Merge pull request 'release: promote stage -> main (ci publish hardening for sdk-release)' (#77) from stage into main

Mirror the PR #72 fix (applied to ci-dev.yaml) onto ci-stage.yaml and
sdk-release.yaml. The stage and prod release jobs reported FAILURE even
when the el-runtime-c/-h publish SUCCEEDED, because the ancillary ci-base
Docker rebuild (a CI-cache optimization on the fragile host-mode GCE
runner) reddened the whole job.

- Rebuild ci-base step: continue-on-error: true — never blocks/reddens
  the job; the SDK publish is the deliverable.
- Publish step: set -euo pipefail + empty-key guard + active-account echo
  so a real publish failure still fails loud and is diagnosable.

.gitea/workflows/ci-dev.yaml

@@ -214,9 +214,18 @@ 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}"
 
@@ -268,6 +277,12 @@ 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

@@ -212,12 +212,21 @@ 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}"
 
@@ -253,6 +262,12 @@ 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

@@ -288,12 +288,21 @@ 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}"
 
@@ -345,6 +354,12 @@ 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 \
+          cc -std=c11 -O2 -DHAVE_CURL \
             -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 \
+          cc -std=c11 -O2 -DHAVE_CURL \
             -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 \
+          cc -std=c11 -O2 -DHAVE_CURL \
             -I /usr/local/lib/el \
             -o dist/engram \
             dist/engram.c \

epm/src/install.el

@@ -17,6 +17,16 @@
 //   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

@@ -14,6 +14,15 @@
 //   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

@@ -6,6 +6,15 @@
 // 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

@@ -75,6 +75,7 @@ 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

@@ -710,6 +710,7 @@ static void jb_init(JsonBuf* b);
 static void jb_putc(JsonBuf* b, char c);
 static void jb_puts(JsonBuf* b, const char* s);
 static void jb_emit_escaped(JsonBuf* b, const char* s);
+static char* jb_finish(JsonBuf* b);
 static int  looks_like_string(el_val_t v);
 static const char* json_find_key(const char* s, const char* key);
 static const char* json_skip_value(const char* p);
@@ -1962,8 +1963,9 @@ 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;
-    setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof(yes));
+    /* Win32/mingw setsockopt takes optval as (const char*); the cast is portable on POSIX too. */
-    setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, &no, sizeof(no));
+    setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, (const char*)&yes, sizeof(yes));
+    setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, (const char*)&no, sizeof(no));
     struct sockaddr_in6 addr;
     memset(&addr, 0, sizeof(addr));
     addr.sin6_family = AF_INET6;
@@ -2014,7 +2016,7 @@ el_val_t http_response(el_val_t status, el_val_t headers_json, el_val_t body) {
     jb_puts(&out, ",\"body\":");
     jb_emit_escaped(&out, b);
     jb_putc(&out, '}');
-    return el_wrap_str(out.buf);
+    return el_wrap_str(jb_finish(&out));
 }
 
 /* ── Filesystem ──────────────────────────────────────────────────────────── */
@@ -2121,7 +2123,7 @@ el_val_t exec_capture(el_val_t cmdv) {
     char buf[4096];
     while (fgets(buf, sizeof(buf), f)) jb_puts(&b, buf);
     pclose(f);
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 // exec — run a shell command via /bin/sh, capture stdout, return as String.
@@ -2145,7 +2147,7 @@ el_val_t exec(el_val_t cmdv) {
         jb_puts(&b, buf);
     }
     pclose(f);
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 // exec_bg — run a shell command in background, return PID as String.
@@ -3292,6 +3294,23 @@ static void jb_puts(JsonBuf* b, const char* s) {
     b->buf[b->len] = '\0';
 }
 
+/* jb_finish — call exactly once, at the point a JsonBuf's buffer is handed
+ * off as an el_val_t return value (via el_wrap_str). JsonBuf allocates through
+ * raw malloc/realloc (jb_init/jb_reserve), NOT el_strdup/el_strbuf, so unlike
+ * those helpers it was never registered with the per-request arena — every
+ * JsonBuf-built JSON string leaked unconditionally, whether or not a request
+ * arena was active, because nothing ever freed it. el_arena_track() is a
+ * documented no-op when no arena is active, so this is safe to call from any
+ * context (HTTP request handler, CLI arena-scope, or no arena at all — the
+ * last case still leaks, same as before, until the caller's context is itself
+ * arena-scoped). Do NOT call this on a JsonBuf whose ->buf is consumed by
+ * another builder (jb_puts(&other, x.buf)) and freed manually — only on the
+ * one actually returned via el_wrap_str. */
+static char* jb_finish(JsonBuf* b) {
+    el_arena_track(b->buf);
+    return b->buf;
+}
+
 static void jb_emit_escaped(JsonBuf* b, const char* s) {
     jb_putc(b, '"');
     const unsigned char* p = (const unsigned char*)s;
@@ -3400,7 +3419,7 @@ static void jb_emit_value(JsonBuf* b, el_val_t v) {
 el_val_t json_stringify(el_val_t v) {
     JsonBuf b; jb_init(&b);
     jb_emit_value(&b, v);
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* ── JSON substring accessors ────────────────────────────────────────────── */
@@ -3587,7 +3606,7 @@ el_val_t json_set(el_val_t json_str, el_val_t key, el_val_t value) {
         jb_putc(&b, ':');
         jb_puts(&b, raw_val);
         jb_putc(&b, '}');
-        return el_wrap_str(b.buf);
+        return el_wrap_str(jb_finish(&b));
     }
     const char* existing = json_find_key(json, k);
     JsonBuf b; jb_init(&b);
@@ -3601,7 +3620,7 @@ el_val_t json_set(el_val_t json_str, el_val_t key, el_val_t value) {
         b.buf[b.len] = '\0';
         jb_puts(&b, raw_val);
         jb_puts(&b, end);
-        return el_wrap_str(b.buf);
+        return el_wrap_str(jb_finish(&b));
     }
     /* Insert before closing '}'. Find last '}' */
     size_t jl = strlen(json);
@@ -3633,7 +3652,7 @@ el_val_t json_set(el_val_t json_str, el_val_t key, el_val_t value) {
     jb_puts(&b, raw_val);
     /* Append from close_idx onward */
     jb_puts(&b, json + close_idx);
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t json_array_len(el_val_t json_str) {
@@ -3771,7 +3790,7 @@ el_val_t json_build_object(el_val_t kvs) {
         jb_putc(&b, '"');
     }
     jb_putc(&b, '}');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* json_build_array — build a JSON array from a list of raw JSON values.
@@ -3789,7 +3808,7 @@ el_val_t json_build_array(el_val_t items) {
         jb_puts(&b, vs);
     }
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* ── Time ────────────────────────────────────────────────────────────────── */
@@ -5193,7 +5212,7 @@ el_val_t state_keys(void) {
     }
     jb_putc(&b, ']');
     pthread_mutex_unlock(&_state_mu);
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* Returns 1 (true) if the key is present in the state store, else 0 (false). */
@@ -5399,7 +5418,7 @@ el_val_t str_format(el_val_t fmt, el_val_t data) {
         jb_putc(&b, *p);
         p++;
     }
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t str_lower(el_val_t s) { return str_to_lower(s); }
@@ -5815,7 +5834,7 @@ el_val_t list_join(el_val_t listv, el_val_t sep) {
             jb_puts(&b, tmp);
         }
     }
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t list_range(el_val_t start, el_val_t end) {
@@ -6465,12 +6484,12 @@ el_val_t engram_node(el_val_t content, el_val_t node_type, el_val_t salience) {
     n->id        = engram_new_id();
     const char* c  = EL_CSTR(content);
     const char* nt = EL_CSTR(node_type);
-    n->content   = el_strdup(c ? c : "");
+    n->content   = el_strdup_persist(c ? c : "");
-    n->node_type = el_strdup(nt && *nt ? nt : "Memory");
+    n->node_type = el_strdup_persist(nt && *nt ? nt : "Memory");
-    n->label     = engram_first_n_chars(c, 60);
+    { char* _lb = engram_first_n_chars(c, 60); n->label = el_strdup_persist(_lb); } /* persist: stored field must outlive request arena */
-    n->tier      = el_strdup("Working");
+    n->tier      = el_strdup_persist("Working");
-    n->tags      = el_strdup("");
+    n->tags      = el_strdup_persist("");
-    n->metadata  = el_strdup("{}");
+    n->metadata  = el_strdup_persist("{}");
     n->salience    = engram_decode_score(salience);
     if (n->salience <= 0.0 || n->salience > 1.0) n->salience = 0.5;
     n->importance  = 0.5;
@@ -6597,11 +6616,11 @@ el_val_t engram_node_layered(el_val_t content, el_val_t node_type, el_val_t labe
     const char* lb = EL_CSTR(label);
     const char* tg = EL_CSTR(tags);
     const char* st = EL_CSTR(status);
-    n->content   = el_strdup(c ? c : "");
+    n->content   = el_strdup_persist(c ? c : "");
-    n->node_type = el_strdup(nt && *nt ? nt : "Memory");
+    n->node_type = el_strdup_persist(nt && *nt ? nt : "Memory");
-    n->label     = el_strdup(lb && *lb ? lb : (c ? engram_first_n_chars(c, 60) : ""));
+    n->label     = el_strdup_persist(lb && *lb ? lb : (c ? engram_first_n_chars(c, 60) : ""));
-    n->tier      = el_strdup("Working");
+    n->tier      = el_strdup_persist("Working");
-    n->tags      = el_strdup(tg ? tg : "");
+    n->tags      = el_strdup_persist(tg ? tg : "");
     if (st && *st) {
         /* Minimal metadata payload: {"status":"..."}. Keep it cheap so
          * callers using `status` don't pay JSON parse cost on every read. */
@@ -6610,7 +6629,7 @@ el_val_t engram_node_layered(el_val_t content, el_val_t node_type, el_val_t labe
         snprintf(meta, sl, "{\"status\":\"%s\"}", st);
         n->metadata = meta;
     } else {
-        n->metadata = el_strdup("{}");
+        n->metadata = el_strdup_persist("{}");
     }
     n->salience    = engram_decode_score(salience);
     n->importance  = engram_decode_score(certainty);
@@ -6808,6 +6827,312 @@ static int istr_contains(const char* hay, const char* needle) {
     return 0;
 }
 
+/* ── Tokenized query matching ───────────────────────────────────────────
+ * The engram query surface (search / activate / goal-bias) historically
+ * matched the ENTIRE raw query string as a single case-insensitive
+ * substring via istr_contains(field, q). That is Ctrl-F, not search:
+ * a multi-word query like "windows msi signing" only matched a node whose
+ * text contained that exact contiguous run, so real multi-word queries
+ * returned zero. istr_contains stays as the per-TOKEN primitive; these
+ * helpers split the query on whitespace and match ANY token, then rank by
+ * how many DISTINCT tokens a node covers. Single-token queries are a strict
+ * special case (score is 0 or 1) so single-word callers never regress. */
+#define ENGRAM_MAX_QTOKENS 32
+#define ENGRAM_QTOK_LEN    256
+
+/* Split q on whitespace into up to ENGRAM_MAX_QTOKENS distinct
+ * (case-insensitive) tokens. Returns the token count. Over-long tokens are
+ * truncated to ENGRAM_QTOK_LEN-1; over-count tokens are ignored. */
+static int engram_tokenize_query(const char* q,
+                                 char toks[][ENGRAM_QTOK_LEN], int maxtok) {
+    int n = 0;
+    if (!q) return 0;
+    const char* p = q;
+    while (*p && n < maxtok) {
+        while (*p && isspace((unsigned char)*p)) p++;
+        if (!*p) break;
+        char buf[ENGRAM_QTOK_LEN];
+        size_t tl = 0;
+        while (*p && !isspace((unsigned char)*p)) {
+            if (tl < sizeof(buf) - 1) buf[tl++] = *p;
+            p++;
+        }
+        buf[tl] = '\0';
+        if (tl == 0) continue;
+        int dup = 0;
+        for (int s = 0; s < n; s++) {
+            if (strcasecmp(toks[s], buf) == 0) { dup = 1; break; }
+        }
+        if (dup) continue;
+        memcpy(toks[n], buf, tl + 1);
+        n++;
+    }
+    return n;
+}
+
+/* Count how many of the ntok distinct query tokens appear (case-insensitive)
+ * in the node's content, label, or tags. 0 == no match. */
+static int engram_node_match_score(const EngramNode* n,
+                                   char toks[][ENGRAM_QTOK_LEN], int ntok) {
+    int score = 0;
+    for (int t = 0; t < ntok; t++) {
+        if (istr_contains(n->content, toks[t]) ||
+            istr_contains(n->label,   toks[t]) ||
+            istr_contains(n->tags,    toks[t]))
+            score++;
+    }
+    return score;
+}
+
+/* Rank entry: distinct-token match count (primary, desc) then salience
+ * (tiebreak, desc). */
+typedef struct { int64_t idx; int score; double salience; } EngramRankEntry;
+static int engram_rank_cmp(const void* a, const void* b) {
+    const EngramRankEntry* ea = (const EngramRankEntry*)a;
+    const EngramRankEntry* eb = (const EngramRankEntry*)b;
+    if (ea->score != eb->score) return eb->score - ea->score;      /* desc */
+    if (ea->salience < eb->salience) return 1;
+    if (ea->salience > eb->salience) return -1;
+    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);
@@ -6815,21 +7140,45 @@ el_val_t engram_search(el_val_t query, el_val_t limit) {
     if (lim <= 0) lim = 100;
     el_val_t lst = el_list_empty();
     if (!q || !*q) return lst;
-    int64_t found = 0;
+    char toks[ENGRAM_MAX_QTOKENS][ENGRAM_QTOK_LEN];
-    for (int64_t i = 0; i < g->node_count && found < lim; i++) {
+    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; }
+    int64_t nhits = 0;
+    for (int64_t i = 0; i < g->node_count; i++) {
         EngramNode* n = &g->nodes[i];
         /* Filter transparent layers: nodes whose layer is `transparent=1`
          * shape output but are invisible to introspection ("what do you
          * know about yourself"). They still surface via engram_activate
          * + engram_compile_layered_json — that's the legitimate path. */
         if (engram_layer_is_transparent(n->layer_id)) continue;
-        if (istr_contains(n->content, q) ||
+        int sc = engram_node_match_score(n, toks, ntok);
-            istr_contains(n->label,   q) ||
+        double sem = qvec ? engram_node_cosine(n, qvec, qdim) : 0.0;
-            istr_contains(n->tags,    q)) {
+        if (sc > 0 || sem >= sem_min) {
-            lst = el_list_append(lst, engram_node_to_map(n));
+            hits[nhits].idx      = i;
-            found++;
+            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. */
+    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;
 }
 
@@ -6883,10 +7232,10 @@ void engram_connect(el_val_t from_id, el_val_t to_id, el_val_t weight, el_val_t
     EngramEdge* e = &g->edges[g->edge_count];
     memset(e, 0, sizeof(*e));
     e->id        = engram_new_id();
-    e->from_id   = el_strdup(f);
+    e->from_id   = el_strdup_persist(f);
-    e->to_id     = el_strdup(t);
+    e->to_id     = el_strdup_persist(t);
-    e->relation  = el_strdup(r && *r ? r : "associate");
+    e->relation  = el_strdup_persist(r && *r ? r : "associate");
-    e->metadata  = el_strdup("{}");
+    e->metadata  = el_strdup_persist("{}");
     e->weight    = engram_decode_score(weight);
     if (e->weight <= 0.0 || e->weight > 1.0) e->weight = 0.5;
     e->confidence = 1.0;
@@ -7106,10 +7455,14 @@ static double engram_temporal_proximity_bonus(int64_t node_created,
 static double engram_goal_bias(const EngramNode* n, const char* query) {
     if (!query || !*query) return 1.0;
     double bias = 1.0;
-    /* Direct lexical overlap: node content/label/tags share text with query. */
+    /* Direct lexical overlap, graded by token coverage: a node covering all
-    if (istr_contains(n->content, query) || istr_contains(n->label, query) ||
+     * query tokens gets the full +0.5; partial coverage gets a proportional
-        istr_contains(n->tags, query)) {
+     * share. Single-token queries → full +0.5 on match, identical to before. */
-        bias += 0.5;
+    {
+        char toks[ENGRAM_MAX_QTOKENS][ENGRAM_QTOK_LEN];
+        int ntok = engram_tokenize_query(query, toks, ENGRAM_MAX_QTOKENS);
+        int sc   = engram_node_match_score(n, toks, ntok);
+        if (sc > 0 && ntok > 0) bias += 0.5 * ((double)sc / (double)ntok);
     }
     /* Node-type resonance with query intent. */
     int technical_query = istr_contains(query, "code")         ||
@@ -7175,14 +7528,31 @@ 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. */
+    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];
-        if (istr_contains(n->content, q) ||
+        int sc = engram_node_match_score(n, toks, ntok);
-            istr_contains(n->label,   q) ||
+        double sem = q_vec ? engram_node_cosine(n, q_vec, q_dim) : 0.0;
-            istr_contains(n->tags,    q)) {
+        if (sc > 0 || sem >= q_sem_min) {
             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 */
             seeds[seed_count].idx        = i;
             seeds[seed_count].act        = act;
             seeds[seed_count].created_at = n->created_at;
@@ -7192,6 +7562,7 @@ 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) {
@@ -7519,11 +7890,11 @@ el_val_t engram_save(el_val_t path) {
 /* Helper: extract a string field from a JSON object substring. */
 static char* eg_get_str_field(const char* obj, const char* key) {
     const char* p = json_find_key(obj, key);
-    if (!p) return el_strdup("");
+    if (!p) return el_strdup_persist("");
-    if (*p != '"') return el_strdup("");
+    if (*p != '"') return el_strdup_persist("");
     JsonParser jp = { .p = p, .end = p + strlen(p), .err = 0 };
     char* out = jp_parse_string_raw(&jp);
-    if (jp.err) { free(out); return el_strdup(""); }
+    if (jp.err) { free(out); return el_strdup_persist(""); }
     return out;
 }
 
@@ -7597,7 +7968,7 @@ el_val_t engram_load(el_val_t path) {
                 nn->tier      = eg_get_str_field(obj, "tier");
                 nn->tags      = eg_get_str_field(obj, "tags");
                 nn->metadata  = eg_get_str_field(obj, "metadata");
-                if (!nn->metadata || !*nn->metadata) { free(nn->metadata); nn->metadata = el_strdup("{}"); }
+                if (!nn->metadata || !*nn->metadata) { free(nn->metadata); nn->metadata = el_strdup_persist("{}"); }
                 nn->salience           = eg_get_num_field(obj, "salience");
                 nn->importance         = eg_get_num_field(obj, "importance");
                 nn->confidence         = eg_get_num_field(obj, "confidence");
@@ -7648,7 +8019,7 @@ el_val_t engram_load(el_val_t path) {
                 ee->to_id    = eg_get_str_field(obj, "to_id");
                 ee->relation = eg_get_str_field(obj, "relation");
                 ee->metadata = eg_get_str_field(obj, "metadata");
-                if (!ee->metadata || !*ee->metadata) { free(ee->metadata); ee->metadata = el_strdup("{}"); }
+                if (!ee->metadata || !*ee->metadata) { free(ee->metadata); ee->metadata = el_strdup_persist("{}"); }
                 ee->weight     = eg_get_num_field(obj, "weight");
                 ee->confidence = eg_get_num_field(obj, "confidence");
                 ee->created_at = eg_get_int_field(obj, "created_at");
@@ -7740,7 +8111,36 @@ el_val_t engram_get_node_json(el_val_t id) {
     if (!n) return el_wrap_str(el_strdup("{}"));
     JsonBuf b; jb_init(&b);
     engram_emit_node_json(&b, n);
-    return el_wrap_str(b.buf);
+    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) {
@@ -7750,25 +8150,53 @@ 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, '[');
-    int first = 1;
+    if (q && *q && g->node_count > 0) {
-    int64_t found = 0;
+        /* Collect candidates from the UNION of tokenized-lexical and semantic
-    if (q && *q) {
+         * matches, score each, rank by score, emit the top `lim`. A node is a
-        for (int64_t i = 0; i < g->node_count && found < lim; i++) {
+         * candidate if it covers >=1 query token (tokenized-lexical, #66) OR its
-            EngramNode* n = &g->nodes[i];
+         * query cosine clears the threshold (#67). Lexical score is the distinct
-            /* Filter transparent layers — same as engram_search. */
+         * token count (>=1), so any lexical hit outranks a pure-semantic hit
-            if (engram_layer_is_transparent(n->layer_id)) continue;
+         * (cosine < 1); pure-semantic hits are scored by cosine alone. When
-            if (istr_contains(n->content, q) ||
+         * semantic is unavailable qvec is NULL, sem is 0, only tokenized-lexical
-                istr_contains(n->label,   q) ||
+         * hits are collected, and the stable insertion sort preserves order. */
-                istr_contains(n->tags,    q)) {
+        char toks[ENGRAM_MAX_QTOKENS][ENGRAM_QTOK_LEN];
-                if (!first) jb_putc(&b, ',');
+        int ntok = engram_tokenize_query(q, toks, ENGRAM_MAX_QTOKENS);
-                engram_emit_node_json(&b, n);
+        int qdim = 0;
-                first = 0;
+        float* qvec = engram_embed_query(q, &qdim);
-                found++;
+        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++;
+                }
             }
+            /* 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(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t engram_scan_nodes_json(el_val_t limit, el_val_t offset) {
@@ -7777,9 +8205,9 @@ el_val_t engram_scan_nodes_json(el_val_t limit, el_val_t offset) {
     int64_t off = (int64_t)offset; if (off < 0) off = 0;
     JsonBuf b; jb_init(&b);
     jb_putc(&b, '[');
-    if (g->node_count == 0) { jb_putc(&b, ']'); return el_wrap_str(b.buf); }
+    if (g->node_count == 0) { jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b)); }
     int64_t* idx = malloc((size_t)g->node_count * sizeof(int64_t));
-    if (!idx) { jb_putc(&b, ']'); return el_wrap_str(b.buf); }
+    if (!idx) { jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b)); }
     /* Skip transparent layers — introspection filter, same as engram_scan_nodes. */
     int64_t live = 0;
     for (int64_t i = 0; i < g->node_count; i++) {
@@ -7797,7 +8225,7 @@ el_val_t engram_scan_nodes_json(el_val_t limit, el_val_t offset) {
     }
     free(idx);
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* engram_scan_nodes_by_type_json — filter by node_type before paginating.
@@ -7813,9 +8241,9 @@ el_val_t engram_scan_nodes_by_type_json(el_val_t type_v, el_val_t limit, el_val_
     int64_t off = (int64_t)offset; if (off < 0) off = 0;
     JsonBuf b; jb_init(&b);
     jb_putc(&b, '[');
-    if (g->node_count == 0) { jb_putc(&b, ']'); return el_wrap_str(b.buf); }
+    if (g->node_count == 0) { jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b)); }
     int64_t* idx = malloc((size_t)g->node_count * sizeof(int64_t));
-    if (!idx) { jb_putc(&b, ']'); return el_wrap_str(b.buf); }
+    if (!idx) { jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b)); }
     int64_t live = 0;
     for (int64_t i = 0; i < g->node_count; i++) {
         if (engram_layer_is_transparent(g->nodes[i].layer_id)) continue;
@@ -7834,7 +8262,7 @@ el_val_t engram_scan_nodes_by_type_json(el_val_t type_v, el_val_t limit, el_val_
     }
     free(idx);
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t direction) {
@@ -7849,7 +8277,7 @@ el_val_t engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t di
     int allow_in  = (strcmp(dir, "in")  == 0) || (strcmp(dir, "both") == 0);
     JsonBuf b; jb_init(&b);
     jb_putc(&b, '[');
-    if (!sid || !*sid) { jb_putc(&b, ']'); return el_wrap_str(b.buf); }
+    if (!sid || !*sid) { jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b)); }
 
     /* Frontier of (node_id, hops). Cap to a sane size. */
     char** frontier = calloc(1024, sizeof(char*));
@@ -7859,7 +8287,7 @@ el_val_t engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t di
     int64_t vc = 0;
     if (!frontier || !frontier_h || !visited) {
         free(frontier); free(frontier_h); free(visited);
-        jb_putc(&b, ']'); return el_wrap_str(b.buf);
+        jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b));
     }
     /* Use plain strdup (not el_strdup) so arena doesn't track these pointers.
      * The BFS loop manually frees them below — arena would double-free them. */
@@ -7902,7 +8330,7 @@ el_val_t engram_neighbors_json(el_val_t node_id, el_val_t max_depth, el_val_t di
     for (int64_t i = 0; i < vc; i++) free(visited[i]);
     free(frontier); free(frontier_h); free(visited);
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t engram_activate_json(el_val_t query, el_val_t depth) {
@@ -7943,7 +8371,7 @@ el_val_t engram_activate_json(el_val_t query, el_val_t depth) {
         }
     }
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 el_val_t engram_stats_json(void) {
@@ -7963,7 +8391,7 @@ el_val_t engram_list_layers_json(void) {
     jb_putc(&b, '[');
     /* Build a sorted index over live layers. */
     size_t* idx = malloc((g->layer_count + 1) * sizeof(size_t));
-    if (!idx) { jb_putc(&b, ']'); return el_wrap_str(b.buf); }
+    if (!idx) { jb_putc(&b, ']'); return el_wrap_str(jb_finish(&b)); }
     size_t live = 0;
     for (size_t i = 0; i < g->layer_count; i++) {
         if (g->layers[i].name) idx[live++] = i;
@@ -8000,7 +8428,7 @@ el_val_t engram_list_layers_json(void) {
     }
     free(idx);
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* engram_compile_layered_json — produce a prompt-ready context block split
@@ -8086,7 +8514,7 @@ el_val_t engram_compile_layered_json(el_val_t intent, el_val_t depth) {
         free(b.buf);
         return el_wrap_str(el_strdup(""));
     }
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* engram_query_range — temporal range query.
@@ -8131,7 +8559,7 @@ el_val_t engram_query_range(el_val_t start_ms_v, el_val_t end_ms_v) {
     }
     jb_putc(&b, ']');
     free(idx);
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 /* engram_load_merge — like engram_load but WITHOUT resetting the store.
@@ -8382,7 +8810,7 @@ el_val_t engram_wm_top_json(el_val_t n_v) {
     }
     free(idx);
     jb_putc(&b, ']');
-    return el_wrap_str(b.buf);
+    return el_wrap_str(jb_finish(&b));
 }
 
 #ifdef HAVE_CURL
@@ -8778,13 +9206,13 @@ static int64_t dharma_find_or_create_relation_edge(const char* peer_base, int cr
         engram_grow_nodes();
         EngramNode* n = &g->nodes[g->node_count];
         memset(n, 0, sizeof(*n));
-        n->id        = el_strdup(self_id);
+        n->id        = el_strdup_persist(self_id);
-        n->content   = el_strdup(_el_cgi_dharma_id ? _el_cgi_dharma_id : "(self)");
+        n->content   = el_strdup_persist(_el_cgi_dharma_id ? _el_cgi_dharma_id : "(self)");
-        n->node_type = el_strdup("DharmaSelf");
+        n->node_type = el_strdup_persist("DharmaSelf");
-        n->label     = el_strdup("dharma:self");
+        n->label     = el_strdup_persist("dharma:self");
-        n->tier      = el_strdup("Working");
+        n->tier      = el_strdup_persist("Working");
-        n->tags      = el_strdup("dharma");
+        n->tags      = el_strdup_persist("dharma");
-        n->metadata  = el_strdup("{}");
+        n->metadata  = el_strdup_persist("{}");
         n->salience  = 1.0; n->importance = 1.0; n->confidence = 1.0;
         int64_t now  = engram_now_ms();
         n->created_at = now; n->updated_at = now; n->last_activated = now;
@@ -8795,13 +9223,13 @@ static int64_t dharma_find_or_create_relation_edge(const char* peer_base, int cr
         engram_grow_nodes();
         EngramNode* n = &g->nodes[g->node_count];
         memset(n, 0, sizeof(*n));
-        n->id        = el_strdup(peer_node);
+        n->id        = el_strdup_persist(peer_node);
-        n->content   = el_strdup(peer_base);
+        n->content   = el_strdup_persist(peer_base);
-        n->node_type = el_strdup("DharmaPeer");
+        n->node_type = el_strdup_persist("DharmaPeer");
-        n->label     = el_strdup(peer_node);
+        n->label     = el_strdup_persist(peer_node);
-        n->tier      = el_strdup("Working");
+        n->tier      = el_strdup_persist("Working");
-        n->tags      = el_strdup("dharma");
+        n->tags      = el_strdup_persist("dharma");
-        n->metadata  = el_strdup("{}");
+        n->metadata  = el_strdup_persist("{}");
         n->salience  = 0.5; n->importance = 0.5; n->confidence = 1.0;
         int64_t now  = engram_now_ms();
         n->created_at = now; n->updated_at = now; n->last_activated = now;
@@ -8813,10 +9241,10 @@ static int64_t dharma_find_or_create_relation_edge(const char* peer_base, int cr
     EngramEdge* e = &g->edges[g->edge_count];
     memset(e, 0, sizeof(*e));
     e->id        = engram_new_id();
-    e->from_id   = el_strdup(self_id);
+    e->from_id   = el_strdup_persist(self_id);
-    e->to_id     = el_strdup(peer_node);
+    e->to_id     = el_strdup_persist(peer_node);
-    e->relation  = el_strdup("dharma-relation");
+    e->relation  = el_strdup_persist("dharma-relation");
-    e->metadata  = el_strdup("{}");
+    e->metadata  = el_strdup_persist("{}");
     e->weight    = 0.0;
     e->confidence = 1.0;
     int64_t now  = engram_now_ms();
@@ -9109,7 +9537,7 @@ static el_val_t llm_extract_text(el_val_t resp_val) {
         free(obj);
         p = end;
     }
-    return el_wrap_str(out.buf);
+    return el_wrap_str(jb_finish(&out));
 }
 
 el_val_t llm_call(el_val_t model, el_val_t prompt) {

lang/el-compiler/runtime/el_runtime.h

@@ -632,6 +632,7 @@ 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

@@ -1072,6 +1072,7 @@ 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

@@ -226,6 +226,7 @@ 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

@@ -2670,6 +2670,7 @@ 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 }

lang/el-compiler/src/parser.el

@@ -23,10 +23,29 @@ fn tok_at(tokens: [Any], pos: Int) -> Map<String, Any> {
 }
 
 fn tok_kind(tokens: [Any], pos: Int) -> String {
+    // Out-of-range reads must report the Eof sentinel so every `== "Eof"`
+    // termination guard in the parser fires. Without this, reading past the
+    // single trailing Eof token returns runtime null (el_list_get OOB -> 0),
+    // which matches no delimiter, letting inner parse loops append AST nodes
+    // forever on malformed input -> unbounded allocation -> OOM.
+    let n: Int = native_list_len(tokens) / 2
+    if pos < 0 {
+        return "Eof"
+    }
+    if pos >= n {
+        return "Eof"
+    }
     native_list_get(tokens, pos * 2)
 }
 
 fn tok_value(tokens: [Any], pos: Int) -> String {
+    let n: Int = native_list_len(tokens) / 2
+    if pos < 0 {
+        return ""
+    }
+    if pos >= n {
+        return ""
+    }
     native_list_get(tokens, pos * 2 + 1)
 }
 
@@ -35,7 +54,12 @@ fn expect(tokens: [Any], pos: Int, kind: String) -> Int {
     if k == kind {
         return pos + 1
     }
-    // On mismatch just advance; error recovery is best-effort
+    // On mismatch, error recovery is best-effort. But never step PAST the Eof
+    // sentinel: once at Eof a mismatch means the input ended early, and
+    // advancing would run the cursor off the token list.
+    if k == "Eof" {
+        return pos
+    }
     pos + 1
 }