Final stretch of standardizing Butina

[Matthew-Neba]

Summary

• Standardize Butina clustering behavior with RDKit.
• Ensure deterministic higher-index tie-breaking and consistent cluster ordering (exactly matching RDkit), took a small hit in performance due to the extra work done

[greptile-apps]

Greptile Summary

Standardizes Butina clustering output with RDKit.

• Uses higher point indices to break equal-neighbor-count ties consistently.
• Orders cluster IDs and centroids by descending cluster size and centroid index.
• Consolidates packed-candidate encoding and decoding across fused and non-fused implementations.
• Replaces version-dependent ArgMax handling with a deterministic packed 64-bit reduction.

Confidence Score: 5/5

The PR appears safe to merge with no actionable correctness, security, or compatibility failures identified.

The packed reduction preserves descending neighbor-count and higher-index ordering, temporary centroid storage remains correctly sized and stream-safe, and cluster IDs stay aligned with reordered centroids.

Important Files Changed

Filename	Overview
src/butina.cu	Implements deterministic higher-index tie-breaking, coordinated cluster/centroid remapping, and internal centroid storage when centroid output is omitted; no actionable defect identified.
src/butina_common.cuh	Adds shared helpers for decoding packed Butina candidates while preserving inactive-candidate handling.
src/fused_butina.cu	Updates fused centroid selection to use the renamed shared candidate-decoding helper without changing behavior.

_{Reviews (1): Last reviewed commit: "Final stretch of standardizing Butina" | Re-trigger Greptile}

[scal444]

This backup path was here because older versions of CCCL ship with olders versions of CUDA. The first CUDA that natively ships sufficient CCCL is 12.9 https://docs.nvidia.com/cuda/archive/12.9.1/cuda-toolkit-release-notes/index.html#cuda-toolkit-major-components

I'm hoping to support 12.5 for a few more release cycles due to some users being stuck on older AWS instances. Can you check on this?

EDIT: I see your replacement is in-kernel cubmax, not devicemax - so it may be this concern is invalid. If you can quickly double-check the API is supported in older cccl, this is probably good to go. Thanks!

[Matthew-Neba]

Yeah, I just double checked it on 12.5, it works. Since the O(n^2) distance thresholding step dominates the runtime anyways (>60%), and some added complexity would'hv been needed to make the new changes work with DeviceReduce (which would'hv improved runtime by 1.5% at 30k fingerprints), decided to go for the in kernel cubmax.