Add minimal factoring Cube-and-Conquer workflow

[isPANN]

Summary

• Reduce the benchmark surface to the factoring workflow: deterministic SAT/UNSAT instance generation in CircuitSAT JSON and DIMACS CNF.
• Add direct Kissat solving and march-generated cube conquering from Python.
• Add Rust cube export and streaming conquer modes with bounded parallel Kissat workers.
• Terminate in-flight Kissat processes after an early SAT result and stream CNFs through stdin to avoid per-cube disk I/O.
• Remove the superseded hard-regime calibration, matrix, HPC, verification, and reporting pipeline.

Why

The previous benchmark stack mixed several historical experiment paths. This change keeps the paper-facing factoring workflow small and makes the Rust solver responsible for structural cubing and optional online conquer, while Python remains focused on reproducible experiment orchestration and analysis.

User impact

The retained workflow can generate matching circuit/CNF factoring corpora, solve CNF directly with Kissat, generate and conquer march cubes, export all cubes from the structural Rust solver, or stream those cubes directly into a parallel Kissat pool. SAT streaming runs stop and terminate outstanding solver processes as soon as a model is found.

Validation

• cargo test --all-targets — 109 passed.
• PYTHONPATH=. uv run pytest — 30 passed.
• Linux 26-bit UNSAT streaming benchmark, 72 Kissat workers, CnC threshold 1024 — UNSAT in 21.69 s; 248/248 cubes closed with zero errors.

Closes #51

[isPANN]

HPC2 debug smoke completed successfully.

• Slurm job: 10006741
• State / exit: COMPLETED (0:0)
• Elapsed: 00:03:14
• Allocation: debug, 1 task, 4 CPUs, 16 GiB; batch MaxRSS 1,314,732 KiB
• Checks: frozen contract; 24 targeted Python tests; 3 Rust trace tests; release build; 39 deterministic prime targets; independent primality, width, and calibration/held-out split validation

No pilot or full-table job has been submitted yet.

[isPANN]

HPC2 materialization checkpoint

Slurm job 10006802 completed successfully on HPC2.

• submission: debug, 00:30:00, 1 task × 4 CPUs, 16 GiB, no GPU
• result: COMPLETED (0:0) in 49 s; batch peak RSS 96,692 KiB
• build: pinned Kissat 8af8e56f174b778aef3aa45af9f739b2a5f492c2 and march_cu 705b60c6491ef2b61988b3ce6ac674be1b90571d
• Kissat SHA-256: 07da6866c339de638458a7867f8d11a17162046878b552cadd22b265a8b5e7b1
• march_cu SHA-256: a0236ad5634fb680fa84220ad6806e5def15bca4d9c8850f1abbc397cedec416
• toolchain lock SHA-256: 9be804ff3e8bb99c6bbeb92e80ed1947919a4425857fe7ae7934414196e2800c
• corpus: 39 manifest rows, 39 CircuitSAT files, 39 CNFs, 39 metadata files
• validation: deterministic target generation, 64/32 + 72/36 + 80/40 width semantics, disjoint calibration/held-out splits, and independent SymPy/Miller–Rabin primality checks all passed

Logs remain under artifacts/cnc-hard-regime/hpc/logs/materialize_10006802.{out,err} on HPC2. This is setup evidence, not a scientific outcome run; subsequent calibration and measurement commands will be recorded through the H1 runscribe hypothesis.