(el_val_t method, el_val_t path, el_val_t body);
-el_val_t route_create_ise(el_val_t method, el_val_t path, el_val_t body);
-el_val_t route_sync(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_save(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_load(el_val_t method, el_val_t path, el_val_t body);
 el_val_t route_health(el_val_t method, el_val_t path, el_val_t body);
+el_val_t route_sync(el_val_t method, el_val_t path, el_val_t body);
+el_val_t route_load_merge(el_val_t method, el_val_t path, el_val_t body);
+el_val_t route_emit_ise(el_val_t method, el_val_t path, el_val_t body);
+el_val_t route_capture_knowledge(el_val_t method, el_val_t path, el_val_t body);
 el_val_t check_auth_ok(el_val_t method, el_val_t body);
 el_val_t handle_request(el_val_t method, el_val_t path, el_val_t body);
 
+el_val_t bind_raw;
 el_val_t bind_str;
 el_val_t port;
+el_val_t data_dir_raw;
 el_val_t data_dir;
 el_val_t snapshot_path;
+el_val_t boot_snap;
 
 el_val_t parse_port(el_val_t bind) {
   el_val_t colon = str_index_of(bind, EL_STR(":"));
@@ -110,17 +116,22 @@ el_val_t route_stats(el_val_t method, el_val_t path, el_val_t body) {
   return 0;
 }
 
+el_val_t persist_canonical(void) {
+  el_val_t dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
+  el_val_t dir = ({ el_val_t _if_result_1 = 0; if (str_eq(dir_raw, EL_STR(""))) { _if_result_1 = (EL_STR("/tmp/engram")); } else { _if_result_1 = (dir_raw); } _if_result_1; });
+  engram_save(el_str_concat(dir, EL_STR("/snapshot.json")));
+  return 1;
+  return 0;
+}
+
 el_val_t route_create_node(el_val_t method, el_val_t path, el_val_t body) {
   el_val_t content = json_get_string(body, EL_STR("content"));
-  el_val_t node_type = json_get_string(body, EL_STR("node_type"));
-  if (str_eq(node_type, EL_STR(""))) {
-    node_type = EL_STR("Memory");
-  }
-  el_val_t salience = json_get_float(body, EL_STR("salience"));
-  if (salience == el_from_float(0.0)) {
-    salience = el_from_float(0.5);
-  }
+  el_val_t nt_raw = json_get_string(body, EL_STR("node_type"));
+  el_val_t node_type = ({ el_val_t _if_result_2 = 0; if (str_eq(nt_raw, EL_STR(""))) { _if_result_2 = (EL_STR("Memory")); } else { _if_result_2 = (nt_raw); } _if_result_2; });
+  el_val_t sal_raw = json_get_float(body, EL_STR("salience"));
+  el_val_t salience = ({ el_val_t _if_result_3 = 0; if ((sal_raw == el_from_float(0.0))) { _if_result_3 = (el_from_float(0.5)); } else { _if_result_3 = (sal_raw); } _if_result_3; });
   el_val_t id = engram_node(content, node_type, salience);
+  el_val_t saved = persist_canonical();
   return el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(EL_STR("{\"id\":\""), id), EL_STR("\",\"content\":\"")), content), EL_STR("\",\"node_type\":\"")), node_type), EL_STR("\"}"));
   return 0;
 }
@@ -146,11 +157,9 @@ el_val_t route_scan_nodes(el_val_t method, el_val_t path, el_val_t body) {
 }
 
 el_val_t route_scan_edges(el_val_t method, el_val_t path, el_val_t body) {
-  el_val_t dir = env(EL_STR("ENGRAM_DATA_DIR"));
-  if (str_eq(dir, EL_STR(""))) {
-    dir = EL_STR("/tmp/engram");
-  }
-  el_val_t snap_path = el_str_concat(dir, EL_STR("/snapshot.json"));
+  el_val_t dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
+  el_val_t dir = ({ el_val_t _if_result_4 = 0; if (str_eq(dir_raw, EL_STR(""))) { _if_result_4 = (EL_STR("/tmp/engram")); } else { _if_result_4 = (dir_raw); } _if_result_4; });
+  el_val_t snap_path = el_str_concat(dir, EL_STR("/.scan-export.json"));
   engram_save(snap_path);
   el_val_t snap = fs_read(snap_path);
   if (str_eq(snap, EL_STR(""))) {
@@ -165,36 +174,22 @@ el_val_t route_scan_edges(el_val_t method, el_val_t path, el_val_t body) {
 }
 
 el_val_t route_search(el_val_t method, el_val_t path, el_val_t body) {
-  el_val_t q = EL_STR("");
-  if (str_eq(method, EL_STR("GET"))) {
-    q = query_param(path, EL_STR("q"));
-  } else {
-    q = json_get_string(body, EL_STR("query"));
-  }
-  el_val_t limit = query_int(path, EL_STR("limit"), 20);
-  if (limit == 0) {
-    limit = json_get_int(body, EL_STR("limit"));
-  }
-  if (limit == 0) {
-    limit = 20;
-  }
+  el_val_t q = ({ el_val_t _if_result_5 = 0; if (str_eq(method, EL_STR("GET"))) { _if_result_5 = (query_param(path, EL_STR("q"))); } else { _if_result_5 = (json_get_string(body, EL_STR("query"))); } _if_result_5; });
+  el_val_t lim_url = query_int(path, EL_STR("limit"), 0);
+  el_val_t lim_body = json_get_int(body, EL_STR("limit"));
+  el_val_t lim_either = ({ el_val_t _if_result_6 = 0; if ((lim_url > 0)) { _if_result_6 = (lim_url); } else { _if_result_6 = (lim_body); } _if_result_6; });
+  el_val_t limit = ({ el_val_t _if_result_7 = 0; if ((lim_either > 0)) { _if_result_7 = (lim_either); } else { _if_result_7 = (20); } _if_result_7; });
   return engram_search_json(q, limit);
   return 0;
 }
 
 el_val_t route_activate(el_val_t method, el_val_t path, el_val_t body) {
-  el_val_t q = EL_STR("");
-  el_val_t depth = 3;
-  if (str_eq(method, EL_STR("GET"))) {
-    q = query_param(path, EL_STR("q"));
-    depth = query_int(path, EL_STR("depth"), 3);
-  } else {
-    q = json_get_string(body, EL_STR("query"));
-    el_val_t bd = json_get_int(body, EL_STR("depth"));
-    if (bd > 0) {
-      depth = bd;
-    }
+  el_val_t q = ({ el_val_t _if_result_8 = 0; if (str_eq(method, EL_STR("GET"))) { _if_result_8 = (query_param(path, EL_STR("q"))); } else { _if_result_8 = (json_get_string(body, EL_STR("query"))); } _if_result_8; });
+  if (str_eq(q, EL_STR(""))) {
+    return err_json(EL_STR("missing query"));
   }
+  el_val_t d_raw = ({ el_val_t _if_result_9 = 0; if (str_eq(method, EL_STR("GET"))) { _if_result_9 = (query_int(path, EL_STR("depth"), 3)); } else { _if_result_9 = (json_get_int(body, EL_STR("depth"))); } _if_result_9; });
+  el_val_t depth = ({ el_val_t _if_result_10 = 0; if ((d_raw > 0)) { _if_result_10 = (d_raw); } else { _if_result_10 = (3); } _if_result_10; });
   return el_str_concat(el_str_concat(EL_STR("{\"results\":"), engram_activate_json(q, depth)), EL_STR("}"));
   return 0;
 }
@@ -202,15 +197,12 @@ el_val_t route_activate(el_val_t method, el_val_t path, el_val_t body) {
 el_val_t route_create_edge(el_val_t method, el_val_t path, el_val_t body) {
   el_val_t from_id = json_get_string(body, EL_STR("from_id"));
   el_val_t to_id = json_get_string(body, EL_STR("to_id"));
-  el_val_t relation = json_get_string(body, EL_STR("relation"));
-  if (str_eq(relation, EL_STR(""))) {
-    relation = EL_STR("associates");
-  }
-  el_val_t weight = json_get_float(body, EL_STR("weight"));
-  if (weight == el_from_float(0.0)) {
-    weight = el_from_float(0.5);
-  }
+  el_val_t rel_raw = json_get_string(body, EL_STR("relation"));
+  el_val_t relation = ({ el_val_t _if_result_11 = 0; if (str_eq(rel_raw, EL_STR(""))) { _if_result_11 = (EL_STR("associates")); } else { _if_result_11 = (rel_raw); } _if_result_11; });
+  el_val_t w_raw = json_get_float(body, EL_STR("weight"));
+  el_val_t weight = ({ el_val_t _if_result_12 = 0; if ((w_raw == el_from_float(0.0))) { _if_result_12 = (el_from_float(0.5)); } else { _if_result_12 = (w_raw); } _if_result_12; });
   engram_connect(from_id, to_id, weight, relation);
+  el_val_t saved = persist_canonical();
   return el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"from_id\":\""), from_id), EL_STR("\",\"to_id\":\"")), to_id), EL_STR("\",\"relation\":\"")), relation), EL_STR("\"}"));
   return 0;
 }
@@ -231,6 +223,7 @@ el_val_t route_strengthen(el_val_t method, el_val_t path, el_val_t body) {
     return err_json(EL_STR("missing node_id"));
   }
   engram_strengthen(id);
+  el_val_t saved = persist_canonical();
   return ok_json();
   return 0;
 }
@@ -241,29 +234,40 @@ el_val_t route_forget(el_val_t method, el_val_t path, el_val_t body) {
     return err_json(EL_STR("missing id"));
   }
   engram_forget(id);
+  el_val_t saved = persist_canonical();
   return ok_json();
   return 0;
 }
 
-el_val_t route_create_ise(el_val_t method, el_val_t path, el_val_t body) {
-  el_val_t content = json_get_string(body, EL_STR("content"));
-  if (str_eq(content, EL_STR(""))) {
-    return err_json(EL_STR("missing content"));
-  }
-  el_val_t sal = el_from_float(0.3);
-  el_val_t imp = el_from_float(0.3);
-  el_val_t conf = el_from_float(0.8);
-  el_val_t id = engram_node_full(content, EL_STR("InternalStateEvent"), EL_STR("state-event"), sal, imp, conf, EL_STR("Episodic"), EL_STR("[\"internal-state\",\"InternalStateEvent\"]"));
-  return el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"id\":\""), id), EL_STR("\"}"));
+el_val_t route_save(el_val_t method, el_val_t path, el_val_t body) {
+  el_val_t p_raw = json_get_string(body, EL_STR("path"));
+  el_val_t dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
+  el_val_t dir = ({ el_val_t _if_result_13 = 0; if (str_eq(dir_raw, EL_STR(""))) { _if_result_13 = (EL_STR("/tmp/engram")); } else { _if_result_13 = (dir_raw); } _if_result_13; });
+  el_val_t p = ({ el_val_t _if_result_14 = 0; if (str_eq(p_raw, EL_STR(""))) { _if_result_14 = (el_str_concat(dir, EL_STR("/snapshot.json"))); } else { _if_result_14 = (p_raw); } _if_result_14; });
+  engram_save(p);
+  return el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"path\":\""), p), EL_STR("\"}"));
+  return 0;
+}
+
+el_val_t route_load(el_val_t method, el_val_t path, el_val_t body) {
+  el_val_t p_raw = json_get_string(body, EL_STR("path"));
+  el_val_t dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
+  el_val_t dir = ({ el_val_t _if_result_15 = 0; if (str_eq(dir_raw, EL_STR(""))) { _if_result_15 = (EL_STR("/tmp/engram")); } else { _if_result_15 = (dir_raw); } _if_result_15; });
+  el_val_t p = ({ el_val_t _if_result_16 = 0; if (str_eq(p_raw, EL_STR(""))) { _if_result_16 = (el_str_concat(dir, EL_STR("/snapshot.json"))); } else { _if_result_16 = (p_raw); } _if_result_16; });
+  engram_load(p);
+  return ok_json();
+  return 0;
+}
+
+el_val_t route_health(el_val_t method, el_val_t path, el_val_t body) {
+  return EL_STR("{\"status\":\"ok\",\"engine\":\"engram-runtime-native\"}");
   return 0;
 }
 
 el_val_t route_sync(el_val_t method, el_val_t path, el_val_t body) {
-  el_val_t dir = env(EL_STR("ENGRAM_DATA_DIR"));
-  if (str_eq(dir, EL_STR(""))) {
-    dir = EL_STR("/tmp/engram");
-  }
-  el_val_t snap_path = el_str_concat(dir, EL_STR("/sync-export.json"));
+  el_val_t dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
+  el_val_t dir = ({ el_val_t _if_result_17 = 0; if (str_eq(dir_raw, EL_STR(""))) { _if_result_17 = (EL_STR("/tmp/engram")); } else { _if_result_17 = (dir_raw); } _if_result_17; });
+  el_val_t snap_path = el_str_concat(dir, EL_STR("/.sync-export.json"));
   engram_save(snap_path);
   el_val_t snap = fs_read(snap_path);
   if (str_eq(snap, EL_STR(""))) {
@@ -273,36 +277,68 @@ el_val_t route_sync(el_val_t method, el_val_t path, el_val_t body) {
   return 0;
 }
 
-el_val_t route_save(el_val_t method, el_val_t path, el_val_t body) {
+el_val_t route_load_merge(el_val_t method, el_val_t path, el_val_t body) {
   el_val_t p = json_get_string(body, EL_STR("path"));
   if (str_eq(p, EL_STR(""))) {
-    el_val_t dir = env(EL_STR("ENGRAM_DATA_DIR"));
-    if (str_eq(dir, EL_STR(""))) {
-      dir = EL_STR("/tmp/engram");
-    }
-    p = el_str_concat(dir, EL_STR("/snapshot.json"));
+    return err_json(EL_STR("path is required"));
   }
-  engram_save(p);
-  return el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"path\":\""), p), EL_STR("\"}"));
+  if (str_eq(fs_read(p), EL_STR(""))) {
+    return err_json(EL_STR("file missing or empty"));
+  }
+  el_val_t before_n = engram_node_count();
+  el_val_t before_e = engram_edge_count();
+  engram_load_merge(p);
+  el_val_t added_n = (engram_node_count() - before_n);
+  el_val_t added_e = (engram_edge_count() - before_e);
+  el_val_t saved = persist_canonical();
+  return el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"nodes_added\":"), int_to_str(added_n)), EL_STR(",\"edges_added\":")), int_to_str(added_e)), EL_STR(",\"node_count\":")), int_to_str(engram_node_count())), EL_STR("}"));
   return 0;
 }
 
-el_val_t route_load(el_val_t method, el_val_t path, el_val_t body) {
-  el_val_t p = json_get_string(body, EL_STR("path"));
-  if (str_eq(p, EL_STR(""))) {
-    el_val_t dir = env(EL_STR("ENGRAM_DATA_DIR"));
-    if (str_eq(dir, EL_STR(""))) {
-      dir = EL_STR("/tmp/engram");
-    }
-    p = el_str_concat(dir, EL_STR("/snapshot.json"));
+el_val_t route_emit_ise(el_val_t method, el_val_t path, el_val_t body) {
+  el_val_t content = json_get_string(body, EL_STR("content"));
+  if (str_eq(content, EL_STR(""))) {
+    return err_json(EL_STR("missing content"));
   }
-  engram_load(p);
-  return ok_json();
+  el_val_t sal = el_from_float(0.3);
+  el_val_t imp = el_from_float(0.3);
+  el_val_t conf = el_from_float(0.8);
+  el_val_t id = engram_node_full(content, EL_STR("InternalStateEvent"), EL_STR("state-event"), sal, imp, conf, EL_STR("Episodic"), EL_STR("[\"internal-state\",\"InternalStateEvent\"]"));
+  el_val_t ret_raw = env(EL_STR("ENGRAM_ISE_RETENTION_MS"));
+  el_val_t ret_ms = ({ el_val_t _if_result_18 = 0; if (str_eq(ret_raw, EL_STR(""))) { _if_result_18 = (172800000); } else { _if_result_18 = (str_to_int(ret_raw)); } _if_result_18; });
+  el_val_t pruned = engram_prune_telemetry(ret_ms);
+  return el_str_concat(el_str_concat(el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"id\":\""), id), EL_STR("\",\"pruned\":")), int_to_str(pruned)), EL_STR("}"));
   return 0;
 }
 
-el_val_t route_health(el_val_t method, el_val_t path, el_val_t body) {
-  return EL_STR("{\"status\":\"ok\",\"engine\":\"engram-runtime-native\"}");
+el_val_t route_capture_knowledge(el_val_t method, el_val_t path, el_val_t body) {
+  el_val_t content = json_get_string(body, EL_STR("content"));
+  if (str_eq(content, EL_STR(""))) {
+    return err_json(EL_STR("missing content"));
+  }
+  el_val_t title = json_get_string(body, EL_STR("title"));
+  el_val_t label = ({ el_val_t _if_result_19 = 0; if (str_eq(title, EL_STR(""))) { _if_result_19 = (str_slice(content, 0, 60)); } else { _if_result_19 = (title); } _if_result_19; });
+  el_val_t category_raw = json_get_string(body, EL_STR("category"));
+  el_val_t category = ({ el_val_t _if_result_20 = 0; if (str_eq(category_raw, EL_STR(""))) { _if_result_20 = (EL_STR("other")); } else { _if_result_20 = (category_raw); } _if_result_20; });
+  el_val_t ktier_raw = json_get_string(body, EL_STR("tier"));
+  el_val_t ktier = ({ el_val_t _if_result_21 = 0; if (str_eq(ktier_raw, EL_STR(""))) { _if_result_21 = (EL_STR("note")); } else { _if_result_21 = (ktier_raw); } _if_result_21; });
+  el_val_t project = json_get_string(body, EL_STR("project"));
+  el_val_t tags_raw = json_get_raw(body, EL_STR("tags"));
+  el_val_t tags_base = ({ el_val_t _if_result_22 = 0; if (str_eq(tags_raw, EL_STR(""))) { _if_result_22 = (EL_STR("[]")); } else { _if_result_22 = (tags_raw); } _if_result_22; });
+  el_val_t base_len = str_len(tags_base);
+  el_val_t head = str_slice(tags_base, 0, (base_len - 1));
+  el_val_t sep = ({ el_val_t _if_result_23 = 0; if (str_eq(head, EL_STR("["))) { _if_result_23 = (EL_STR("")); } else { _if_result_23 = (EL_STR(",")); } _if_result_23; });
+  el_val_t safe_cat = str_replace(category, EL_STR("\""), EL_STR("'"));
+  el_val_t safe_tier = str_replace(ktier, EL_STR("\""), EL_STR("'"));
+  el_val_t safe_proj = str_replace(project, EL_STR("\""), EL_STR("'"));
+  el_val_t proj_tag = ({ el_val_t _if_result_24 = 0; if (str_eq(safe_proj, EL_STR(""))) { _if_result_24 = (EL_STR("")); } else { _if_result_24 = (el_str_concat(el_str_concat(EL_STR(",\"project:"), safe_proj), EL_STR("\""))); } _if_result_24; });
+  el_val_t tags = el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(head, sep), EL_STR("\"category:")), safe_cat), EL_STR("\",\"tier:")), safe_tier), EL_STR("\"")), proj_tag), EL_STR("]"));
+  el_val_t sal = el_from_float(0.5);
+  el_val_t imp = el_from_float(0.5);
+  el_val_t conf = el_from_float(0.9);
+  el_val_t id = engram_node_full(content, EL_STR("Knowledge"), label, sal, imp, conf, EL_STR("Semantic"), tags);
+  el_val_t saved = persist_canonical();
+  return el_str_concat(el_str_concat(EL_STR("{\"ok\":true,\"id\":\""), id), EL_STR("\"}"));
   return 0;
 }
 
@@ -329,12 +365,15 @@ el_val_t handle_request(el_val_t method, el_val_t path, el_val_t body) {
       return route_health(method, path, body);
     }
   }
-  if (str_eq(method, EL_STR("POST")) && str_starts_with(clean, EL_STR("/api/neuron/state-events"))) {
-    return route_create_ise(method, path, body);
+  if (str_eq(method, EL_STR("POST")) && str_eq(clean, EL_STR("/api/neuron/state-events"))) {
+    return route_emit_ise(method, path, body);
   }
   if (!check_auth_ok(method, body)) {
     return err_json(EL_STR("unauthorized"));
   }
+  if (str_eq(method, EL_STR("POST")) && str_eq(clean, EL_STR("/api/neuron/knowledge/capture"))) {
+    return route_capture_knowledge(method, path, body);
+  }
   if (str_eq(method, EL_STR("GET")) && (str_eq(clean, EL_STR("/api/stats")) || str_eq(clean, EL_STR("/stats")))) {
     return route_stats(method, path, body);
   }
@@ -374,32 +413,40 @@ el_val_t handle_request(el_val_t method, el_val_t path, el_val_t body) {
   if (str_eq(method, EL_STR("POST")) && (str_eq(clean, EL_STR("/api/strengthen")) || str_eq(clean, EL_STR("/strengthen")))) {
     return route_strengthen(method, path, body);
   }
-  if (str_eq(method, EL_STR("GET")) && (str_eq(clean, EL_STR("/api/sync")) || str_eq(clean, EL_STR("/sync")))) {
-    return route_sync(method, path, body);
-  }
   if (str_eq(method, EL_STR("POST")) && (str_eq(clean, EL_STR("/api/save")) || str_eq(clean, EL_STR("/save")))) {
     return route_save(method, path, body);
   }
   if (str_eq(method, EL_STR("POST")) && (str_eq(clean, EL_STR("/api/load")) || str_eq(clean, EL_STR("/load")))) {
     return route_load(method, path, body);
   }
+  if (str_eq(method, EL_STR("POST")) && (str_eq(clean, EL_STR("/api/load-merge")) || str_eq(clean, EL_STR("/load-merge")))) {
+    return route_load_merge(method, path, body);
+  }
+  if (str_eq(method, EL_STR("GET")) && str_eq(clean, EL_STR("/api/sync"))) {
+    return route_sync(method, path, body);
+  }
   return el_str_concat(el_str_concat(EL_STR("{\"error\":\"not found\",\"path\":\""), clean), EL_STR("\"}"));
   return 0;
 }
 
 int main(int _argc, char** _argv) {
   el_runtime_init_args(_argc, _argv);
-  bind_str = env(EL_STR("ENGRAM_BIND"));
-  if (str_eq(bind_str, EL_STR(""))) {
-    bind_str = EL_STR(":8742");
-  }
+  bind_raw = env(EL_STR("ENGRAM_BIND"));
+  bind_str = ({ el_val_t _if_result_25 = 0; if (str_eq(bind_raw, EL_STR(""))) { _if_result_25 = (EL_STR(":8742")); } else { _if_result_25 = (bind_raw); } _if_result_25; });
   port = parse_port(bind_str);
-  data_dir = env(EL_STR("ENGRAM_DATA_DIR"));
-  if (str_eq(data_dir, EL_STR(""))) {
-    data_dir = EL_STR("/tmp/engram");
-  }
+  data_dir_raw = env(EL_STR("ENGRAM_DATA_DIR"));
+  data_dir = ({ el_val_t _if_result_26 = 0; if (str_eq(data_dir_raw, EL_STR(""))) { _if_result_26 = (EL_STR("/tmp/engram")); } else { _if_result_26 = (data_dir_raw); } _if_result_26; });
   snapshot_path = el_str_concat(data_dir, EL_STR("/snapshot.json"));
   engram_load(snapshot_path);
+  boot_snap = fs_read(snapshot_path);
+  if (!str_eq(boot_snap, EL_STR(""))) {
+    if (engram_node_count() == 0) {
+      println(EL_STR("[engram] WARNING: snapshot.json is non-empty but load produced 0 nodes \xe2\x80\x94 preserving copy at snapshot.failed-load.json"));
+      fs_write(el_str_concat(data_dir, EL_STR("/snapshot.failed-load.json")), boot_snap);
+    } else {
+      fs_write(el_str_concat(data_dir, EL_STR("/snapshot.boot-backup.json")), boot_snap);
+    }
+  }
   println(EL_STR("[engram] runtime-native graph engine"));
   println(el_str_concat(EL_STR("[engram] data_dir="), data_dir));
   println(el_str_concat(EL_STR("[engram] node_count="), int_to_str(engram_node_count())));

engram/src/server.el

@@ -76,13 +76,43 @@ fn route_stats(method: String, path: String, body: String) -> String {
     engram_stats_json()
 }
 
+// (2026-07-18 self-review) Scoping sweep: `let` inside an if-block creates an
+// inner scope only — it does NOT mutate the outer binding (documented with
+// evidence in awareness.el, 2026-05-25). Every default/reassignment below used
+// that broken pattern, so defaults never applied: nodes were created with
+// node_type="" and salience=0.0, /api/search and /api/activate ALWAYS ran with
+// q="" regardless of input, edges defaulted to relation=""/weight=0.0, and
+// save/load with no "path" hit engram_save(""). Rewritten to the
+// `let x = if cond { a } else { b }` expression form (the pattern the newer
+// routes route_emit_ise/route_capture_knowledge already use correctly).
+// persist_canonical — save the canonical snapshot after a durable write.
+//
+// WHY (2026-07-22 self-review): the 2026-07-21 fix correctly stopped READ
+// routes from writing the canonical snapshot.json — but nothing was left
+// that saved it on WRITE. Every mutation (node create, edge create,
+// knowledge capture, forget, merge) lived only in RAM until someone POSTed
+// /api/save manually; a process restart silently discarded everything since
+// the last manual save. Observed live: two engram restarts during the
+// 2026-07-22 review reverted the store to a ~17h-old snapshot, destroying
+// same-day writes. Reads must never write the canonical; writes must always
+// persist it. ISE telemetry is deliberately excluded (48h-pruned, loss-
+// tolerant, ~2/min — snapshotting the whole store per heartbeat is waste;
+// any durable write that follows persists the pruning too).
+fn persist_canonical() -> Int {
+    let dir_raw: String = env("ENGRAM_DATA_DIR")
+    let dir: String = if str_eq(dir_raw, "") { "/tmp/engram" } else { dir_raw }
+    engram_save(dir + "/snapshot.json")
+    return 1
+}
+
 fn route_create_node(method: String, path: String, body: String) -> String {
     let content: String = json_get_string(body, "content")
-    let node_type: String = json_get_string(body, "node_type")
-    if str_eq(node_type, "") { let node_type = "Memory" }
-    let salience: Float = json_get_float(body, "salience")
-    if salience == 0.0 { let salience = 0.5 }
+    let nt_raw: String = json_get_string(body, "node_type")
+    let node_type: String = if str_eq(nt_raw, "") { "Memory" } else { nt_raw }
+    let sal_raw: Float = json_get_float(body, "salience")
+    let salience: Float = if sal_raw == 0.0 { 0.5 } else { sal_raw }
     let id: String = engram_node(content, node_type, salience)
+    let saved: Int = persist_canonical()
     "{\"id\":\"" + id + "\",\"content\":\"" + content + "\",\"node_type\":\"" + node_type + "\"}"
 }
 
@@ -103,13 +133,14 @@ fn route_scan_nodes(method: String, path: String, body: String) -> String {
 }
 
 // route_scan_edges — bulk export of all edges as a JSON array. Implemented
-// via engram_save → fs_read of the canonical on-disk snapshot, which the
-// runtime keeps in lockstep with the in-memory graph. Live against the
-// running graph, not a stale export.
+// via engram_save → fs_read of a SCRATCH export path. (2026-07-21 self-review:
+// previously this saved over the canonical snapshot.json on every GET — if the
+// process ever booted with a partial/empty store, the first read request
+// clobbered the good snapshot. Read routes must never write the canonical path.)
 fn route_scan_edges(method: String, path: String, body: String) -> String {
-    let dir: String = env("ENGRAM_DATA_DIR")
-    if str_eq(dir, "") { let dir = "/tmp/engram" }
-    let snap_path: String = dir + "/snapshot.json"
+    let dir_raw: String = env("ENGRAM_DATA_DIR")
+    let dir: String = if str_eq(dir_raw, "") { "/tmp/engram" } else { dir_raw }
+    let snap_path: String = dir + "/.scan-export.json"
     engram_save(snap_path)
     let snap: String = fs_read(snap_path)
     if str_eq(snap, "") { return "[]" }
@@ -122,40 +153,34 @@ fn route_scan_edges(method: String, path: String, body: String) -> String {
 }
 
 fn route_search(method: String, path: String, body: String) -> String {
-    let q: String = ""
-    if str_eq(method, "GET") {
-        let q = query_param(path, "q")
-    } else {
-        let q = json_get_string(body, "query")
-    }
-    let limit: Int = query_int(path, "limit", 20)
-    if limit == 0 { let limit = json_get_int(body, "limit") }
-    if limit == 0 { let limit = 20 }
+    let q: String = if str_eq(method, "GET") { query_param(path, "q") } else { json_get_string(body, "query") }
+    let lim_url: Int = query_int(path, "limit", 0)
+    let lim_body: Int = json_get_int(body, "limit")
+    let lim_either: Int = if lim_url > 0 { lim_url } else { lim_body }
+    let limit: Int = if lim_either > 0 { lim_either } else { 20 }
     return engram_search_json(q, limit)
 }
 
 fn route_activate(method: String, path: String, body: String) -> String {
-    let q: String = ""
-    let depth: Int = 3
-    if str_eq(method, "GET") {
-        let q = query_param(path, "q")
-        let depth = query_int(path, "depth", 3)
-    } else {
-        let q = json_get_string(body, "query")
-        let bd: Int = json_get_int(body, "depth")
-        if bd > 0 { let depth = bd }
-    }
+    let q: String = if str_eq(method, "GET") { query_param(path, "q") } else { json_get_string(body, "query") }
+    // Guard: engram_activate with an empty query matches zero seeds, which
+    // zeroes ALL carried working-memory weights (documented in awareness.el
+    // perceive()). Never let an empty activation through to wipe WM.
+    if str_eq(q, "") { return err_json("missing query") }
+    let d_raw: Int = if str_eq(method, "GET") { query_int(path, "depth", 3) } else { json_get_int(body, "depth") }
+    let depth: Int = if d_raw > 0 { d_raw } else { 3 }
     return "{\"results\":" + engram_activate_json(q, depth) + "}"
 }
 
 fn route_create_edge(method: String, path: String, body: String) -> String {
     let from_id: String = json_get_string(body, "from_id")
     let to_id: String = json_get_string(body, "to_id")
-    let relation: String = json_get_string(body, "relation")
-    if str_eq(relation, "") { let relation = "associates" }
-    let weight: Float = json_get_float(body, "weight")
-    if weight == 0.0 { let weight = 0.5 }
+    let rel_raw: String = json_get_string(body, "relation")
+    let relation: String = if str_eq(rel_raw, "") { "associates" } else { rel_raw }
+    let w_raw: Float = json_get_float(body, "weight")
+    let weight: Float = if w_raw == 0.0 { 0.5 } else { w_raw }
     engram_connect(from_id, to_id, weight, relation)
+    let saved: Int = persist_canonical()
     "{\"ok\":true,\"from_id\":\"" + from_id + "\",\"to_id\":\"" + to_id + "\",\"relation\":\"" + relation + "\"}"
 }
 
@@ -170,6 +195,7 @@ fn route_strengthen(method: String, path: String, body: String) -> String {
     let id: String = json_get_string(body, "node_id")
     if str_eq(id, "") { return err_json("missing node_id") }
     engram_strengthen(id)
+    let saved: Int = persist_canonical()
     ok_json()
 }
 
@@ -177,27 +203,24 @@ fn route_forget(method: String, path: String, body: String) -> String {
     let id: String = extract_id(path, "/api/nodes/")
     if str_eq(id, "") { return err_json("missing id") }
     engram_forget(id)
+    let saved: Int = persist_canonical()
     ok_json()
 }
 
 fn route_save(method: String, path: String, body: String) -> String {
-    let p: String = json_get_string(body, "path")
-    if str_eq(p, "") {
-        let dir: String = env("ENGRAM_DATA_DIR")
-        if str_eq(dir, "") { let dir = "/tmp/engram" }
-        let p = dir + "/snapshot.json"
-    }
+    let p_raw: String = json_get_string(body, "path")
+    let dir_raw: String = env("ENGRAM_DATA_DIR")
+    let dir: String = if str_eq(dir_raw, "") { "/tmp/engram" } else { dir_raw }
+    let p: String = if str_eq(p_raw, "") { dir + "/snapshot.json" } else { p_raw }
     engram_save(p)
     "{\"ok\":true,\"path\":\"" + p + "\"}"
 }
 
 fn route_load(method: String, path: String, body: String) -> String {
-    let p: String = json_get_string(body, "path")
-    if str_eq(p, "") {
-        let dir: String = env("ENGRAM_DATA_DIR")
-        if str_eq(dir, "") { let dir = "/tmp/engram" }
-        let p = dir + "/snapshot.json"
-    }
+    let p_raw: String = json_get_string(body, "path")
+    let dir_raw: String = env("ENGRAM_DATA_DIR")
+    let dir: String = if str_eq(dir_raw, "") { "/tmp/engram" } else { dir_raw }
+    let p: String = if str_eq(p_raw, "") { dir + "/snapshot.json" } else { p_raw }
     engram_load(p)
     ok_json()
 }
@@ -219,15 +242,36 @@ fn route_health(method: String, path: String, body: String) -> String {
 // (it skips nodes already present by ID). Auth-exempt: same-host internal call.
 // (2026-06-27 self-review: added this route to fix silent 10-min sync failures)
 fn route_sync(method: String, path: String, body: String) -> String {
-    let dir: String = env("ENGRAM_DATA_DIR")
-    if str_eq(dir, "") { let dir = "/tmp/engram" }
-    let snap_path: String = dir + "/snapshot.json"
+    let dir_raw: String = env("ENGRAM_DATA_DIR")
+    let dir: String = if str_eq(dir_raw, "") { "/tmp/engram" } else { dir_raw }
+    // 2026-07-21 self-review: export to a scratch path, never the canonical
+    // snapshot.json — read routes must not be able to clobber the good snapshot.
+    let snap_path: String = dir + "/.sync-export.json"
     engram_save(snap_path)
     let snap: String = fs_read(snap_path)
     if str_eq(snap, "") { return "{\"nodes\":[],\"edges\":[]}" }
     return snap
 }
 
+// route_load_merge — POST /api/load-merge {"path": "..."} — merge a snapshot
+// file into the live store WITHOUT resetting it (engram_load_merge skips nodes
+// already present by id). Added 2026-07-21 self-review to restore the 244 kn-
+// identity Knowledge nodes lost from the snapshot lineage between 05-13 and
+// 07-13. Requires an explicit path: refuses to run without one so it can never
+// be triggered accidentally against a default.
+fn route_load_merge(method: String, path: String, body: String) -> String {
+    let p: String = json_get_string(body, "path")
+    if str_eq(p, "") { return err_json("path is required") }
+    if str_eq(fs_read(p), "") { return err_json("file missing or empty") }
+    let before_n: Int = engram_node_count()
+    let before_e: Int = engram_edge_count()
+    engram_load_merge(p)
+    let added_n: Int = engram_node_count() - before_n
+    let added_e: Int = engram_edge_count() - before_e
+    let saved: Int = persist_canonical()
+    "{\"ok\":true,\"nodes_added\":" + int_to_str(added_n) + ",\"edges_added\":" + int_to_str(added_e) + ",\"node_count\":" + int_to_str(engram_node_count()) + "}"
+}
+
 // route_emit_ise — write an InternalStateEvent node from the soul daemon.
 //
 // Endpoint: POST /api/neuron/state-events
@@ -241,10 +285,20 @@ fn route_sync(method: String, path: String, body: String) -> String {
 //
 // Salience/importance set to match engram_node_full ISE defaults used by the
 // in-process fallback path in awareness.el (salience=0.3, importance=0.3,
-// confidence=0.8, tier=Episodic). High temporal_decay_rate (1.617) — ISEs
-// are inherently transient; they should decay faster than structural knowledge.
+// confidence=0.8, tier=Episodic).
 // (2026-06-26 self-review: added this route after discovering ise_post was
 // silently failing — the soul posts here but the endpoint didn't exist.)
+//
+// Retention (2026-07-16 self-review): an earlier comment here claimed ISEs
+// got temporal_decay_rate=1.617 — that was never implemented (engram_node_full
+// hardcodes 0.0), and per-node decay only dampens activation anyway; it never
+// removes nodes. By 2026-07-16 ISEs were 75% of the store (10,175 of 13,522
+// nodes, ~4,300/day, unbounded). ISEs are already WM-excluded in
+// engram_activate, so the fix is retention, not decay: every insert calls
+// engram_prune_telemetry(), a single O(nodes+edges) compaction pass that
+// removes ISEs older than ENGRAM_ISE_RETENTION_MS (default 48h), protecting
+// "session-start" labels and self_review events as durable history. At
+// ~3 ISEs/min this bounds telemetry at ~8.6k nodes instead of growing forever.
 fn route_emit_ise(method: String, path: String, body: String) -> String {
     let content: String = json_get_string(body, "content")
     if str_eq(content, "") { return err_json("missing content") }
@@ -256,6 +310,64 @@ fn route_emit_ise(method: String, path: String, body: String) -> String {
         sal, imp, conf,
         "Episodic", "[\"internal-state\",\"InternalStateEvent\"]"
     )
+    let ret_raw: String = env("ENGRAM_ISE_RETENTION_MS")
+    let ret_ms: Int = if str_eq(ret_raw, "") { 172800000 } else { str_to_int(ret_raw) }
+    let pruned: Int = engram_prune_telemetry(ret_ms)
+    "{\"ok\":true,\"id\":\"" + id + "\",\"pruned\":" + int_to_str(pruned) + "}"
+}
+
+// ── Knowledge capture ─────────────────────────────────────────────────────────
+//
+// route_capture_knowledge — direct Knowledge-node capture over HTTP.
+//
+// Endpoint: POST /api/neuron/knowledge/capture   (auth required: "_auth" in body)
+// Body:     {"content": "...", "title": "...", "category": "...",
+//            "tier": "note|lesson|canonical", "tags": [...], "project": "...",
+//            "_auth": "<key>"}
+//
+// WHY (2026-07-15 self-review): the world-ingestor integrator was designed
+// against this endpoint (its MCP-unavailable fallback), but the route never
+// existed — every direct push 404'd, and because the auth gate ran before
+// routing, the failure surfaced as {"error":"unauthorized"} and was
+// misdiagnosed for two weeks while world knowledge silently dropped.
+// POST /api/nodes was no substitute: it discards label/tags/tier, which
+// makes captured knowledge invisible to tag-scoped search and curiosity.
+//
+// The incoming knowledge tier (note/lesson/canonical) is preserved as a
+// "tier:<x>" tag rather than mapped onto Engram's cognitive tiers — Knowledge
+// nodes land in Semantic (stable reference), and the epistemic tier stays
+// queryable without inventing a lossy mapping.
+fn route_capture_knowledge(method: String, path: String, body: String) -> String {
+    let content: String = json_get_string(body, "content")
+    if str_eq(content, "") { return err_json("missing content") }
+    let title: String = json_get_string(body, "title")
+    let label: String = if str_eq(title, "") { str_slice(content, 0, 60) } else { title }
+    let category_raw: String = json_get_string(body, "category")
+    let category: String = if str_eq(category_raw, "") { "other" } else { category_raw }
+    let ktier_raw: String = json_get_string(body, "tier")
+    let ktier: String = if str_eq(ktier_raw, "") { "note" } else { ktier_raw }
+    let project: String = json_get_string(body, "project")
+    let tags_raw: String = json_get_raw(body, "tags")
+    let tags_base: String = if str_eq(tags_raw, "") { "[]" } else { tags_raw }
+    // Merge category/tier/project markers into the tag array. Search matches
+    // against the tags string, so these make captures findable by facet.
+    let base_len: Int = str_len(tags_base)
+    let head: String = str_slice(tags_base, 0, base_len - 1)
+    let sep: String = if str_eq(head, "[") { "" } else { "," }
+    let safe_cat: String = str_replace(category, "\"", "'")
+    let safe_tier: String = str_replace(ktier, "\"", "'")
+    let safe_proj: String = str_replace(project, "\"", "'")
+    let proj_tag: String = if str_eq(safe_proj, "") { "" } else { ",\"project:" + safe_proj + "\"" }
+    let tags: String = head + sep + "\"category:" + safe_cat + "\",\"tier:" + safe_tier + "\"" + proj_tag + "]"
+    let sal: Float = 0.5
+    let imp: Float = 0.5
+    let conf: Float = 0.9
+    let id: String = engram_node_full(
+        content, "Knowledge", label,
+        sal, imp, conf,
+        "Semantic", tags
+    )
+    let saved: Int = persist_canonical()
     "{\"ok\":true,\"id\":\"" + id + "\"}"
 }
 
@@ -295,6 +407,12 @@ fn handle_request(method: String, path: String, body: String) -> String {
         return err_json("unauthorized")
     }
 
+    // Knowledge capture (auth enforced above; the world-ingestor integrator
+    // and any headless session without MCP push knowledge through this)
+    if str_eq(method, "POST") && str_eq(clean, "/api/neuron/knowledge/capture") {
+        return route_capture_knowledge(method, path, body)
+    }
+
     // Stats
     if str_eq(method, "GET") && (str_eq(clean, "/api/stats") || str_eq(clean, "/stats")) {
         return route_stats(method, path, body)
@@ -351,6 +469,9 @@ fn handle_request(method: String, path: String, body: String) -> String {
     if str_eq(method, "POST") && (str_eq(clean, "/api/load") || str_eq(clean, "/load")) {
         return route_load(method, path, body)
     }
+    if str_eq(method, "POST") && (str_eq(clean, "/api/load-merge") || str_eq(clean, "/load-merge")) {
+        return route_load_merge(method, path, body)
+    }
 
     // Sync — soul daemon periodic pull of non-ISE knowledge into in-process graph
     if str_eq(method, "GET") && str_eq(clean, "/api/sync") {
@@ -362,16 +483,31 @@ fn handle_request(method: String, path: String, body: String) -> String {
 
 // ── Entry ─────────────────────────────────────────────────────────────────────
 
-let bind_str: String = env("ENGRAM_BIND")
-if str_eq(bind_str, "") { let bind_str = ":8742" }
+let bind_raw: String = env("ENGRAM_BIND")
+let bind_str: String = if str_eq(bind_raw, "") { ":8742" } else { bind_raw }
 let port: Int = parse_port(bind_str)
 
 // On startup, try to load any existing snapshot (best effort).
-let data_dir: String = env("ENGRAM_DATA_DIR")
-if str_eq(data_dir, "") { let data_dir = "/tmp/engram" }
+let data_dir_raw: String = env("ENGRAM_DATA_DIR")
+let data_dir: String = if str_eq(data_dir_raw, "") { "/tmp/engram" } else { data_dir_raw }
 let snapshot_path: String = data_dir + "/snapshot.json"
 engram_load(snapshot_path)
 
+// 2026-07-21 self-review boot guard: if the snapshot file has content but the
+// load produced 0 nodes, something is wrong (corrupt file / parse failure).
+// Preserve the evidence and warn loudly — and since read routes no longer write
+// the canonical path, a bad boot can no longer clobber the good snapshot.
+let boot_snap: String = fs_read(snapshot_path)
+if !str_eq(boot_snap, "") {
+    if engram_node_count() == 0 {
+        println("[engram] WARNING: snapshot.json is non-empty but load produced 0 nodes — preserving copy at snapshot.failed-load.json")
+        fs_write(data_dir + "/snapshot.failed-load.json", boot_snap)
+    } else {
+        // Good load: keep a boot-time backup of the snapshot as loaded.
+        fs_write(data_dir + "/snapshot.boot-backup.json", boot_snap)
+    }
+}
+
 println("[engram] runtime-native graph engine")
 println("[engram] data_dir=" + data_dir)
 println("[engram] node_count=" + int_to_str(engram_node_count()))

epm/src/install.el

@@ -17,16 +17,6 @@
 //   4. Append dep to order after all its transitive deps
 //   5. Deduplicate: skip already-ordered vessels
 
-// ── Cross-module forward declarations ─────────────────────────────────────────
-// Defined in sibling epm modules; resolved at link time. The `extern fn` decls
-// give elc the C prototypes so generated install.c compiles cleanly under strict
-// compilers (gcc>=14 / clang) that reject implicit function declarations.
-extern fn manifest_name(src: String) -> String                        // manifest.el
-extern fn manifest_deps(src: String) -> String                        // manifest.el
-extern fn registry_token() -> String                                  // registry.el
-extern fn registry_find(name: String, version: String) -> String     // registry.el
-extern fn registry_latest_version(name: String) -> String            // registry.el
-
 // ── Install paths ─────────────────────────────────────────────────────────────
 
 // packages_dir returns the root directory for installed vessels.

epm/src/registry.el

@@ -14,15 +14,6 @@
 //   EPM_REGISTRY_ORG — org name that hosts vessel repos (default: neuron-technologies)
 //   EPM_TOKEN      — Gitea personal access token (required for publish)
 
-// ── Cross-module forward declarations ─────────────────────────────────────────
-// These symbols are defined in sibling epm modules or the El runtime and are
-// resolved at link time. The `extern fn` decls give elc the C prototype so the
-// generated registry.c compiles cleanly under strict compilers (gcc>=14 / clang)
-// that reject implicit function declarations. Signature arity must match the
-// definition; return/param types are informational (all lower to el_val_t).
-extern fn config(key: String) -> String          // El runtime builtin
-extern fn read_installed() -> String              // install.el
-
 // ── Config helpers ────────────────────────────────────────────────────────────
 
 // registry_api_url returns the Gitea API base URL with no trailing slash.

epm/src/update.el

@@ -6,15 +6,6 @@
 // Depends on: registry.el (registry_latest_version, registry_find),
 //             install.el  (read_installed, install_vessel, installed_version)
 
-// ── Cross-module forward declarations ─────────────────────────────────────────
-// Defined in sibling epm modules; resolved at link time. The `extern fn` decls
-// give elc the C prototypes so generated update.c compiles cleanly under strict
-// compilers (gcc>=14 / clang) that reject implicit function declarations.
-extern fn read_installed() -> String                                  // install.el
-extern fn installed_version(name: String) -> String                   // install.el
-extern fn install_vessel(name: String, version: String) -> Bool      // install.el
-extern fn registry_latest_version(name: String) -> String            // registry.el
-
 // ── Semver helpers ────────────────────────────────────────────────────────────
 
 // semver_part extracts the Nth dot-separated component from a semver string.

lang/el-compiler/runtime/el_platform_win.h

@@ -75,7 +75,6 @@ static inline void* el_win_dlsym(void* handle, const char* name) {
 #include <direct.h>            /* _mkdir */
 #define mkdir(path, mode) _mkdir(path)     /* POSIX mkdir(path,mode) → _mkdir(path) */
 #define timegm _mkgmtime                   /* UTC tm → time_t */
-#define fsync(fd) _commit(fd)              /* no fsync() on Windows; _commit() (<io.h>) is the equiv */
 
 /* setenv/unsetenv: not in the Windows CRT; map to _putenv_s / SetEnvironmentVariable. */
 static inline int setenv(const char* name, const char* value, int overwrite) {

lang/el-compiler/runtime/el_runtime.c

@@ -82,14 +82,8 @@ static _Thread_local ElArena _tl_arena = {NULL, 0, 0};
 static _Thread_local int     _tl_arena_active = 0;
 
 /* Binary-safe fs_read length — set by fs_read, consumed by http_send_response.
- * Allows serving PNGs and other binary files without strlen truncation.
- * PAIRED with the buffer pointer it describes: the length may only be applied
- * to the exact buffer fs_read returned. Without the pairing, any handler that
- * fs_read a file and then WRAPPED it into a larger response had that response
- * truncated to the file's length (Content-Length lied AND the send stopped
- * short) — the safety-contact onboarding trap, 2026-07-17. */
-static _Thread_local size_t      _tl_fs_read_len  = 0;
-static _Thread_local const char* _tl_fs_read_buf  = NULL;
+ * Allows serving PNGs and other binary files without strlen truncation. */
+static _Thread_local size_t  _tl_fs_read_len  = 0;
 
 static void el_arena_track(char* p) {
     if (!_tl_arena_active || !p) return;
@@ -107,8 +101,6 @@ static void el_arena_track(char* p) {
 void el_request_start(void) {
     _tl_arena.count  = 0;
     _tl_arena_active = 1;
-    _tl_fs_read_len  = 0;     /* never let a previous request's file length */
-    _tl_fs_read_buf  = NULL;  /* leak into this response's byte accounting  */
 }
 
 /* Called by http_worker after the El handler returns and the response is sent.
@@ -1492,14 +1484,11 @@ static void http_send_response(int fd, const char* body) {
     }
 
     const char* eff_body = is_envelope ? env_body : body;
-    /* Use the real byte count from fs_read ONLY when this body IS the exact
-     * buffer fs_read returned (binary files with embedded null bytes — PNG,
-     * WOFF2, etc.). Any other body — wrapped, enveloped, or derived — must be
-     * measured with strlen, or it is truncated/over-read to the file's size. */
-    size_t blen = (_tl_fs_read_len > 0 && eff_body == _tl_fs_read_buf)
-                      ? _tl_fs_read_len : strlen(eff_body);
+    /* Use the real byte count from fs_read if available (handles binary files
+     * with embedded null bytes — PNG, WOFF2, etc.). Fall back to strlen for
+     * normal text/JSON responses where _tl_fs_read_len is 0. */
+    size_t blen = (_tl_fs_read_len > 0) ? _tl_fs_read_len : strlen(eff_body);
     _tl_fs_read_len = 0;  /* consume — one-shot per response */
-    _tl_fs_read_buf = NULL;
     int head_only = _tl_http_head_only;
 
     JsonBuf hdrs; jb_init(&hdrs);
@@ -1579,22 +1568,11 @@ static void* http_worker(void* arg) {
             const char* rs = EL_CSTR(r);
             /* Copy response out BEFORE arena teardown.
              * For binary files, _tl_fs_read_len holds the real byte count —
-             * use memcpy instead of strdup so null bytes are preserved.
-             * The stored length applies ONLY when the response IS the exact
-             * fs_read buffer; a wrapped/derived response must use strlen or
-             * it gets truncated (or over-read) to the file's length. */
-            size_t rlen;
-            if (_tl_fs_read_len > 0 && rs && rs == _tl_fs_read_buf) {
-                rlen = _tl_fs_read_len;   /* raw file bytes — binary-safe */
-            } else {
-                rlen = rs ? strlen(rs) : 0;
-                _tl_fs_read_len = 0;      /* hint doesn't describe this body */
-                _tl_fs_read_buf = NULL;
-            }
+             * use memcpy instead of strdup so null bytes are preserved. */
+            size_t rlen = _tl_fs_read_len > 0 ? _tl_fs_read_len : (rs ? strlen(rs) : 0);
             response = malloc(rlen + 1);
             if (response && rs) { memcpy(response, rs, rlen); response[rlen] = '\0'; }
             else if (response) { response[0] = '\0'; }
-            if (_tl_fs_read_len > 0) _tl_fs_read_buf = response; /* hint follows the copy */
         } else {
             response = el_strdup_persist("el-runtime: no http handler registered");
         }
@@ -1844,20 +1822,10 @@ static void* http_worker_v2(void* arg) {
             el_val_t hmap = http_build_headers_map(hdr_block ? hdr_block : "");
             el_val_t r = h(EL_STR(dispatch_method), EL_STR(path), hmap, EL_STR(body));
             const char* rs = EL_CSTR(r);
-            /* Same pairing rule as the v1 worker: the fs_read length is only
-             * trustworthy for the exact buffer fs_read returned. */
-            size_t rlen;
-            if (_tl_fs_read_len > 0 && rs && rs == _tl_fs_read_buf) {
-                rlen = _tl_fs_read_len;   /* raw file bytes — binary-safe */
-            } else {
-                rlen = rs ? strlen(rs) : 0;
-                _tl_fs_read_len = 0;      /* hint doesn't describe this body */
-                _tl_fs_read_buf = NULL;
-            }
+            size_t rlen = _tl_fs_read_len > 0 ? _tl_fs_read_len : (rs ? strlen(rs) : 0);
             response = malloc(rlen + 1);
             if (response && rs) { memcpy(response, rs, rlen); response[rlen] = '\0'; }
             else if (response) { response[0] = '\0'; }
-            if (_tl_fs_read_len > 0) _tl_fs_read_buf = response; /* hint follows the copy */
             el_release(hmap);
         } else {
             response = el_strdup_persist(
@@ -1995,9 +1963,8 @@ void http_serve_async(el_val_t port, el_val_t handler) {
     int sock = socket(AF_INET6, SOCK_STREAM, 0);
     if (sock < 0) { perror("socket"); return; }
     int yes = 1; int no = 0;
-    /* Win32/mingw setsockopt takes optval as (const char*); the cast is portable on POSIX too. */
-    setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, (const char*)&yes, sizeof(yes));
-    setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, (const char*)&no, sizeof(no));
+    setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof(yes));
+    setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, &no, sizeof(no));
     struct sockaddr_in6 addr;
     memset(&addr, 0, sizeof(addr));
     addr.sin6_family = AF_INET6;
@@ -2056,7 +2023,6 @@ el_val_t http_response(el_val_t status, el_val_t headers_json, el_val_t body) {
 el_val_t fs_read(el_val_t pathv) {
     const char* path = EL_CSTR(pathv);
     _tl_fs_read_len = 0;
-    _tl_fs_read_buf = NULL;
     if (!path) return el_wrap_str(el_strdup(""));
     FILE* f = fopen(path, "rb");
     if (!f) return el_wrap_str(el_strdup(""));
@@ -2068,7 +2034,6 @@ el_val_t fs_read(el_val_t pathv) {
     size_t got = fread(buf, 1, (size_t)sz, f);
     buf[got] = '\0';
     _tl_fs_read_len = got;   /* store real byte count for binary-safe send */
-    _tl_fs_read_buf = buf;   /* ...valid ONLY for this exact buffer */
     fclose(f);
     return el_wrap_str(buf);
 }
@@ -3611,10 +3576,8 @@ el_val_t json_get_raw(el_val_t json_str, el_val_t key) {
     const char* k    = EL_CSTR(key);
     const char* p = json_find_key(json, k);
     /* Clear fs_read binary-length hint — result is a fresh null-terminated
-     * string, not the raw file bytes, so Content-Length must use strlen.
-     * (Kept although the pointer pairing now makes this redundant.) */
+     * string, not the raw file bytes, so Content-Length must use strlen. */
     _tl_fs_read_len = 0;
-    _tl_fs_read_buf = NULL;
     if (!p) return el_wrap_str(el_strdup(""));
     const char* end = json_skip_value(p);
     size_t n = (size_t)(end - p);
@@ -6932,243 +6895,6 @@ static int engram_rank_cmp(const void* a, const void* b) {
     return 0;
 }
 
-/* ══════════════════════════════════════════════════════════════════════════
- * SEMANTIC SEARCH LAYER — nomic-embed-text via Ollama /api/embeddings
- * ──────────────────────────────────────────────────────────────────────────
- * Augments the lexical (istr_contains) matcher with dense-vector retrieval.
- * Node content and the query are embedded through a local Ollama server;
- * nodes are ranked by cosine similarity and UNIONED with lexical hits. This
- * lets a paraphrase query surface a node whose words never appear in it.
- *
- * DEGRADABLE BY DESIGN. The whole layer is gated on HAVE_CURL plus a one-shot
- * runtime probe of the embedding endpoint. If curl is not compiled in, or
- * Ollama is unreachable, or ENGRAM_SEMANTIC=0, every entry point returns
- * "no semantic signal" and callers fall back to pure lexical behaviour —
- * byte-for-byte the pre-existing search.
- *
- * CACHE. Node embeddings are computed lazily on first use and cached in
- * process memory keyed by node id, with an FNV-1a content hash for
- * invalidation (edited content re-embeds). The query is embedded once per
- * search call. This is what "avoid re-embedding the whole graph every query"
- * buys us: a warm cache serves cosine from RAM. (A cold process still pays
- * O(N) embed calls the first time each node is scanned — persisting the cache
- * to a snapshot sidecar is the documented next step, not done here.)
- *
- * nomic task prefixes ("search_query:" / "search_document:") are applied
- * because nomic-embed-text is trained with them; they materially improve
- * retrieval separation (empirically: paraphrase 0.72 vs distractors <0.48).
- *
- * ENV:
- *   ENGRAM_SEMANTIC       "0" disables; unset/other = auto-probe
- *   ENGRAM_EMBED_URL      default http://localhost:11434/api/embeddings
- *   ENGRAM_EMBED_MODEL    default nomic-embed-text
- *   ENGRAM_SEMANTIC_MIN   cosine threshold for a pure-semantic match (def 0.6)
- * ════════════════════════════════════════════════════════════════════════ */
-
-static double engram_semantic_min(void) {
-    static double v = -1.0;
-    if (v >= 0.0) return v;
-    const char* s = getenv("ENGRAM_SEMANTIC_MIN");
-    double d = 0.6;
-    if (s && *s) { char* e = NULL; double t = strtod(s, &e);
-                   if (e != s && t >= 0.0 && t <= 1.0) d = t; }
-    v = d; return v;
-}
-
-#ifdef HAVE_CURL
-
-typedef struct { char* id; uint64_t hash; float* vec; int dim; } EngramEmbEntry;
-static EngramEmbEntry* g_emb_items = NULL;
-static int64_t g_emb_count = 0, g_emb_cap = 0;
-static int g_emb_state = 0;   /* 0=unprobed, 1=available, -1=disabled */
-
-static uint64_t engram_fnv1a(const char* s) {
-    uint64_t h = 1469598103934665603ULL;
-    if (s) for (const unsigned char* p = (const unsigned char*)s; *p; p++) {
-        h ^= *p; h *= 1099511628211ULL;
-    }
-    return h;
-}
-
-/* Parse "embedding":[f,f,...] from an Ollama response. malloc'd vec, or NULL. */
-static float* engram_parse_embedding(const char* json, int* out_dim) {
-    if (!json) return NULL;
-    const char* p = strstr(json, "\"embedding\"");
-    if (!p) return NULL;
-    p = strchr(p, '[');
-    if (!p) return NULL;
-    p++;
-    int cap = 1024, n = 0;
-    float* v = malloc((size_t)cap * sizeof(float));
-    if (!v) return NULL;
-    while (*p && *p != ']') {
-        while (*p == ' ' || *p == '\t' || *p == '\n' || *p == '\r' || *p == ',') p++;
-        if (*p == ']' || !*p) break;
-        char* e = NULL;
-        double d = strtod(p, &e);
-        if (e == p) break;
-        if (n >= cap) { cap *= 2; float* nv = realloc(v, (size_t)cap * sizeof(float));
-                        if (!nv) { free(v); return NULL; } v = nv; }
-        v[n++] = (float)d;
-        p = e;
-    }
-    if (n == 0) { free(v); return NULL; }
-    *out_dim = n;
-    return v;
-}
-
-/* JSON-escape src into a malloc'd buffer (no surrounding quotes). */
-static char* engram_json_escape(const char* src) {
-    if (!src) src = "";
-    size_t n = strlen(src);
-    char* out = malloc(n * 2 + 1);
-    if (!out) return NULL;
-    size_t j = 0;
-    for (size_t i = 0; i < n; i++) {
-        unsigned char c = (unsigned char)src[i];
-        if      (c == '"')  { out[j++] = '\\'; out[j++] = '"'; }
-        else if (c == '\\') { out[j++] = '\\'; out[j++] = '\\'; }
-        else if (c == '\n') { out[j++] = '\\'; out[j++] = 'n'; }
-        else if (c == '\r') { out[j++] = '\\'; out[j++] = 'r'; }
-        else if (c == '\t') { out[j++] = '\\'; out[j++] = 't'; }
-        else if (c < 0x20)  { /* drop other control bytes */ }
-        else                { out[j++] = (char)c; }
-    }
-    out[j] = '\0';
-    return out;
-}
-
-/* Embed `prefix+text` via Ollama. Returns malloc'd vec (caller frees), or NULL. */
-static float* engram_embed_raw(const char* prefix, const char* text, int* out_dim) {
-    if (!text) return NULL;
-    const char* url = getenv("ENGRAM_EMBED_URL");
-    if (!url || !*url) url = "http://localhost:11434/api/embeddings";
-    const char* model = getenv("ENGRAM_EMBED_MODEL");
-    if (!model || !*model) model = "nomic-embed-text";
-    /* Bound content length to keep latency/memory sane on huge nodes. */
-    char* trunc = NULL;
-    size_t maxlen = 8192;
-    if (strlen(text) > maxlen) {
-        trunc = malloc(maxlen + 1);
-        if (trunc) { memcpy(trunc, text, maxlen); trunc[maxlen] = '\0'; text = trunc; }
-    }
-    char* esc_prefix = engram_json_escape(prefix ? prefix : "");
-    char* esc = engram_json_escape(text);
-    free(trunc);
-    if (!esc || !esc_prefix) { free(esc); free(esc_prefix); return NULL; }
-    size_t blen = strlen(esc) + strlen(esc_prefix) + strlen(model) + 64;
-    char* body = malloc(blen);
-    if (!body) { free(esc); free(esc_prefix); return NULL; }
-    snprintf(body, blen, "{\"model\":\"%s\",\"prompt\":\"%s%s\"}", model, esc_prefix, esc);
-    free(esc); free(esc_prefix);
-
-    CURL* c = curl_easy_init();
-    if (!c) { free(body); return NULL; }
-    HttpBuf rb; httpbuf_init(&rb);
-    struct curl_slist* h = curl_slist_append(NULL, "Content-Type: application/json");
-    char errbuf[CURL_ERROR_SIZE]; errbuf[0] = '\0';
-    curl_easy_setopt(c, CURLOPT_URL, url);
-    curl_easy_setopt(c, CURLOPT_WRITEFUNCTION, http_write_cb);
-    curl_easy_setopt(c, CURLOPT_WRITEDATA, &rb);
-    curl_easy_setopt(c, CURLOPT_POST, 1L);
-    curl_easy_setopt(c, CURLOPT_POSTFIELDS, body);
-    curl_easy_setopt(c, CURLOPT_POSTFIELDSIZE, (long)strlen(body));
-    curl_easy_setopt(c, CURLOPT_HTTPHEADER, h);
-    curl_easy_setopt(c, CURLOPT_TIMEOUT_MS, el_http_timeout_ms());
-    curl_easy_setopt(c, CURLOPT_NOSIGNAL, 1L);
-    curl_easy_setopt(c, CURLOPT_ERRORBUFFER, errbuf);
-    CURLcode rc = curl_easy_perform(c);
-    curl_slist_free_all(h);
-    curl_easy_cleanup(c);
-    free(body);
-    if (rc != CURLE_OK) { free(rb.data); return NULL; }
-    float* v = engram_parse_embedding(rb.data, out_dim);
-    free(rb.data);
-    return v;
-}
-
-/* One-shot probe: is semantic search available? Caches the verdict. */
-static int engram_semantic_enabled(void) {
-    if (g_emb_state != 0) return g_emb_state == 1;
-    const char* s = getenv("ENGRAM_SEMANTIC");
-    if (s && strcmp(s, "0") == 0) { g_emb_state = -1; return 0; }
-    int dim = 0;
-    float* v = engram_embed_raw("search_query: ", "probe", &dim);
-    if (v && dim > 0) { free(v); g_emb_state = 1; return 1; }
-    free(v);
-    g_emb_state = -1; return 0;
-}
-
-/* Embed the query. Returns malloc'd vec (caller frees), or NULL if semantic off. */
-static float* engram_embed_query(const char* q, int* dim) {
-    if (!engram_semantic_enabled()) return NULL;
-    if (!q || !*q) return NULL;
-    return engram_embed_raw("search_query: ", q, dim);
-}
-
-/* Cached node embedding. Returns a pointer OWNED BY THE CACHE — do not free. */
-static const float* engram_node_vec(EngramNode* n, int* out_dim) {
-    if (!n || !n->id) return NULL;
-    uint64_t h = engram_fnv1a(n->content);
-    for (int64_t i = 0; i < g_emb_count; i++) {
-        if (g_emb_items[i].id && strcmp(g_emb_items[i].id, n->id) == 0) {
-            if (g_emb_items[i].hash == h && g_emb_items[i].vec) {
-                *out_dim = g_emb_items[i].dim; return g_emb_items[i].vec;
-            }
-            /* content changed → re-embed in place */
-            int dim = 0;
-            float* v = engram_embed_raw("search_document: ", n->content ? n->content : "", &dim);
-            if (!v) return NULL;
-            free(g_emb_items[i].vec);
-            g_emb_items[i].vec = v; g_emb_items[i].dim = dim; g_emb_items[i].hash = h;
-            *out_dim = dim; return v;
-        }
-    }
-    int dim = 0;
-    float* v = engram_embed_raw("search_document: ", n->content ? n->content : "", &dim);
-    if (!v) return NULL;
-    if (g_emb_count >= g_emb_cap) {
-        int64_t nc = g_emb_cap ? g_emb_cap * 2 : 256;
-        EngramEmbEntry* ni = realloc(g_emb_items, (size_t)nc * sizeof(EngramEmbEntry));
-        if (!ni) { free(v); return NULL; }
-        g_emb_items = ni; g_emb_cap = nc;
-    }
-    g_emb_items[g_emb_count].id   = strdup(n->id);
-    g_emb_items[g_emb_count].hash = h;
-    g_emb_items[g_emb_count].vec  = v;
-    g_emb_items[g_emb_count].dim  = dim;
-    g_emb_count++;
-    *out_dim = dim; return v;
-}
-
-static double engram_cosine(const float* a, const float* b, int dim) {
-    double dot = 0, na = 0, nb = 0;
-    for (int i = 0; i < dim; i++) { dot += (double)a[i] * b[i];
-                                    na  += (double)a[i] * a[i];
-                                    nb  += (double)b[i] * b[i]; }
-    if (na <= 0 || nb <= 0) return 0.0;
-    return dot / (sqrt(na) * sqrt(nb));
-}
-
-/* Cosine of node n against the query vector; 0 if unavailable / dim mismatch. */
-static double engram_node_cosine(EngramNode* n, const float* qvec, int qdim) {
-    if (!qvec || qdim <= 0) return 0.0;
-    int ndim = 0;
-    const float* nv = engram_node_vec(n, &ndim);
-    if (!nv || ndim != qdim) return 0.0;
-    return engram_cosine(qvec, nv, qdim);
-}
-
-#else  /* !HAVE_CURL — semantic layer compiled out; callers stay pure-lexical.
-        * Only the two boundary functions the always-compiled search/activate
-        * code calls are stubbed; the query embed always yields NULL so every
-        * cosine is 0 and every caller collapses to lexical-only. */
-static float* engram_embed_query(const char* q, int* dim) { (void)q; (void)dim; return NULL; }
-static double engram_node_cosine(EngramNode* n, const float* qvec, int qdim) {
-    (void)n; (void)qvec; (void)qdim; return 0.0;
-}
-#endif /* HAVE_CURL */
-
 el_val_t engram_search(el_val_t query, el_val_t limit) {
     EngramStore* g = engram_get();
     const char* q = EL_CSTR(query);
@@ -7179,15 +6905,8 @@ el_val_t engram_search(el_val_t query, el_val_t limit) {
     char toks[ENGRAM_MAX_QTOKENS][ENGRAM_QTOK_LEN];
     int ntok = engram_tokenize_query(q, toks, ENGRAM_MAX_QTOKENS);
     if (ntok == 0) return lst;
-    /* Semantic augmentation: embed the query once; a node is a hit if it covers
-     * >=1 query token (tokenized-lexical, #66) OR its cosine clears the
-     * threshold (#67). qvec is NULL (cosine 0) when semantic is unavailable →
-     * pure tokenized-lexical, byte-identical to the lexical-only behaviour. */
-    int qdim = 0;
-    float* qvec = engram_embed_query(q, &qdim);
-    double sem_min = engram_semantic_min();
     EngramRankEntry* hits = malloc((size_t)g->node_count * sizeof(EngramRankEntry));
-    if (!hits) { free(qvec); return lst; }
+    if (!hits) return lst;
     int64_t nhits = 0;
     for (int64_t i = 0; i < g->node_count; i++) {
         EngramNode* n = &g->nodes[i];
@@ -7197,24 +6916,20 @@ el_val_t engram_search(el_val_t query, el_val_t limit) {
          * + engram_compile_layered_json — that's the legitimate path. */
         if (engram_layer_is_transparent(n->layer_id)) continue;
         int sc = engram_node_match_score(n, toks, ntok);
-        double sem = qvec ? engram_node_cosine(n, qvec, qdim) : 0.0;
-        if (sc > 0 || sem >= sem_min) {
+        if (sc > 0) {
             hits[nhits].idx      = i;
             hits[nhits].score    = sc;
             hits[nhits].salience = n->salience;
             nhits++;
         }
     }
-    /* Rank by distinct tokens matched (desc) then salience (desc), then cap.
-     * Pure-semantic hits (token score 0) sort after every lexical hit — a
-     * lexical ∪ semantic union with lexical precedence. */
+    /* Rank by distinct tokens matched (desc) then salience (desc), then cap. */
     qsort(hits, (size_t)nhits, sizeof(EngramRankEntry), engram_rank_cmp);
     int64_t end = nhits < lim ? nhits : lim;
     for (int64_t k = 0; k < end; k++) {
         lst = el_list_append(lst, engram_node_to_map(&g->nodes[hits[k].idx]));
     }
     free(hits);
-    free(qvec);
     return lst;
 }
 
@@ -7564,31 +7279,21 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
     if (!seeds) {
         free(best_bg); free(best_hops); free(reached); return out;
     }
-    /* Tokenized + semantic seeding: a node seeds if it covers >=1 query token
-     * (tokenized-lexical, #66) OR its cosine clears the threshold (#67). A
-     * lexical seed's activation is scaled by token coverage (fraction of
-     * distinct query tokens covered) so a node matching all words seeds more
-     * strongly than one matching a single word; single-word queries → coverage
-     * 1.0. A pure-semantic seed (no token match) is instead down-weighted by
-     * its cosine so paraphrase matches spread without overpowering exact seeds.
-     * q_vec is NULL (cosine 0) when semantic is unavailable → the seed set is
-     * exactly the tokenized-lexical one. q_vec is freed right after this loop
-     * so the many downstream early-returns need no cleanup change. */
+    /* Tokenize once: a node seeds if it matches ANY query token, and its seed
+     * activation is scaled by token coverage (fraction of distinct query
+     * tokens it contains) so a node matching all words seeds more strongly
+     * than one matching a single word. Single-word queries → coverage 1.0,
+     * identical to the prior whole-query behavior. */
     char toks[ENGRAM_MAX_QTOKENS][ENGRAM_QTOK_LEN];
     int ntok = engram_tokenize_query(q, toks, ENGRAM_MAX_QTOKENS);
-    int q_dim = 0;
-    float* q_vec = engram_embed_query(q, &q_dim);
-    double q_sem_min = engram_semantic_min();
     for (int64_t i = 0; i < g->node_count; i++) {
         EngramNode* n = &g->nodes[i];
         int sc = engram_node_match_score(n, toks, ntok);
-        double sem = q_vec ? engram_node_cosine(n, q_vec, q_dim) : 0.0;
-        if (sc > 0 || sem >= q_sem_min) {
+        if (sc > 0) {
             double tdecay = engram_temporal_decay(n, now_ms);
             double dampen = engram_activation_dampen(n);
-            double act    = n->salience * tdecay * dampen;
-            if (sc > 0) act *= (ntok > 0 ? (double)sc / (double)ntok : 1.0);
-            else        act *= sem;  /* pure-semantic seed: down-weight by cosine */
+            double cover  = ntok > 0 ? (double)sc / (double)ntok : 1.0;
+            double act    = n->salience * tdecay * dampen * cover;
             seeds[seed_count].idx        = i;
             seeds[seed_count].act        = act;
             seeds[seed_count].created_at = n->created_at;
@@ -7598,7 +7303,6 @@ el_val_t engram_activate(el_val_t query, el_val_t depth) {
             reached[i]   = 1;
         }
     }
-    free(q_vec);
     /* Compute mean seed created_at for temporal proximity bonus. */
     int64_t seed_epoch = 0;
     if (seed_count > 0) {
@@ -8150,35 +7854,6 @@ el_val_t engram_get_node_json(el_val_t id) {
     return el_wrap_str(jb_finish(&b));
 }
 
-/* engram_get_node_by_label — find the first node whose label field exactly
- * matches the given string. Returns the node as a JSON object string, or "{}"
- * if no match is found.
- *
- * Used by chat.el to retrieve well-known nodes (e.g. "conv:history",
- * "session:summary") by their stable label rather than by ID, which is immune
- * to vector index drift across restarts.
- *
- * Exact match (strcmp, not istr_contains) because labels like "conv:history"
- * must not collide with nodes whose content happens to contain that substring.
- *
- * Backported verbatim (idiom-adapted to jb_finish) from release runtime
- * v1.0.0-20260501 to unblock the soul regen link: chat.el references this
- * native but the current runtime lacked its definition. */
-el_val_t engram_get_node_by_label(el_val_t label) {
-    const char* lbl = EL_CSTR(label);
-    if (!lbl || !*lbl) return el_wrap_str(el_strdup("{}"));
-    EngramStore* g = engram_get();
-    for (int64_t i = 0; i < g->node_count; i++) {
-        EngramNode* n = &g->nodes[i];
-        if (n->label && strcmp(n->label, lbl) == 0) {
-            JsonBuf b; jb_init(&b);
-            engram_emit_node_json(&b, n);
-            return el_wrap_str(jb_finish(&b));
-        }
-    }
-    return el_wrap_str(el_strdup("{}"));
-}
-
 el_val_t engram_search_json(el_val_t query, el_val_t limit) {
     EngramStore* g = engram_get();
     const char* q = EL_CSTR(query);
@@ -8186,50 +7861,39 @@ el_val_t engram_search_json(el_val_t query, el_val_t limit) {
     if (lim <= 0) lim = 100;
     JsonBuf b; jb_init(&b);
     jb_putc(&b, '[');
-    if (q && *q && g->node_count > 0) {
-        /* Collect candidates from the UNION of tokenized-lexical and semantic
-         * matches, score each, rank by score, emit the top `lim`. A node is a
-         * candidate if it covers >=1 query token (tokenized-lexical, #66) OR its
-         * query cosine clears the threshold (#67). Lexical score is the distinct
-         * token count (>=1), so any lexical hit outranks a pure-semantic hit
-         * (cosine < 1); pure-semantic hits are scored by cosine alone. When
-         * semantic is unavailable qvec is NULL, sem is 0, only tokenized-lexical
-         * hits are collected, and the stable insertion sort preserves order. */
+    int first = 1;
+    if (q && *q) {
         char toks[ENGRAM_MAX_QTOKENS][ENGRAM_QTOK_LEN];
         int ntok = engram_tokenize_query(q, toks, ENGRAM_MAX_QTOKENS);
-        int qdim = 0;
-        float* qvec = engram_embed_query(q, &qdim);
-        double sem_min = engram_semantic_min();
-        typedef struct { int64_t idx; double score; } Cand;
-        Cand* cand = malloc((size_t)g->node_count * sizeof(Cand));
-        if (cand) {
-            int64_t nc = 0;
-            for (int64_t i = 0; i < g->node_count; i++) {
-                EngramNode* n = &g->nodes[i];
-                if (engram_layer_is_transparent(n->layer_id)) continue;
-                int sc = engram_node_match_score(n, toks, ntok);
-                double sem = qvec ? engram_node_cosine(n, qvec, qdim) : 0.0;
-                if (sc > 0 || sem >= sem_min) {
-                    cand[nc].idx   = i;
-                    cand[nc].score = (double)sc + sem;
-                    nc++;
+        if (ntok > 0) {
+            EngramRankEntry* hits =
+                malloc((size_t)g->node_count * sizeof(EngramRankEntry));
+            if (hits) {
+                int64_t nhits = 0;
+                for (int64_t i = 0; i < g->node_count; i++) {
+                    EngramNode* n = &g->nodes[i];
+                    /* Filter transparent layers — same as engram_search. */
+                    if (engram_layer_is_transparent(n->layer_id)) continue;
+                    int sc = engram_node_match_score(n, toks, ntok);
+                    if (sc > 0) {
+                        hits[nhits].idx      = i;
+                        hits[nhits].score    = sc;
+                        hits[nhits].salience = n->salience;
+                        nhits++;
+                    }
                 }
+                /* Rank by distinct tokens matched (desc) then salience (desc). */
+                qsort(hits, (size_t)nhits, sizeof(EngramRankEntry),
+                      engram_rank_cmp);
+                int64_t end = nhits < lim ? nhits : lim;
+                for (int64_t k = 0; k < end; k++) {
+                    if (!first) jb_putc(&b, ',');
+                    engram_emit_node_json(&b, &g->nodes[hits[k].idx]);
+                    first = 0;
+                }
+                free(hits);
             }
-            /* Insertion sort by score desc; stable for equal scores. */
-            for (int64_t i = 1; i < nc; i++) {
-                Cand k = cand[i]; int64_t j = i - 1;
-                while (j >= 0 && cand[j].score < k.score) { cand[j + 1] = cand[j]; j--; }
-                cand[j + 1] = k;
-            }
-            int first = 1;
-            for (int64_t i = 0; i < nc && i < lim; i++) {
-                if (!first) jb_putc(&b, ',');
-                engram_emit_node_json(&b, &g->nodes[cand[i].idx]);
-                first = 0;
-            }
-            free(cand);
         }
-        free(qvec);
     }
     jb_putc(&b, ']');
     return el_wrap_str(jb_finish(&b));

lang/el-compiler/runtime/el_runtime.h

@@ -632,7 +632,6 @@ el_val_t  engram_load(el_val_t path);
  * can pass results straight through without round-tripping ElList/ElMap
  * through json_stringify. */
 el_val_t  engram_get_node_json(el_val_t id);
-el_val_t  engram_get_node_by_label(el_val_t label);
 el_val_t  engram_search_json(el_val_t query, el_val_t limit);
 el_val_t  engram_scan_nodes_json(el_val_t limit, el_val_t offset);
 el_val_t  engram_scan_nodes_by_type_json(el_val_t node_type, el_val_t limit, el_val_t offset);

lang/el-compiler/runtime/el_seed.c

@@ -1072,7 +1072,6 @@ el_val_t __engram_save(el_val_t path)                       { return engram_save
 el_val_t __engram_load(el_val_t path)                       { return engram_load(path); }
 
 el_val_t __engram_get_node_json(el_val_t id)               { return engram_get_node_json(id); }
-el_val_t __engram_get_node_by_label(el_val_t label)        { return engram_get_node_by_label(label); }
 
 el_val_t __engram_search_json(el_val_t query, el_val_t limit) {
     return engram_search_json(query, limit);

lang/el-compiler/runtime/el_seed.h

@@ -226,7 +226,6 @@ el_val_t  __engram_activate(el_val_t query, el_val_t depth);
 el_val_t  __engram_save(el_val_t path);
 el_val_t  __engram_load(el_val_t path);
 el_val_t  __engram_get_node_json(el_val_t id);
-el_val_t  __engram_get_node_by_label(el_val_t label);
 el_val_t  __engram_search_json(el_val_t query, el_val_t limit);
 el_val_t  __engram_scan_nodes_json(el_val_t limit, el_val_t offset);
 el_val_t  __engram_scan_nodes_by_type_json(el_val_t node_type, el_val_t limit, el_val_t offset);

lang/el-compiler/src/codegen.el

@@ -2670,7 +2670,6 @@ fn builtin_arity(name: String) -> Int {
     if str_eq(name, "engram_save") { return 1 }
     if str_eq(name, "engram_load") { return 1 }
     if str_eq(name, "engram_get_node_json") { return 1 }
-    if str_eq(name, "engram_get_node_by_label") { return 1 }
     if str_eq(name, "engram_search_json") { return 2 }
     if str_eq(name, "engram_scan_nodes_json") { return 2 }
     if str_eq(name, "engram_neighbors_json") { return 3 }