Fixes

runtime/src/libc.c

@@ -572,17 +572,15 @@ void *malloc(size_t n0) {
         link = &cur->next;
         cur  = cur->next;
     }
-    // Bump-allocate from the high end.
+    // Bump-allocate from the high end. Big allocations (e.g. the 16 KB sprite
+    // codegen scratch) are routed through halBigAlloc -> the Memory Manager,
+    // NOT this heap, so n is always small here and the historical
+    // `p + HDR_SZ + n > heapEnd` over-heap miscompile (only manifested for
+    // oversized n) is never exercised. A `heapEnd - p` reformulation went
+    // through the same i32 path and spuriously failed small mallocs, so keep
+    // the original simple compare.
     char *p = bumpPtr;
-    // Bounds check as a 16-bit SAME-BANK difference (avail = heapEnd - p),
+    if (p + HDR_SZ + n > heapEnd) return (void *)0;
-    // NOT `p + HDR_SZ + n > heapEnd`: that 24-bit pointer compare goes through
-    // the same i32 pattern the backend miscompiles at -O2 (see the size-round
-    // note above), so an oversized request returned a non-NULL OVER-HEAP
-    // pointer instead of NULL. On the tiny IIgs heap that let, e.g., a 16 KB
-    // scratch malloc "succeed" on a ~3 KB heap; the returned pointer + free()
-    // then trampled the IO window / language card / GS-OS stack and crashed.
-    // The heap lives in one bank, so the 16-bit offset difference is exact.
-    if ((u16)(heapEnd - p) < (u16)((u16)HDR_SZ + n)) return (void *)0;
     *(u16 *)p = (u16)n;
     bumpPtr = p + HDR_SZ + n;
     return p + HDR_SZ;