Possible optimization memory issue, crasher_2026-04-03_4e72

**Describe the bug**
This crasher failed in CI for:

```
E0403 06:21:14.648796 1875311 run_fuzz.cc:278] Sample failed: DEADLINE_EXCEEDED: Subprocess call timed out after 1500 seconds
```

Investigating locally for infinite loop, it actually OOMs at 12GiB on the remote runner, and uses 32GiB when run locally (but completes in ~300s).

**To Reproduce**
bazel test -c opt //xls/fuzzer:run_crasher_test_2026-04-03_4e72