diff --git a/asm/lumbda.s b/asm/lumbda.s
index c0e68ca..4080f4e 100644
--- a/asm/lumbda.s
+++ b/asm/lumbda.s
@@ -1044,6 +1044,20 @@ heap_alloc:
     movq %r15, gc_free_list(%rip)
     addq %rax, %r15            # %r15 now == %r13
 .ha_grow_no_pad:
+    # Pin the abandoned chunk's stored end to %r15 (its high-water).
+    # In the padded path above this is a no-op (r15 already == r13).
+    # In the un-padded path (tail < 16) r15 sits short of r13 and the
+    # gap holds mmap-zeros. Without this update, gc_sweep would still
+    # walk to the original mmap end, decode the zero pages as fake
+    # dead blocks, and on the final iteration write the free-list
+    # next-pointer to (chunk_end - 8) + 8 == chunk_end — the first
+    # byte of an unmapped page — and segfault in .gsw_dead. Tying
+    # chunk_end to r15 stops the walker at the high-water mark.
+    # Discovered while diagnosing ecdsa task #45, 2026-06-05.
+    movq gc_chunk_count(%rip), %rdx
+    decq %rdx
+    leaq gc_chunk_end(%rip), %rcx
+    movq %r15, (%rcx,%rdx,8)
     # Chunk size: at least HEAP_SIZE, but large enough for this one
     # block. A single big allocation (e.g. file->string on a ~3 MB
     # PDF) would otherwise loop forever because each mmap'd 1 MB