ci: run the aarch64 suite with one qemu insn per tb, tcg optimization falsely trips the ubsan pointer check

.gitea/workflows/gates.yml

@@ -101,8 +101,15 @@ jobs:
           sudo apt-get update -qq
           sudo apt-get install -qq -y --no-install-recommends qemu-user
 
+      # QEMU_ONE_INSN_PER_TB: qemu's TCG with translation-block optimization
+      # (observed on 8.2 and 11.1) takes the panic branch of Zig's UBSan
+      # pointer-overflow check on a valid in-bounds pointer (subs+b.cc is
+      # correct AArch64; real silicon and one-insn-per-tb both pass). The
+      # false abort makes qemu write a ~380 MB guest core per crashing test
+      # and the respawn loop ran the job into the runner's 3 h kill. ~10x
+      # slower, correct by construction.
       - name: Run test suite under qemu (plain suite, no -Dintegration)
-        run: zig build test-aarch64 -fqemu
+        run: QEMU_ONE_INSN_PER_TB=1 zig build test-aarch64 -fqemu
 
   frontend:
     # The light runner, a second act_runner at capacity 1 that advertises only