diff --git a/source/mach.c b/source/mach.c
index ea08beda..fa74b598 100644
--- a/source/mach.c
+++ b/source/mach.c
@@ -536,13 +536,10 @@ JSValue js_new_register_function(JSContext *ctx, JSCodeRegister *code, JSValue e
   return out;
 }
 
-/* Create a native (QBE-compiled) function */
-JSValue js_new_native_function(JSContext *ctx, void *fn_ptr, void *dl_handle,
-                               uint16_t nr_slots, int arity, JSValue outer_frame) {
+JSValue js_new_native_function_with_code(JSContext *ctx, JSValue code_obj, int arity, JSValue outer_frame) {
   JSGCRef frame_ref;
   JSGCRef fn_ref;
   JSFunction *fn;
-  JSValue code_obj;
   JS_PushGCRef(ctx, &frame_ref);
   frame_ref.val = outer_frame;
   JS_AddGCRef(ctx, &fn_ref);
@@ -560,12 +557,6 @@ JSValue js_new_native_function(JSContext *ctx, void *fn_ptr, void *dl_handle,
   fn->kind = JS_FUNC_KIND_NATIVE;
   fn->length = arity;
   fn->name = JS_NULL;
-  code_obj = js_new_native_code(ctx, fn_ptr, dl_handle, nr_slots, arity);
-  if (JS_IsException(code_obj)) {
-    JS_DeleteGCRef(ctx, &fn_ref);
-    JS_PopGCRef(ctx, &frame_ref);
-    return JS_EXCEPTION;
-  }
   fn = JS_VALUE_GET_FUNCTION(fn_ref.val);
   fn->u.cell.code = code_obj;
   fn->u.cell.env_record = JS_NULL;
@@ -577,6 +568,15 @@ JSValue js_new_native_function(JSContext *ctx, void *fn_ptr, void *dl_handle,
   return out;
 }
 
+/* Create a native (QBE-compiled) function */
+JSValue js_new_native_function(JSContext *ctx, void *fn_ptr, void *dl_handle,
+                               uint16_t nr_slots, int arity, JSValue outer_frame) {
+  JSValue code_obj = js_new_native_code(ctx, fn_ptr, dl_handle, nr_slots, arity);
+  if (JS_IsException(code_obj))
+    return JS_EXCEPTION;
+  return js_new_native_function_with_code(ctx, code_obj, arity, outer_frame);
+}
+
 /* Binary operations helper */
 static JSValue reg_vm_binop(JSContext *ctx, int op, JSValue a, JSValue b) {
   /* Fast path for integers */
diff --git a/source/qbe_helpers.c b/source/qbe_helpers.c
index 9f96f2b9..ba90379a 100644
--- a/source/qbe_helpers.c
+++ b/source/qbe_helpers.c
@@ -12,12 +12,6 @@
 #include <stdlib.h>
 #include <string.h>
 
-#if defined(__GNUC__) || defined(__clang__)
-#define CELL_THREAD_LOCAL __thread
-#else
-#define CELL_THREAD_LOCAL _Thread_local
-#endif
-
 /* Non-inline wrappers for static inline functions in quickjs.h */
 JSValue qbe_new_float64(JSContext *ctx, double d) {
   return __JS_NewFloat64(ctx, d);
@@ -230,36 +224,72 @@ JSValue qbe_shift_shr(JSContext *ctx, JSValue a, JSValue b) {
 
 /* --- Property access --- */
 
-/* Current module handle for active native dispatch. */
-static CELL_THREAD_LOCAL void *g_current_dl_handle = NULL;
-
 typedef struct {
   void *dl_handle;
   JSContext *ctx;
   JSGCRef *vals;
   int count;
 } AOTLiteralPool;
+typedef struct {
+  const char *name;
+  JSValue key;
+} AOTKeyCacheEntry;
+typedef struct {
+  void *dl_handle;
+  int64_t fn_idx;
+  JSValue code;
+} AOTCodeCacheEntry;
 
-static CELL_THREAD_LOCAL AOTLiteralPool g_aot_lit_pool = {0};
+typedef struct AOTGCRefChunk AOTGCRefChunk;
 
-static void aot_clear_lit_pool(void) {
-  if (g_aot_lit_pool.vals) {
-    if (g_aot_lit_pool.ctx) {
-      for (int i = 0; i < g_aot_lit_pool.count; i++)
-        JS_DeleteGCRef(g_aot_lit_pool.ctx, &g_aot_lit_pool.vals[i]);
-    }
-    free(g_aot_lit_pool.vals);
+typedef struct {
+  void *current_dl_handle;
+  AOTLiteralPool lit_pool;
+  AOTKeyCacheEntry *key_cache;
+  int key_cache_count;
+  int key_cache_cap;
+  AOTCodeCacheEntry *code_cache;
+  int code_cache_count;
+  int code_cache_cap;
+  JSGCRef native_env_ref;
+  int has_native_env;
+  int native_env_ref_inited;
+  AOTGCRefChunk **gc_ref_chunks;
+  int gc_ref_chunk_count;
+  int aot_depth;
+  JSValue pending_callee_frame;
+  int pending_is_tail;
+} NativeRTState;
+
+static NativeRTState *native_state(JSContext *ctx) {
+  NativeRTState *st = (NativeRTState *)ctx->native_state;
+  if (st) return st;
+  st = js_mallocz_rt(sizeof(*st));
+  if (!st) {
+    JS_RaiseOOM(ctx);
+    return NULL;
   }
-  g_aot_lit_pool.dl_handle = NULL;
-  g_aot_lit_pool.ctx = NULL;
-  g_aot_lit_pool.vals = NULL;
-  g_aot_lit_pool.count = 0;
+  ctx->native_state = st;
+  return st;
 }
 
-static int aot_load_lit_pool(JSContext *ctx, void *dl_handle) {
-  aot_clear_lit_pool();
-  g_aot_lit_pool.dl_handle = dl_handle;
-  g_aot_lit_pool.ctx = ctx;
+static void aot_clear_lit_pool(JSContext *ctx, NativeRTState *st) {
+  if (!st) return;
+  if (st->lit_pool.vals) {
+    for (int i = 0; i < st->lit_pool.count; i++)
+      JS_DeleteGCRef(ctx, &st->lit_pool.vals[i]);
+    free(st->lit_pool.vals);
+  }
+  st->lit_pool.dl_handle = NULL;
+  st->lit_pool.ctx = NULL;
+  st->lit_pool.vals = NULL;
+  st->lit_pool.count = 0;
+}
+
+static int aot_load_lit_pool(JSContext *ctx, NativeRTState *st, void *dl_handle) {
+  aot_clear_lit_pool(ctx, st);
+  st->lit_pool.dl_handle = dl_handle;
+  st->lit_pool.ctx = ctx;
   if (!dl_handle)
     return 1;
 
@@ -269,20 +299,20 @@ static int aot_load_lit_pool(JSContext *ctx, void *dl_handle) {
   if (count <= 0 || !table_ptr)
     return 1;
 
-  g_aot_lit_pool.vals = (JSGCRef *)calloc((size_t)count, sizeof(JSGCRef));
-  if (!g_aot_lit_pool.vals) {
+  st->lit_pool.vals = (JSGCRef *)calloc((size_t)count, sizeof(JSGCRef));
+  if (!st->lit_pool.vals) {
     JS_RaiseOOM(ctx);
     return 0;
   }
-  g_aot_lit_pool.count = 0;
+  st->lit_pool.count = 0;
 
   for (int i = 0; i < count; i++) {
     const char *cstr = table_ptr[i] ? table_ptr[i] : "";
-    JS_AddGCRef(ctx, &g_aot_lit_pool.vals[i]);
-    g_aot_lit_pool.count = i + 1;
-    g_aot_lit_pool.vals[i].val = js_key_new(ctx, cstr);
-    if (JS_IsException(g_aot_lit_pool.vals[i].val)) {
-      aot_clear_lit_pool();
+    JS_AddGCRef(ctx, &st->lit_pool.vals[i]);
+    st->lit_pool.count = i + 1;
+    st->lit_pool.vals[i].val = js_key_new(ctx, cstr);
+    if (JS_IsException(st->lit_pool.vals[i].val)) {
+      aot_clear_lit_pool(ctx, st);
       return 0;
     }
   }
@@ -290,70 +320,56 @@ static int aot_load_lit_pool(JSContext *ctx, void *dl_handle) {
 }
 
 static JSValue aot_lit_from_index(JSContext *ctx, int64_t lit_idx) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return JS_EXCEPTION;
   if (lit_idx < 0) {
     JS_RaiseDisrupt(ctx, "literal index out of range");
     return JS_EXCEPTION;
   }
 
-  if (g_aot_lit_pool.dl_handle != g_current_dl_handle || g_aot_lit_pool.ctx != ctx) {
-    if (!aot_load_lit_pool(ctx, g_current_dl_handle))
+  if (st->lit_pool.dl_handle != st->current_dl_handle || st->lit_pool.ctx != ctx) {
+    if (!aot_load_lit_pool(ctx, st, st->current_dl_handle))
       return JS_EXCEPTION;
   }
 
-  if (lit_idx >= g_aot_lit_pool.count) {
+  if (lit_idx >= st->lit_pool.count) {
     JS_RaiseDisrupt(ctx, "literal index out of range");
     return JS_EXCEPTION;
   }
-  return g_aot_lit_pool.vals[lit_idx].val;
+  return st->lit_pool.vals[lit_idx].val;
 }
 
-typedef struct {
-  const char *name;
-  JSValue key;
-} AOTKeyCacheEntry;
-
-static CELL_THREAD_LOCAL JSContext *g_aot_key_cache_ctx = NULL;
-static CELL_THREAD_LOCAL AOTKeyCacheEntry *g_aot_key_cache = NULL;
-static CELL_THREAD_LOCAL int g_aot_key_cache_count = 0;
-static CELL_THREAD_LOCAL int g_aot_key_cache_cap = 0;
-
 /* Convert a static C string to an interned JSValue key.
-   Uses a small thread-local cache keyed by C-string pointer to avoid
+   Uses a small per-actor cache keyed by C-string pointer to avoid
    repeated UTF-8 decoding in hot property paths. */
 static JSValue aot_key_from_cstr(JSContext *ctx, const char *name) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return JS_EXCEPTION;
   if (!name)
     return JS_NULL;
 
-  if (g_aot_key_cache_ctx != ctx) {
-    free(g_aot_key_cache);
-    g_aot_key_cache = NULL;
-    g_aot_key_cache_count = 0;
-    g_aot_key_cache_cap = 0;
-    g_aot_key_cache_ctx = ctx;
-  }
-
-  for (int i = 0; i < g_aot_key_cache_count; i++) {
-    if (g_aot_key_cache[i].name == name)
-      return g_aot_key_cache[i].key;
+  for (int i = 0; i < st->key_cache_count; i++) {
+    if (st->key_cache[i].name == name)
+      return st->key_cache[i].key;
   }
 
   JSValue key = js_key_new(ctx, name);
   if (JS_IsNull(key))
     return JS_RaiseDisrupt(ctx, "invalid property key");
 
-  if (g_aot_key_cache_count >= g_aot_key_cache_cap) {
-    int new_cap = g_aot_key_cache_cap ? (g_aot_key_cache_cap * 2) : 64;
+  if (st->key_cache_count >= st->key_cache_cap) {
+    int new_cap = st->key_cache_cap ? (st->key_cache_cap * 2) : 64;
     AOTKeyCacheEntry *new_cache =
-        (AOTKeyCacheEntry *)realloc(g_aot_key_cache, (size_t)new_cap * sizeof(*new_cache));
+        (AOTKeyCacheEntry *)realloc(st->key_cache, (size_t)new_cap * sizeof(*new_cache));
     if (!new_cache)
       return JS_RaiseOOM(ctx);
-    g_aot_key_cache = new_cache;
-    g_aot_key_cache_cap = new_cap;
+    st->key_cache = new_cache;
+    st->key_cache_cap = new_cap;
   }
 
-  g_aot_key_cache[g_aot_key_cache_count].name = name;
-  g_aot_key_cache[g_aot_key_cache_count].key = key;
-  g_aot_key_cache_count++;
+  st->key_cache[st->key_cache_count].name = name;
+  st->key_cache[st->key_cache_count].key = key;
+  st->key_cache_count++;
   return key;
 }
 
@@ -460,42 +476,44 @@ int cell_rt_store_index(JSContext *ctx, JSValue val, JSValue arr,
 
 /* --- Intrinsic/global lookup --- */
 
-/* Native module environment — set before executing a native module's cell_main.
-   Contains runtime functions (starts_with, ends_with, etc.) and use(). */
-static CELL_THREAD_LOCAL JSGCRef g_native_env_ref;
-static CELL_THREAD_LOCAL int g_has_native_env = 0;
-
 void cell_rt_set_native_env(JSContext *ctx, JSValue env) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return;
   if (!JS_IsNull(env) && !JS_IsStone(env)) {
     fprintf(stderr, "cell_rt_set_native_env: ERROR env not stone\n");
     abort();
   }
   /* Drop module literal pool roots before switching native env/module. */
-  aot_clear_lit_pool();
+  aot_clear_lit_pool(ctx, st);
 
-  /* Native module boundary: clear per-thread key cache so stale keys
+  /* Native module boundary: clear per-actor key cache so stale keys
      cannot survive across context/module lifetimes. */
-  free(g_aot_key_cache);
-  g_aot_key_cache = NULL;
-  g_aot_key_cache_count = 0;
-  g_aot_key_cache_cap = 0;
-  g_aot_key_cache_ctx = ctx;
+  free(st->key_cache);
+  st->key_cache = NULL;
+  st->key_cache_count = 0;
+  st->key_cache_cap = 0;
 
-  if (g_has_native_env)
-    JS_DeleteGCRef(ctx, &g_native_env_ref);
+  if (st->has_native_env && st->native_env_ref_inited) {
+    JS_DeleteGCRef(ctx, &st->native_env_ref);
+    st->native_env_ref_inited = 0;
+  }
   if (!JS_IsNull(env)) {
-    JS_AddGCRef(ctx, &g_native_env_ref);
-    g_native_env_ref.val = env;
-    g_has_native_env = 1;
+    JS_AddGCRef(ctx, &st->native_env_ref);
+    st->native_env_ref_inited = 1;
+    st->native_env_ref.val = env;
+    st->has_native_env = 1;
   } else {
-    g_has_native_env = 0;
+    st->has_native_env = 0;
+    st->native_env_ref.val = JS_NULL;
   }
 }
 
 static JSValue cell_rt_get_intrinsic_key(JSContext *ctx, JSValue key) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return JS_EXCEPTION;
   /* Check native env first (runtime-provided functions like log) */
-  if (g_has_native_env) {
-    JSValue v = JS_GetProperty(ctx, g_native_env_ref.val, key);
+  if (st->has_native_env) {
+    JSValue v = JS_GetProperty(ctx, st->native_env_ref.val, key);
     if (!JS_IsNull(v))
       return v;
   }
@@ -586,75 +604,51 @@ typedef struct AOTGCRefChunk {
   uint8_t inited[AOT_GC_REF_CHUNK_SIZE];
 } AOTGCRefChunk;
 
-static CELL_THREAD_LOCAL AOTGCRefChunk **g_aot_gc_ref_chunks = NULL;
-static CELL_THREAD_LOCAL int g_aot_gc_ref_chunk_count = 0;
-static CELL_THREAD_LOCAL int g_aot_depth = 0;
-static CELL_THREAD_LOCAL JSContext *g_aot_gc_ref_ctx = NULL;
-
-int cell_rt_native_active(void) {
-  return g_aot_depth > 0;
+int cell_rt_native_active(JSContext *ctx) {
+  NativeRTState *st = (NativeRTState *)ctx->native_state;
+  return st ? (st->aot_depth > 0) : 0;
 }
 
-static int ensure_aot_gc_ref_slot(JSContext *ctx, int depth_index) {
+static int ensure_aot_gc_ref_slot(JSContext *ctx, NativeRTState *st, int depth_index) {
   if (depth_index < 0)
     return 0;
   int needed_chunks = (depth_index / AOT_GC_REF_CHUNK_SIZE) + 1;
-  if (needed_chunks <= g_aot_gc_ref_chunk_count)
+  if (needed_chunks <= st->gc_ref_chunk_count)
     return 1;
   AOTGCRefChunk **new_chunks =
-      (AOTGCRefChunk **)realloc(g_aot_gc_ref_chunks,
+      (AOTGCRefChunk **)realloc(st->gc_ref_chunks,
                                 (size_t)needed_chunks * sizeof(*new_chunks));
   if (!new_chunks) {
     JS_RaiseOOM(ctx);
     return 0;
   }
-  g_aot_gc_ref_chunks = new_chunks;
-  for (int i = g_aot_gc_ref_chunk_count; i < needed_chunks; i++) {
-    g_aot_gc_ref_chunks[i] = (AOTGCRefChunk *)calloc(1, sizeof(AOTGCRefChunk));
-    if (!g_aot_gc_ref_chunks[i]) {
+  st->gc_ref_chunks = new_chunks;
+  for (int i = st->gc_ref_chunk_count; i < needed_chunks; i++) {
+    st->gc_ref_chunks[i] = (AOTGCRefChunk *)calloc(1, sizeof(AOTGCRefChunk));
+    if (!st->gc_ref_chunks[i]) {
       JS_RaiseOOM(ctx);
       return 0;
     }
   }
-  g_aot_gc_ref_chunk_count = needed_chunks;
+  st->gc_ref_chunk_count = needed_chunks;
   return 1;
 }
 
-static inline JSGCRef *aot_gc_ref_at(int depth_index) {
+static inline JSGCRef *aot_gc_ref_at(NativeRTState *st, int depth_index) {
   int chunk_index = depth_index / AOT_GC_REF_CHUNK_SIZE;
   int slot_index = depth_index % AOT_GC_REF_CHUNK_SIZE;
-  return &g_aot_gc_ref_chunks[chunk_index]->refs[slot_index];
+  return &st->gc_ref_chunks[chunk_index]->refs[slot_index];
 }
 
-static inline uint8_t *aot_gc_ref_inited_at(int depth_index) {
+static inline uint8_t *aot_gc_ref_inited_at(NativeRTState *st, int depth_index) {
   int chunk_index = depth_index / AOT_GC_REF_CHUNK_SIZE;
   int slot_index = depth_index % AOT_GC_REF_CHUNK_SIZE;
-  return &g_aot_gc_ref_chunks[chunk_index]->inited[slot_index];
+  return &st->gc_ref_chunks[chunk_index]->inited[slot_index];
 }
 
-/* GC refs are owned by a specific JSContext. If context changes on this thread,
-   unregister previous refs and reset per-slot initialization state. */
-static void aot_gc_ref_reset_ctx(JSContext *ctx) {
-  if (g_aot_gc_ref_ctx == ctx)
-    return;
-  if (g_aot_gc_ref_ctx) {
-    for (int ci = 0; ci < g_aot_gc_ref_chunk_count; ci++) {
-      AOTGCRefChunk *chunk = g_aot_gc_ref_chunks[ci];
-      for (int si = 0; si < AOT_GC_REF_CHUNK_SIZE; si++) {
-        if (chunk->inited[si]) {
-          JS_DeleteGCRef(g_aot_gc_ref_ctx, &chunk->refs[si]);
-          chunk->inited[si] = 0;
-          chunk->refs[si].val = JS_NULL;
-        }
-      }
-    }
-  }
-  g_aot_gc_ref_ctx = ctx;
-}
-
-static inline void aot_gc_ref_activate(JSContext *ctx, int depth_index) {
-  JSGCRef *ref = aot_gc_ref_at(depth_index);
-  uint8_t *inited = aot_gc_ref_inited_at(depth_index);
+static inline void aot_gc_ref_activate(JSContext *ctx, NativeRTState *st, int depth_index) {
+  JSGCRef *ref = aot_gc_ref_at(st, depth_index);
+  uint8_t *inited = aot_gc_ref_inited_at(st, depth_index);
   if (!*inited) {
     JS_AddGCRef(ctx, ref);
     *inited = 1;
@@ -662,42 +656,45 @@ static inline void aot_gc_ref_activate(JSContext *ctx, int depth_index) {
 }
 
 JSValue *cell_rt_enter_frame(JSContext *ctx, int64_t nr_slots) {
-  aot_gc_ref_reset_ctx(ctx);
-  if (!ensure_aot_gc_ref_slot(ctx, g_aot_depth)) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return NULL;
+  if (!ensure_aot_gc_ref_slot(ctx, st, st->aot_depth)) {
     return NULL;
   }
   JSFrameRegister *frame = alloc_frame_register(ctx, (int)nr_slots);
   if (!frame) return NULL;
-  aot_gc_ref_activate(ctx, g_aot_depth);
-  JSGCRef *ref = aot_gc_ref_at(g_aot_depth);
+  aot_gc_ref_activate(ctx, st, st->aot_depth);
+  JSGCRef *ref = aot_gc_ref_at(st, st->aot_depth);
   ref->val = JS_MKPTR(frame);
-  g_aot_depth++;
+  st->aot_depth++;
   return (JSValue *)frame->slots;
 }
 
 /* Push an already-allocated frame onto the active AOT frame stack. */
 static int cell_rt_push_existing_frame(JSContext *ctx, JSValue frame_val) {
-  aot_gc_ref_reset_ctx(ctx);
-  if (!ensure_aot_gc_ref_slot(ctx, g_aot_depth))
+  NativeRTState *st = native_state(ctx);
+  if (!st) return 0;
+  if (!ensure_aot_gc_ref_slot(ctx, st, st->aot_depth))
     return 0;
-  aot_gc_ref_activate(ctx, g_aot_depth);
-  JSGCRef *ref = aot_gc_ref_at(g_aot_depth);
+  aot_gc_ref_activate(ctx, st, st->aot_depth);
+  JSGCRef *ref = aot_gc_ref_at(st, st->aot_depth);
   ref->val = frame_val;
-  g_aot_depth++;
+  st->aot_depth++;
   return 1;
 }
 
 JSValue *cell_rt_refresh_fp(JSContext *ctx) {
-  (void)ctx;
-  if (g_aot_depth <= 0) {
-    fprintf(stderr, "[BUG] cell_rt_refresh_fp: g_aot_depth=%d\n", g_aot_depth);
+  NativeRTState *st = native_state(ctx);
+  if (!st) return NULL;
+  if (st->aot_depth <= 0) {
+    fprintf(stderr, "[BUG] cell_rt_refresh_fp: aot_depth=%d\n", st->aot_depth);
     abort();
   }
-  JSValue val = aot_gc_ref_at(g_aot_depth - 1)->val;
+  JSValue val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
   JSFrameRegister *frame = (JSFrameRegister *)JS_VALUE_GET_PTR(val);
   if (!frame) {
     fprintf(stderr, "[BUG] cell_rt_refresh_fp: frame is NULL at depth=%d val=%lld\n",
-            g_aot_depth, (long long)val);
+            st->aot_depth, (long long)val);
     abort();
   }
   return (JSValue *)frame->slots;
@@ -705,13 +702,15 @@ JSValue *cell_rt_refresh_fp(JSContext *ctx) {
 
 /* Combined refresh + exception check in a single call. */
 JSValue *cell_rt_refresh_fp_checked(JSContext *ctx) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return NULL;
   if (JS_HasException(ctx))
     return NULL;
-  if (g_aot_depth <= 0) {
-    fprintf(stderr, "[BUG] cell_rt_refresh_fp_checked: g_aot_depth=%d\n", g_aot_depth);
+  if (st->aot_depth <= 0) {
+    fprintf(stderr, "[BUG] cell_rt_refresh_fp_checked: aot_depth=%d\n", st->aot_depth);
     abort();
   }
-  JSValue val = aot_gc_ref_at(g_aot_depth - 1)->val;
+  JSValue val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
   JSFrameRegister *frame = (JSFrameRegister *)JS_VALUE_GET_PTR(val);
   if (!frame) {
     fprintf(stderr, "[BUG] cell_rt_refresh_fp_checked: frame is NULL\n");
@@ -721,34 +720,26 @@ JSValue *cell_rt_refresh_fp_checked(JSContext *ctx) {
 }
 
 void cell_rt_leave_frame(JSContext *ctx) {
-  (void)ctx;
-  if (g_aot_depth <= 0) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return;
+  if (st->aot_depth <= 0) {
     fprintf(stderr, "[BUG] cell_rt_leave_frame underflow\n");
     abort();
   }
-  g_aot_depth--;
-  aot_gc_ref_at(g_aot_depth)->val = JS_NULL;
+  st->aot_depth--;
+  aot_gc_ref_at(st, st->aot_depth)->val = JS_NULL;
 }
 
 /* --- Function creation and calling --- */
 
 typedef JSValue (*cell_compiled_fn)(JSContext *ctx, void *fp);
 
-/* Set before executing a native module's cell_main —
-   used by cell_rt_make_function to resolve fn_ptr via dlsym */
-/* g_current_dl_handle is defined near property/literal helpers. */
-
 /* ============================================================
    Dispatch loop — the core of native function execution.
    Each compiled cell_fn_N returns to this loop when it needs
    to call another function (instead of recursing via C stack).
    ============================================================ */
 
-/* Pending call state — set by cell_rt_signal_call / cell_rt_signal_tail_call,
-   read by the dispatch loop. */
-static CELL_THREAD_LOCAL JSValue g_pending_callee_frame = 0;  /* JSFrameRegister ptr */
-static CELL_THREAD_LOCAL int g_pending_is_tail = 0;
-
 /* Poll pause state on taken backward jumps (AOT backedges).
    MACH can suspend/resume a register VM frame at pc granularity; native AOT
    does not currently have an equivalent resume point, so we acknowledge timer
@@ -762,34 +753,38 @@ int cell_rt_check_backedge(JSContext *ctx) {
 }
 
 void cell_rt_signal_call(JSContext *ctx, void *fp, int64_t frame_slot) {
-  (void)ctx;
+  NativeRTState *st = native_state(ctx);
+  if (!st) return;
   JSValue *slots = (JSValue *)fp;
-  g_pending_callee_frame = slots[frame_slot];
-  g_pending_is_tail = 0;
+  st->pending_callee_frame = slots[frame_slot];
+  st->pending_is_tail = 0;
 }
 
 void cell_rt_signal_tail_call(JSContext *ctx, void *fp, int64_t frame_slot) {
-  (void)ctx;
+  NativeRTState *st = native_state(ctx);
+  if (!st) return;
   JSValue *slots = (JSValue *)fp;
-  g_pending_callee_frame = slots[frame_slot];
-  g_pending_is_tail = 1;
+  st->pending_callee_frame = slots[frame_slot];
+  st->pending_is_tail = 1;
 }
 
 /* Entry point called from JS_CallInternal / JS_Call / MACH_INVOKE
    for JS_FUNC_KIND_NATIVE functions. */
 JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
                              JSValue this_obj, int argc, JSValue *argv) {
+  NativeRTState *st = native_state(ctx);
+  if (!st) return JS_EXCEPTION;
   JSFunction *f = JS_VALUE_GET_FUNCTION(func_obj);
   cell_compiled_fn fn = (cell_compiled_fn)JS_VALUE_GET_CODE(f->u.cell.code)->u.native.fn_ptr;
   int nr_slots = JS_VALUE_GET_CODE(f->u.cell.code)->u.native.nr_slots;
   int arity = f->length;
-  void *prev_dl_handle = g_current_dl_handle;
-  g_current_dl_handle = JS_VALUE_GET_CODE(f->u.cell.code)->u.native.dl_handle;
+  void *prev_dl_handle = st->current_dl_handle;
+  st->current_dl_handle = JS_VALUE_GET_CODE(f->u.cell.code)->u.native.dl_handle;
 
 #define RETURN_DISPATCH(v)                                                     \
   do {                                                                         \
     atomic_store_explicit(&ctx->pause_flag, 0, memory_order_relaxed);         \
-    g_current_dl_handle = prev_dl_handle;                                      \
+    st->current_dl_handle = prev_dl_handle;                                    \
     return (v);                                                                \
   } while (0)
 
@@ -820,11 +815,11 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
   JSFrameRegister *frame = (JSFrameRegister *)((char *)fp - offsetof(JSFrameRegister, slots));
   frame->function = func_obj;
 
-  int base_depth = g_aot_depth; /* remember entry depth for return detection */
+  int base_depth = st->aot_depth; /* remember entry depth for return detection */
 
   for (;;) {
-    g_pending_callee_frame = 0;
-    g_pending_is_tail = 0;
+    st->pending_callee_frame = 0;
+    st->pending_is_tail = 0;
     if (atomic_load_explicit(&ctx->pause_flag, memory_order_relaxed) >= 1)
       atomic_store_explicit(&ctx->pause_flag, 0, memory_order_relaxed);
 
@@ -832,26 +827,26 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
     if (JS_IsFunction(frame->function)) {
       JSFunction *cur_fn = JS_VALUE_GET_FUNCTION(frame->function);
       if (cur_fn->kind == JS_FUNC_KIND_NATIVE)
-        g_current_dl_handle = JS_VALUE_GET_CODE(cur_fn->u.cell.code)->u.native.dl_handle;
+        st->current_dl_handle = JS_VALUE_GET_CODE(cur_fn->u.cell.code)->u.native.dl_handle;
     }
 
     JSValue result = fn(ctx, fp);
 
     /* Re-derive frame after potential GC */
-    if (g_aot_depth <= 0) {
+    if (st->aot_depth <= 0) {
       fprintf(stderr, "[BUG] native dispatch lost frame depth after fn call\n");
       abort();
     }
-    JSValue frame_val = aot_gc_ref_at(g_aot_depth - 1)->val;
+    JSValue frame_val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
     frame = (JSFrameRegister *)JS_VALUE_GET_PTR(frame_val);
     fp = (JSValue *)frame->slots;
 
-    if (g_pending_callee_frame != 0) {
+    if (st->pending_callee_frame != 0) {
       /* Function signaled a call — dispatch it */
-      JSValue callee_frame_val = g_pending_callee_frame;
-      g_pending_callee_frame = 0;
-      int pending_is_tail = g_pending_is_tail;
-      g_pending_is_tail = 0;
+      JSValue callee_frame_val = st->pending_callee_frame;
+      st->pending_callee_frame = 0;
+      int pending_is_tail = st->pending_is_tail;
+      st->pending_is_tail = 0;
       JSGCRef callee_ref;
       JS_PushGCRef(ctx, &callee_ref);
       callee_ref.val = callee_frame_val;
@@ -897,7 +892,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
             JS_PopGCRef(ctx, &callee_ref);
             RETURN_DISPATCH(JS_EXCEPTION);
           }
-          frame = (JSFrameRegister *)JS_VALUE_GET_PTR(aot_gc_ref_at(g_aot_depth - 1)->val);
+          frame = (JSFrameRegister *)JS_VALUE_GET_PTR(aot_gc_ref_at(st, st->aot_depth - 1)->val);
           fp = (JSValue *)frame->slots;
           fn = callee_ptr;
         } else {
@@ -917,7 +912,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
 
           if (!cell_rt_push_existing_frame(ctx, callee_ref.val)) {
             /* Resume caller with exception pending */
-            frame_val = aot_gc_ref_at(g_aot_depth - 1)->val;
+            frame_val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
             frame = (JSFrameRegister *)JS_VALUE_GET_PTR(frame_val);
             fp = (JSValue *)frame->slots;
             JSFunction *exc_fn = JS_VALUE_GET_FUNCTION(frame->function);
@@ -926,7 +921,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
             JS_PopGCRef(ctx, &callee_ref);
             continue;
           }
-          frame = (JSFrameRegister *)JS_VALUE_GET_PTR(aot_gc_ref_at(g_aot_depth - 1)->val);
+          frame = (JSFrameRegister *)JS_VALUE_GET_PTR(aot_gc_ref_at(st, st->aot_depth - 1)->val);
           fp = (JSValue *)frame->slots;
           fn = callee_ptr;
         }
@@ -942,7 +937,7 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
                                 callee_argc, &callee_fr->slots[1], 0);
 
         /* Re-derive frame after call */
-        frame_val = aot_gc_ref_at(g_aot_depth - 1)->val;
+        frame_val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
         frame = (JSFrameRegister *)JS_VALUE_GET_PTR(frame_val);
         fp = (JSValue *)frame->slots;
 
@@ -970,18 +965,18 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
         if (pending_is_tail) {
           /* Tail call to non-native: return its result up the chain */
           /* Pop current frame and return to caller */
-          if (g_aot_depth <= base_depth) {
+          if (st->aot_depth <= base_depth) {
             cell_rt_leave_frame(ctx);
             JS_PopGCRef(ctx, &callee_ref);
             RETURN_DISPATCH(ret);
           }
           /* Pop current frame, return to caller frame */
           cell_rt_leave_frame(ctx);
-          if (g_aot_depth < base_depth) {
+          if (st->aot_depth < base_depth) {
             JS_PopGCRef(ctx, &callee_ref);
             RETURN_DISPATCH(ret);
           }
-          frame_val = aot_gc_ref_at(g_aot_depth - 1)->val;
+          frame_val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
           frame = (JSFrameRegister *)JS_VALUE_GET_PTR(frame_val);
           fp = (JSValue *)frame->slots;
           int ret_info = JS_VALUE_GET_INT(frame->address);
@@ -1014,17 +1009,17 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
       if (!JS_HasException(ctx))
         JS_Disrupt(ctx);
 
-      if (g_aot_depth <= base_depth) {
+      if (st->aot_depth <= base_depth) {
         cell_rt_leave_frame(ctx);
         RETURN_DISPATCH(JS_EXCEPTION);
       }
       cell_rt_leave_frame(ctx);
-      if (g_aot_depth < base_depth) {
+      if (st->aot_depth < base_depth) {
         RETURN_DISPATCH(JS_EXCEPTION);
       }
 
       /* Resume caller and tag the return slot with JS_EXCEPTION. */
-      frame_val = aot_gc_ref_at(g_aot_depth - 1)->val;
+      frame_val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
       frame = (JSFrameRegister *)JS_VALUE_GET_PTR(frame_val);
       fp = (JSValue *)frame->slots;
       int ret_info = JS_VALUE_GET_INT(frame->address);
@@ -1038,17 +1033,17 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
     }
 
     /* Normal return — pop frame and store result in caller */
-    if (g_aot_depth <= base_depth) {
+    if (st->aot_depth <= base_depth) {
       cell_rt_leave_frame(ctx);
       RETURN_DISPATCH(result);
     }
     cell_rt_leave_frame(ctx);
-    if (g_aot_depth < base_depth) {
+    if (st->aot_depth < base_depth) {
       RETURN_DISPATCH(result);
     }
 
     /* Return to caller frame */
-    frame_val = aot_gc_ref_at(g_aot_depth - 1)->val;
+    frame_val = aot_gc_ref_at(st, st->aot_depth - 1)->val;
     frame = (JSFrameRegister *)JS_VALUE_GET_PTR(frame_val);
     fp = (JSValue *)frame->slots;
     int ret_info = JS_VALUE_GET_INT(frame->address);
@@ -1064,27 +1059,71 @@ JSValue cell_native_dispatch(JSContext *ctx, JSValue func_obj,
 #undef RETURN_DISPATCH
 }
 
+static JSValue aot_get_or_create_native_code(JSContext *ctx, NativeRTState *st,
+                                             void *dl_handle, int64_t fn_idx,
+                                             int arity, uint16_t nr_slots) {
+  for (int i = 0; i < st->code_cache_count; i++) {
+    AOTCodeCacheEntry *e = &st->code_cache[i];
+    if (e->dl_handle == dl_handle && e->fn_idx == fn_idx)
+      return e->code;
+  }
+
+  char name[64];
+  snprintf(name, sizeof(name), "cell_fn_%lld", (long long)fn_idx);
+  void *fn_ptr = dlsym(dl_handle, name);
+  if (!fn_ptr)
+    return JS_RaiseDisrupt(ctx, "native function %s not found in dylib", name);
+
+  JSCode *code = ct_alloc(ctx, sizeof(JSCode), 8);
+  if (!code)
+    return JS_EXCEPTION;
+  memset(code, 0, sizeof(*code));
+  code->header = objhdr_make(0, OBJ_CODE, 0, 0, 0, 0);
+  code->kind = JS_CODE_KIND_NATIVE;
+  code->arity = (int16_t)arity;
+  code->u.native.fn_ptr = fn_ptr;
+  code->u.native.dl_handle = dl_handle;
+  code->u.native.nr_slots = nr_slots;
+  JSValue code_obj = JS_MKPTR(code);
+
+  if (st->code_cache_count >= st->code_cache_cap) {
+    int new_cap = st->code_cache_cap ? (st->code_cache_cap * 2) : 128;
+    AOTCodeCacheEntry *new_cache =
+        (AOTCodeCacheEntry *)realloc(st->code_cache, (size_t)new_cap * sizeof(*new_cache));
+    if (!new_cache)
+      return JS_RaiseOOM(ctx);
+    st->code_cache = new_cache;
+    st->code_cache_cap = new_cap;
+  }
+
+  st->code_cache[st->code_cache_count].dl_handle = dl_handle;
+  st->code_cache[st->code_cache_count].fn_idx = fn_idx;
+  st->code_cache[st->code_cache_count].code = code_obj;
+  st->code_cache_count++;
+  return code_obj;
+}
+
 /* Create a native function object from a compiled fn_idx.
    Called from QBE-generated code during function creation. */
 JSValue cell_rt_make_function(JSContext *ctx, int64_t fn_idx, void *outer_fp,
                               int64_t nr_args, int64_t nr_slots) {
-  if (!g_current_dl_handle)
+  (void)outer_fp;
+  NativeRTState *st = native_state(ctx);
+  if (!st) return JS_EXCEPTION;
+  if (!st->current_dl_handle)
     return JS_RaiseDisrupt(ctx, "no native module loaded");
 
-  /* Resolve fn_ptr via dlsym at creation time — cached in the function object */
-  char name[64];
-  snprintf(name, sizeof(name), "cell_fn_%lld", (long long)fn_idx);
-  void *fn_ptr = dlsym(g_current_dl_handle, name);
-  if (!fn_ptr)
-    return JS_RaiseDisrupt(ctx, "native function %s not found in dylib", name);
+  JSValue code_obj = aot_get_or_create_native_code(
+      ctx, st, st->current_dl_handle, fn_idx, (int)nr_args, (uint16_t)nr_slots);
+  if (JS_IsException(code_obj))
+    return JS_EXCEPTION;
 
   /* Get the current frame as outer_frame for closures */
   JSValue outer_frame = JS_NULL;
-  if (g_aot_depth > 0)
-    outer_frame = aot_gc_ref_at(g_aot_depth - 1)->val;
+  if (st->aot_depth > 0)
+    outer_frame = aot_gc_ref_at(st, st->aot_depth - 1)->val;
 
-  return js_new_native_function(ctx, fn_ptr, g_current_dl_handle,
-                                (uint16_t)nr_slots, (int)nr_args, outer_frame);
+  return js_new_native_function_with_code(ctx, code_obj, (int)nr_args, outer_frame);
 }
 
 /* --- Frame-based function calling ---
@@ -1338,22 +1377,24 @@ JSValue cell_rt_regexp(JSContext *ctx, const char *pattern, const char *flags) {
 
 /* --- Module entry point ---
    Loads a native .cm module from a dylib handle.
-   Looks up cell_main, builds a heap-allocated frame, sets
-   g_current_dl_handle so closures register in the right module. */
+   Looks up cell_main, builds a heap-allocated frame, and
+   records active module handle in per-actor native state. */
 
 /* Helper: run a native module's entry point through the dispatch loop.
    Creates a temporary JS_FUNC_KIND_NATIVE function so that the full
    dispatch loop (tail calls, closures, etc.) works for module-level code. */
 static JSValue native_module_run(JSContext *ctx, void *dl_handle,
                                  cell_compiled_fn entry, int nr_slots) {
-  void *prev_handle = g_current_dl_handle;
-  g_current_dl_handle = dl_handle;
+  NativeRTState *st = native_state(ctx);
+  if (!st) return JS_EXCEPTION;
+  void *prev_handle = st->current_dl_handle;
+  st->current_dl_handle = dl_handle;
 
   /* Create a native function object for the entry point */
   JSValue func_obj = js_new_native_function(ctx, (void *)entry, dl_handle,
                                             (uint16_t)nr_slots, 0, JS_NULL);
   if (JS_IsException(func_obj)) {
-    g_current_dl_handle = prev_handle;
+    st->current_dl_handle = prev_handle;
     return JS_EXCEPTION;
   }
 
@@ -1362,7 +1403,7 @@ static JSValue native_module_run(JSContext *ctx, void *dl_handle,
     JS_GetException(ctx);
 
   JSValue result = cell_native_dispatch(ctx, func_obj, JS_NULL, 0, NULL);
-  g_current_dl_handle = prev_handle;
+  st->current_dl_handle = prev_handle;
   return result;
 }
 
@@ -1409,6 +1450,38 @@ JSValue cell_rt_native_module_load_named(JSContext *ctx, void *dl_handle, const
   return native_module_run(ctx, dl_handle, fn, nr_slots);
 }
 
+void cell_rt_free_native_state(JSContext *ctx) {
+  NativeRTState *st = (NativeRTState *)ctx->native_state;
+  if (!st) return;
+
+  aot_clear_lit_pool(ctx, st);
+
+  if (st->has_native_env && st->native_env_ref_inited) {
+    JS_DeleteGCRef(ctx, &st->native_env_ref);
+    st->native_env_ref_inited = 0;
+    st->native_env_ref.val = JS_NULL;
+  }
+
+  for (int ci = 0; ci < st->gc_ref_chunk_count; ci++) {
+    AOTGCRefChunk *chunk = st->gc_ref_chunks[ci];
+    if (!chunk) continue;
+    for (int si = 0; si < AOT_GC_REF_CHUNK_SIZE; si++) {
+      if (chunk->inited[si]) {
+        JS_DeleteGCRef(ctx, &chunk->refs[si]);
+        chunk->inited[si] = 0;
+        chunk->refs[si].val = JS_NULL;
+      }
+    }
+    free(chunk);
+  }
+
+  free(st->gc_ref_chunks);
+  free(st->key_cache);
+  free(st->code_cache);
+  js_free_rt(st);
+  ctx->native_state = NULL;
+}
+
 /* Backward-compat: uses RTLD_DEFAULT (works when dylib opened with RTLD_GLOBAL) */
 JSValue cell_rt_module_entry(JSContext *ctx) {
   void *handle = dlopen(NULL, RTLD_LAZY);
diff --git a/source/quickjs-internal.h b/source/quickjs-internal.h
index 73dcdad7..db5546d6 100644
--- a/source/quickjs-internal.h
+++ b/source/quickjs-internal.h
@@ -1121,6 +1121,7 @@ struct JSContext {
   JSGCRef *last_gc_ref; /* used to reference temporary GC roots (list) */
   JSLocalRef *top_local_ref; /* for JS_LOCAL macro - GC updates C locals through pointers */
   CCallRoot *c_call_root; /* stack of auto-rooted C call argv arrays */
+  void *native_state; /* qbe_helpers.c per-actor native runtime state */
 
   int class_count; /* size of class_array and class_proto */
   JSClass *class_array;
@@ -1562,7 +1563,7 @@ static inline void set_value (JSContext *ctx, JSValue *pval, JSValue new_val) {
   *pval = new_val;
 }
 
-int cell_rt_native_active(void);
+int cell_rt_native_active(JSContext *ctx);
 
 static inline __exception int js_poll_interrupts (JSContext *ctx) {
   if (unlikely (atomic_load_explicit (&ctx->pause_flag, memory_order_relaxed) >= 2)) {
@@ -1659,7 +1660,9 @@ JSValue js_key_from_string (JSContext *ctx, JSValue val);
 /* mach.c exports */
 JSValue JS_CallRegisterVM(JSContext *ctx, JSCodeRegister *code, JSValue this_obj, int argc, JSValue *argv, JSValue env, JSValue outer_frame);
 JSValue js_new_native_function(JSContext *ctx, void *fn_ptr, void *dl_handle, uint16_t nr_slots, int arity, JSValue outer_frame);
+JSValue js_new_native_function_with_code(JSContext *ctx, JSValue code_obj, int arity, JSValue outer_frame);
 JSFrameRegister *alloc_frame_register(JSContext *ctx, int slot_count);
+void cell_rt_free_native_state(JSContext *ctx);
 
 
 #endif /* QUICKJS_INTERNAL_H */
diff --git a/source/runtime.c b/source/runtime.c
index 8bc865a4..9bc46d5d 100644
--- a/source/runtime.c
+++ b/source/runtime.c
@@ -2102,6 +2102,7 @@ void JS_FreeContext (JSContext *ctx) {
   JSRuntime *rt = ctx->rt;
   int i;
 
+  cell_rt_free_native_state(ctx);
   JS_DeleteGCRef(ctx, &ctx->suspended_frame_ref);
 
   for (i = 0; i < ctx->class_count; i++) {
@@ -5332,7 +5333,7 @@ JSValue js_regexp_toString (JSContext *ctx, JSValue this_val, int argc, JSValue
 
 int lre_check_timeout (void *opaque) {
   JSContext *ctx = opaque;
-  if (cell_rt_native_active ()) {
+  if (cell_rt_native_active (ctx)) {
     atomic_store_explicit (&ctx->pause_flag, 0, memory_order_relaxed);
     return 0;
   }