[bachase-agent]

Summary

The private CPU replay binding accepted numeric strings such as "0" because nanobind converted entries to uint8_t before Clifft validated them. This reproduces with the locked nanobind 2.11.0 on CPython 3.14.2; numeric bytes and numpy.float32(0) also passed through.

Keep entries as Python objects until applying Python's integer-index protocol and checking that each value is 0 or 1. Python booleans and Python/NumPy integers remain supported, with the existing container and record-length validation preserved. Add regression cases for string, bytes, and floating-point inputs, plus valid boolean/integer branches.

Closes #611. Discovered during #610; this fix changes only the private CPU replay binding and its tests.

Test plan

• Before the fix, the expanded reference tests reproduced five failures with nanobind 2.11.0, including the original string case.
• Built a cp312-abi3 wheel with nanobind 2.11.0. All 47 GPU-reference tests pass on CPython 3.12.14, 3.13.11, and 3.14.2.
• On CPython 3.14.2, GPU-reference, experimental-GPU, GPU-sampling-mode, and record-probability tests: 104 passed, 68 skipped because GPU backends/devices are unavailable.
• uv run --frozen --only-group dev pre-commit run --all-files --show-diff-on-failure passes.

Contributor agreement

• I confirm this contribution is original work and license it under the project's Apache-2.0 license.
• Human review of the AI-assisted changes is pending. The commit includes the Codex Assisted-by: trailer.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-24 16:27 UTC

[github-actions]

🐤 Performance Canary

No confirmed regressions detected. 1 benchmark comparison was inconclusive. This does not block merging.

Benchmark	Base	PR	Runtime change	Assessment
Squeeze 8192 T gates	24.7 us	24.7 us	+0.3%	No material change
Phase pass, 4,096 measurement barriers	3.60 ms	3.58 ms	-0.6%	No material change
Phase pass, 4,096 Clifford regions	5.62 ms	5.59 ms	-0.4%	No material change
Phase pass, 32,000 rank-limited rotations	18.1 ms	18.3 ms	+1.0%	No material change
Phase pass, 4,096 postselected lookahead regions	3.50 ms	3.55 ms	+1.6%	No material change
Compile/plan cultivation d5	13.3 ms	13.3 ms	+0.3%	No material change
Compile+schedule+plan, coherent QEC d5/r5	11.6 ms	11.6 ms	-0.4%	No material change
Schedule pass, coherent QEC d5/r5	8.93 ms	8.91 ms	-0.2%	No material change
Schedule pass, cultivation d5	3.67 ms	3.73 ms	+1.6%	No material change
Schedule dependence build, surface code d7/r7	1.39 ms	1.39 ms	-0.1%	No material change
QV-10, 100 shots	4.98 ms	4.96 ms	-0.4%	No material change
Cultivation d5, 1,000 shots	16.1 ms	15.9 ms	-1.4%	No material change
Early rejection with noisy suffix, 1,000 shots	929 us	932 us	+0.3%	No material change
Coherent QEC d5/r5, width 13, 100 shots	69.2 ms	68.9 ms	-0.4%	No material change
Surface code d7/r7, 10,000 shots	12.1 ms	12.0 ms	-1.0%	No material change
Surface code d5/r5 high noise, 10,000 shots	25.0 ms	24.9 ms	-0.5%	No material change
Surface code d11/r11, 1,000 shots	4.06 ms	4.05 ms	-0.4%	No material change
EXP_VAL 20q/200 probes, 100,000 shots	50.7 ms	47.7 ms	-6.0%	Inconclusive
external_inputs_nonclifford_packed_256_shots	123 us	123 us	+0.1%	No material change
external_inputs_nonclifford_scalar_256_shots	403 us	404 us	+0.2%	No material change
external_inputs_surface_d7_packed_256_shots	194 us	192 us	-0.6%	No material change
external_inputs_surface_d7_scalar_256_shots	7.52 ms	7.54 ms	+0.1%	No material change

_{Compared main (e16fa40) with this PR (3c39ce4) on the same runner using workload-local A/B/B/A ordering. Positive changes are slower. Changes under 5% are reported as no material change; changes of at least 5% but under 10% are notable; changes of at least 10% are possible regressions or improvements. Material changes that do not repeat in both pairs with low noise are inconclusive.}

Environment and method

• Runner CPU: AMD EPYC 9V74 80-Core Processor
• Pinned logical CPU: 0
• ISA: AVX2 runtime backend (forced)
• Compiler: clang version 22.1.8
• Build: Release, x86-64-v2, ThinLTO, lld, and OpenMP enabled.
• Each pass warms every workload for at least 0.2 seconds, then runs 3 Google Benchmark repetitions with at least 0.5 seconds of measured CPU time each.
• One unreported full-suite process per revision runs before the measured pairs.
• Displayed timings are the geometric mean of the two drift-balanced pass medians.
• Material changes must repeat in both pairs with under 5% within-pass relative standard deviation; otherwise they are inconclusive.
• Results apply only to this runner and are intended to detect large regressions, not establish release performance.

View workflow run