fix(datalog): reject forged dl-program handles instead of dereferencing them

src/mem/heap.h

@@ -91,6 +91,15 @@
  * Checked by HNSW builtins before dereferencing.  User must (hnsw-free h). */
 #define RAY_ATTR_HNSW         0x04
 
+/* I64 atom carries an owning dl_program_t* (a Datalog program) in its .i64
+ * slot.  Checked by the dl-* builtins before dereferencing, so a forged/plain
+ * integer or an arithmetic copy (which drops attrs) is rejected with a type
+ * error instead of being reinterpreted as a pointer — see dl_unwrap_program.
+ * Reuses 0x20 (RAY_ATTR_SORTED on vectors / ATTR_QUOTED on -RAY_SYM); the
+ * -RAY_I64 type tag disambiguates, and no free-path or generic check reads
+ * 0x20 on a -RAY_I64 atom.  User must (dl-free h). */
+#define RAY_ATTR_DLPROG       0x20
+
 /* Vector is a linked column.  The 8 bytes of the aux union at offset
  * 8 (i.e. parent->_idx_pad / parent->slice_offset / parent->str_pool
  * slot, depending on which arm is in use) hold an int64

src/ops/datalog.c

@@ -4409,18 +4409,23 @@ ray_t* ray_query_fn(ray_t** args, int64_t n) {
  * Programmatic Datalog API builtins
  * ══════════════════════════════════════════ */
 
-/* Opaque handle for dl_program_t stored in a ray_t atom.
- * We store the pointer in the i64 field. */
+/* Opaque handle for dl_program_t stored in a ray_t atom.  The pointer lives
+ * in the i64 field, tagged with RAY_ATTR_DLPROG so dl_unwrap_program can reject
+ * a plain integer (or an arithmetic copy, which does not carry the attr)
+ * instead of dereferencing an arbitrary value as a dl_program_t*.  Mirrors the
+ * RAY_ATTR_GRAPH / RAY_ATTR_HNSW handle scheme. */
 static ray_t* dl_wrap_program(dl_program_t* prog) {
     ray_t* obj = ray_alloc(0);
     if (!obj || RAY_IS_ERR(obj)) return ray_error("oom", NULL);
     obj->type = -RAY_I64;
     obj->i64 = (int64_t)(uintptr_t)prog;
+    obj->attrs |= RAY_ATTR_DLPROG;
     return obj;
 }
 
 static dl_program_t* dl_unwrap_program(ray_t* obj) {
-    if (!obj || obj->type != -RAY_I64) return NULL;
+    if (!obj || obj->type != -RAY_I64 || !(obj->attrs & RAY_ATTR_DLPROG))
+        return NULL;
     return (dl_program_t*)(uintptr_t)obj->i64;
 }
 
@@ -4514,16 +4519,22 @@ ray_t* ray_dl_provenance_fn(ray_t* prog_obj, ray_t* pred_obj) {
 /* (dl-free prog) — free a dl-program handle created by (dl-program).
  * The handle wraps a raw dl_program_t* with no automatic finalizer, so
  * without an explicit free the program block and all relation tables it
- * owns leak. Idempotent: zeroes the handle so a second call is a no-op
- * (returns false) instead of double-freeing. */
+ * owns leak.  On free we clear RAY_ATTR_DLPROG and zero the pointer, so a
+ * second call is an idempotent no-op (returns false) rather than a
+ * double-free.  A tagged pointer that isn't an i64 is a type error, and a
+ * plain/forged integer (RAY_ATTR_DLPROG absent, non-zero value) is rejected
+ * as a type error instead of being dereferenced as a dl_program_t*. */
 ray_t* ray_dl_free_fn(ray_t* x) {
     if (!x || x->type != -RAY_I64)
         return ray_error("type", "dl-free: arg must be a dl-program");
-    dl_program_t* prog = dl_unwrap_program(x);
-    if (!prog) return ray_bool(false);  /* already freed / null handle */
-    dl_program_free(prog);
-    x->i64 = 0;
-    return ray_bool(true);
+    if (x->attrs & RAY_ATTR_DLPROG) {
+        dl_program_free((dl_program_t*)(uintptr_t)x->i64);
+        x->attrs &= (uint8_t)~RAY_ATTR_DLPROG;
+        x->i64 = 0;
+        return ray_bool(true);
+    }
+    if (x->i64 == 0) return ray_bool(false);  /* already-freed handle: idempotent */
+    return ray_error("type", "dl-free: not a dl-program handle");
 }
 
 /* Reset global Datalog rule storage (called from ray_lang_destroy) */

test/rfl/datalog/datalog_coverage.rfl

@@ -544,6 +544,24 @@
 (dl-query FP 'edge) !- type
 (dl-free 'notprog) !- type
 
+;; --- Claim 5b: forged/copied dl-program handles must not be dereferenced ---
+;; A plain integer is NOT a dl-program handle.  Every dl-* builtin that unwraps
+;; one must reject it with a type error rather than reinterpreting the value as
+;; a dl_program_t* pointer — before the RAY_ATTR_DLPROG tag check, (dl-free 1)
+;; and friends dereferenced address 1 and crashed with SIGSEGV.
+(dl-free 1) !- type
+(dl-stratify 1) !- type
+(dl-eval 1) !- type
+(dl-query 1 'edge) !- type
+(dl-add-edb 1 'edge (table ['a 'b] (list [1] [2])) 2) !- type
+;; An arithmetic copy of a live handle drops the tag (attrs are not carried
+;; through arithmetic), so it can't be freed — which would otherwise
+;; double-free the original.  The original still frees exactly once.
+(set CP (dl-program))
+(set CPcopy (+ CP 0))
+(dl-free CPcopy) !- type
+(dl-free CP) -- true
+
 
 ;; --- Claim 4: fixpoint non-convergence must error loudly, not return a partial result ---
 ;; Reachability over a linear chain needs ~N iterations. The loop caps at 1000.