Optimize a Python-simulated TPU kernel to beat 1487 cycles (Anthropic performance take-home)

1. Read the simulator README + Machine.py and the example kernels in tests/ before touching code.
2. Write your own cycle-accurate trace visualizer — Anthropic intentionally ships none.
3. Optimize one axis at a time: first dependency analysis to pack independent slots into the same VLIW bundle, then loop unrolling and SIMD-style "load 4 elements per cycle" patterns, finally allocating multicore carefully (when re-enabled).
4. Regression-check via python tests/submission_tests.py against fixed thresholds.
5. Never touch tests/ — Anthropic explicitly invalidates solutions that modify the test folder.

The hardest-signal move: notice the competing metric — instruction count vs wall-clock cycles — and pick the one whose bottleneck is binding, rather than chasing the wrong one.