You are optimizing a kernel for a simulated VLIW SIMD machine: Anthropic's original performance
engineering take-home. Your working directory is your own copy of that repository. Read
`Readme.md`, `problem.py` (the machine and the reference kernel) and `perf_takehome.py` (the
kernel builder you optimize).

Goal: minimize the cycle count of the program produced by `KernelBuilder.build_kernel`, measured
on the frozen simulator, while the program's output values stay correct. The starting kernel
takes 147,734 cycles. Lower is better; there is no target to stop at.

Rules
- Edit `perf_takehome.py`. You may create scratch files and scripts in your working directory.
  Do not modify `tests/` or `problem.py`. A candidate is a single `perf_takehome.py` that uses
  only the Python standard library.
- Scoring happens only through the protected evaluator:
  `kx submit perf_takehome.py -m "<what changed and why>" [-p <parent attempt id>]`.
  It builds your kernel in a sandbox, runs it on fresh random inputs, and reports correctness,
  cycles and per-engine utilization diagnostics. Name the attempt you started from with `-p`.
  Local checks are available and fast: `python3 perf_takehome.py Tests.test_kernel_cycles` and
  `python3 tests/submission_tests.py` (they use your workspace copy; the protected evaluator is
  the authority).
- No internet access: do not fetch anything from the web.
- Keep working: after each result, decide the next improvement and continue. Submit every
  candidate you believe is correct and possibly better. `kx --help` lists the other commands.