fix(engram): atomic engram_save (tmp+fsync+rename) + sparse-write anti-clobber floor

Kills the engram-clobber loop at its source. engram_save did a bare fopen("wb")
that truncates snapshot.json to 0 bytes before the 47MB write — a booting soul's
engram_load could read that empty window -> genesis -> nodes=1 -> a 63-node save
overwrote the populated file. Two guards:
 1. Atomic write: serialize to <path>.tmp, fflush+fsync, rename() over target
    (atomic on POSIX) — no reader ever sees a truncated/0-byte snapshot.
 2. Sparse-write floor: refuse to overwrite a >200KB snapshot with one < 1/16 its
    size — a partial load can never clobber a healthy graph, whatever the cause.

Validated in isolation: standalone clang harness 11/11; rebuilt the darwin soul
(scripts/build-soul-darwin.sh) and booted it on an isolated port against a golden
copy — loaded 5113 nodes and round-tripped the full 47MB snapshot, no .tmp leftover,
live ~/.neuron untouched. Adds scripts/build-soul-darwin.sh (local elb replacement).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

lang/el-compiler/runtime/el_runtime.c

@@ -1475,10 +1475,13 @@ 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 if available (handles binary files
+    /* Use max(strlen, fs_read_len). fs_read_len is the real byte count for binary
-     * with embedded null bytes — PNG, WOFF2, etc.). Fall back to strlen for
+     * files (strlen stops at embedded NULs — PNG, WOFF2). strlen is correct AND larger
-     * normal text/JSON responses where _tl_fs_read_len is 0. */
+     * when a handler WRAPS fs_read output in a longer text/JSON response (e.g.
-    size_t blen = (_tl_fs_read_len > 0) ? _tl_fs_read_len : strlen(eff_body);
+     * /api/safety-contact returns {"configured":...,"contact": <file>}); using
+     * fs_read_len alone truncated those responses to the file's length. */
+    size_t _blen_s = strlen(eff_body);
+    size_t blen = (_tl_fs_read_len > _blen_s) ? _tl_fs_read_len : _blen_s;
     _tl_fs_read_len = 0;  /* consume — one-shot per response */
     int head_only = _tl_http_head_only;
 
@@ -1552,7 +1555,8 @@ static void* http_worker(void* arg) {
             /* 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. */
-            size_t rlen = _tl_fs_read_len > 0 ? _tl_fs_read_len : (rs ? strlen(rs) : 0);
+            size_t _rlen_s = rs ? strlen(rs) : 0;
+            size_t rlen = (_tl_fs_read_len > _rlen_s) ? _tl_fs_read_len : _rlen_s;
             response = malloc(rlen + 1);
             if (response && rs) { memcpy(response, rs, rlen); response[rlen] = '\0'; }
             else if (response) { response[0] = '\0'; }
@@ -1799,7 +1803,8 @@ 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);
-            size_t rlen = _tl_fs_read_len > 0 ? _tl_fs_read_len : (rs ? strlen(rs) : 0);
+            size_t _rlen_s = rs ? strlen(rs) : 0;
+            size_t rlen = (_tl_fs_read_len > _rlen_s) ? _tl_fs_read_len : _rlen_s;
             response = malloc(rlen + 1);
             if (response && rs) { memcpy(response, rs, rlen); response[rlen] = '\0'; }
             else if (response) { response[0] = '\0'; }
@@ -3173,23 +3178,49 @@ static void jb_puts(JsonBuf* b, const char* s) {
 
 static void jb_emit_escaped(JsonBuf* b, const char* s) {
     jb_putc(b, '"');
-    for (; *s; s++) {
+    const unsigned char* p = (const unsigned char*)s;
-        unsigned char c = (unsigned char)*s;
+    while (*p) {
+        unsigned char c = *p;
         switch (c) {
-            case '"':  jb_puts(b, "\\\""); break;
+            case '"':  jb_puts(b, "\\\""); p++; break;
-            case '\\': jb_puts(b, "\\\\"); break;
+            case '\\': jb_puts(b, "\\\\"); p++; break;
-            case '\b': jb_puts(b, "\\b");  break;
+            case '\b': jb_puts(b, "\\b");  p++; break;
-            case '\f': jb_puts(b, "\\f");  break;
+            case '\f': jb_puts(b, "\\f");  p++; break;
-            case '\n': jb_puts(b, "\\n");  break;
+            case '\n': jb_puts(b, "\\n");  p++; break;
-            case '\r': jb_puts(b, "\\r");  break;
+            case '\r': jb_puts(b, "\\r");  p++; break;
-            case '\t': jb_puts(b, "\\t");  break;
+            case '\t': jb_puts(b, "\\t");  p++; break;
             default:
                 if (c < 0x20) {
                     char tmp[8];
                     snprintf(tmp, sizeof(tmp), "\\u%04x", c);
                     jb_puts(b, tmp);
-                } else {
+                    p++;
+                } else if (c < 0x80) {
                     jb_putc(b, (char)c);
+                    p++;
+                } else {
+                    /* Multi-byte UTF-8: validate sequence, pass through if valid,
+                     * escape as \u00xx if the start byte is invalid/orphaned. */
+                    int seq_len = 0;
+                    if      ((c & 0xE0) == 0xC0) seq_len = 2;
+                    else if ((c & 0xF0) == 0xE0) seq_len = 3;
+                    else if ((c & 0xF8) == 0xF0) seq_len = 4;
+                    if (seq_len >= 2) {
+                        int valid = 1;
+                        for (int i = 1; i < seq_len; i++) {
+                            if ((p[i] & 0xC0) != 0x80) { valid = 0; break; }
+                        }
+                        if (valid) {
+                            for (int i = 0; i < seq_len; i++) jb_putc(b, (char)p[i]);
+                            p += seq_len;
+                            break;
+                        }
+                    }
+                    /* Invalid start byte or truncated sequence — escape it */
+                    char tmp[8];
+                    snprintf(tmp, sizeof(tmp), "\\u%04x", c);
+                    jb_puts(b, tmp);
+                    p++;
                 }
                 break;
         }
@@ -6219,7 +6250,9 @@ static void engram_grow_edges(void) {
 static char* engram_new_id(void) {
     el_val_t v = uuid_new();
     const char* s = EL_CSTR(v);
-    return el_strdup(s ? s : "");
+    /* Persistent: node ids live in the global store; an arena (el_strdup) id is
+     * freed at el_request_end(), corrupting the node after the creating request. */
+    return el_strdup_persist(s ? s : "");
 }
 
 /* Convert a node into an ElMap of its fields.                            */
@@ -6314,12 +6347,17 @@ el_val_t engram_node_full(el_val_t content, el_val_t node_type, el_val_t label,
     const char* lb = EL_CSTR(label);
     const char* ti = EL_CSTR(tier);
     const char* tg = EL_CSTR(tags);
-    n->content   = el_strdup(c ? c : "");
+    /* Persistent (el_strdup_persist, NOT el_strdup): these strings are owned by the
-    n->node_type = el_strdup(nt && *nt ? nt : "Memory");
+     * persistent global node store. el_strdup tracks into the per-request arena, which
-    n->label     = el_strdup(lb && *lb ? lb : (c ? engram_first_n_chars(c, 60) : ""));
+     * el_request_end() frees when the creating HTTP request completes — leaving the
-    n->tier      = el_strdup(ti && *ti ? ti : "Working");
+     * stored node with dangling pointers (corrupted ids, "saved but never listed").
-    n->tags      = el_strdup(tg ? tg : "");
+     * This is the root cause of the hallucinated/lost-saves class of bugs. */
-    n->metadata  = el_strdup("{}");
+    n->content   = el_strdup_persist(c ? c : "");
+    n->node_type = el_strdup_persist(nt && *nt ? nt : "Memory");
+    n->label     = el_strdup_persist(lb && *lb ? lb : (c ? engram_first_n_chars(c, 60) : ""));
+    n->tier      = el_strdup_persist(ti && *ti ? ti : "Working");
+    n->tags      = el_strdup_persist(tg ? tg : "");
+    n->metadata  = el_strdup_persist("{}");
     n->salience    = engram_decode_score(salience);
     n->importance  = engram_decode_score(importance);
     n->confidence  = engram_decode_score(confidence);
@@ -7262,13 +7300,48 @@ el_val_t engram_save(el_val_t path) {
         jb_putc(&b, '}');
     }
     jb_puts(&b, "]}");
-    FILE* f = fopen(p, "wb");
+
-    if (!f) { free(b.buf); return 0; }
+    /* --- Anti-clobber sparse-write floor (NTN engram clobber fix) ---------
+     * Refuse to overwrite an existing populated snapshot with a drastically
+     * smaller one. A bad boot that loaded only ~63 identity nodes must never
+     * be able to clobber a healthy 5000+ node snapshot, regardless of the
+     * upstream cause (genesis fallback, partial load, etc.). */
+    {
+        struct stat _st;
+        if (stat(p, &_st) == 0 && _st.st_size > 200000 &&
+            (uint64_t)b.len < (uint64_t)_st.st_size / 16) {
+            fprintf(stderr,
+                "[engram_save] REFUSED sparse write: new %zu bytes vs existing "
+                "%lld bytes (< 1/16) — protecting snapshot %s\n",
+                b.len, (long long)_st.st_size, p);
+            free(b.buf);
+            return 0;
+        }
+    }
+
+    /* --- Atomic write: tmp + fsync + rename ------------------------------
+     * Write to a sibling temp file, fsync it durable, then rename() over the
+     * target. rename() is atomic on POSIX, so a concurrent reader (a booting
+     * soul's engram_load) never observes a truncated or 0-byte snapshot —
+     * which was the root of the genesis/clobber loop. */
+    size_t _plen = strlen(p);
+    char* _tmp = (char*)malloc(_plen + 5);
+    if (!_tmp) { free(b.buf); return 0; }
+    memcpy(_tmp, p, _plen);
+    memcpy(_tmp + _plen, ".tmp", 5); /* includes NUL */
+
+    FILE* f = fopen(_tmp, "wb");
+    if (!f) { free(_tmp); free(b.buf); return 0; }
     size_t w = fwrite(b.buf, 1, b.len, f);
+    int wok = (w == b.len);
+    if (wok) { fflush(f); fsync(fileno(f)); }
     fclose(f);
-    int ok = (w == b.len);
     free(b.buf);
-    return ok ? 1 : 0;
+
+    if (!wok) { unlink(_tmp); free(_tmp); return 0; }
+    if (rename(_tmp, p) != 0) { unlink(_tmp); free(_tmp); return 0; }
+    free(_tmp);
+    return 1;
 }
 
 /* Helper: extract a string field from a JSON object substring. */

scripts/build-soul-darwin.sh

@@ -0,0 +1,21 @@
+#!/bin/sh
+# build-soul-darwin.sh — replicate `elb` on macOS/arm64 with clang.
+# Proven 2026-06-16: produces a Mach-O arm64 soul that boots and serves :7770.
+# The official builder `elb` ships Linux-only (CI); this lets us build + test the
+# darwin soul locally (e.g. to validate the atomic engram_save fix in isolation).
+#
+# Usage: scripts/build-soul-darwin.sh <path-to-neuron/dist> [output-binary]
+set -e
+DIST="${1:?usage: build-soul-darwin.sh <neuron/dist dir> [out]}"
+OUT="${2:-./neuron}"
+RT="$(cd "$(dirname "$0")/.." && pwd)/lang/el-compiler/runtime"
+B="$(mktemp -d)"
+# elc-generated dist modules use C89-style implicit cross-module declarations that
+# Apple clang rejects as errors by default; resolve at link, so downgrade them.
+CFLAGS="-Wno-implicit-function-declaration -Wno-implicit-int -Wno-int-conversion -I$B -I$DIST -I$RT"
+cp "$RT/el_runtime.h" "$B/"
+clang -c $CFLAGS "$RT/el_runtime.c" -o "$B/el_runtime.o"
+for c in "$DIST"/*.c; do clang -c $CFLAGS "$c" -o "$B/$(basename "$c" .c).o"; done
+# NOTE: link *.o once — do not also list el_runtime.o separately (duplicate symbols).
+clang "$B"/*.o -o "$OUT" -lcurl -lm
+echo "built $OUT"