Task 08: bound TensorCircuit sampling batches

[QingyunQian]

Take-home insight

Keep the public expert’s TensorCircuit circuit and all 8192 perfect_sampling trajectories, but execute the mapped sampler in contiguous 256-shot blocks. Bounding the mapped shot axis reduces peak live tensor-network intermediates without changing the sampling algorithm.

Complete canonical result

Five alternating pairs ran the complete 8192-shot workload on one sufficient-memory node with identical six-CPU affinity:

Metric	Public expert	256-shot candidate
Passing runs	5/5	5/5
Mean runtime	126.675 s	123.188 s
Pair wins	2/5	3/5
Mean pairwise speedup	—	1.045x
95% Student-t interval	—	[0.818, 1.273]

The interval crosses 1.0, so this PR does not claim a confirmed runtime speedup. The retained contribution is bounded peak memory and reproducible execution. Among feasible full-workload screens, 256-shot blocks were best: 512/256/128-shot blocks took 50.843/44.028/55.182 s.

Reviewer-facing files

• optimized_solutions/challenge-08/README.md
• optimized_solutions/challenge-08/solution_8_bounded_batches.py
• optimized_solutions/challenge-08/factor-ablation.svg

The Python source is byte-identical to the merged Benchmark candidate, including the original expert docstring. Full campaign records, oracle validation, allocation study and raw profiles remain in hmyuuu/OrbitBreakersExpertBenchmarks PR #10.