[trissim]

The 3D intensity batch now uses the existing exact maximum-retention algorithm instead of sorting every foreground intensity. It shares the scalar path’s NumPy 1.24 tie ordering and existing foreground/object-domain owner, without a new kernel, cache or wrapper.

On saved original 60×256×256 production inputs, one complete-backend pair improves 0.534→0.340s (−0.194s). All 88 feature arrays and 1,764 accepted CSV values match exactly. Four existing controls pass on current main db5d6fa. Normal registered preparation installs the actual float32/int64 helper signatures before READY.

One first/only 3D job per source on current main db5d6fa versus this branch 408675a, CPU5/1w1t/INLINE, shows the changed intensity step improving 0.655413→0.564878s (−90.534ms). All four strict comparisons against the two retained native executions pass.

Whole-pipeline timing is worse in this observation: execution 4.843450→6.261316s (+1.417866s), total 6.043962→7.441233s (+1.397271s), peak RSS 1650.664→1980.078MiB (+329.414MiB). Compilation is 0.915580→0.927179s. The six largest slower steps occur before the changed intensity computation at step21; pre-step21 span deltas sum to +1.488868s. This is temporal counterevidence against that computation causing those earlier delays, not proof of a host/GC/JIT cause. Preparation and server startup are excluded from pipeline timing. No end-to-end speedup or new native timing claim.

The original qualification kept the PR draft while the whole-pipeline discrepancy was investigated. Original backend replay evidence remains tied to its recorded source and unchanged module bytes; source-only local tests do not establish runtime improvement.

Closes #634

After the recorded comparison, merged current main dc448b0 into PR HEAD 97cbbad. The intensity module bytes remain identical. The comparison above is still tied to db5d6fa/408675a27 and is not fresh qualification of the later combined source.

Merged current main 5c20b3d (including #648) into this existing branch. Three current intensity admission/batch controls pass, and the production intensity patch remains the same ten-line diff. At the user’s explicit request to merge locally qualified work, this is being merged as an exact algorithm improvement; the retained negative whole-pipeline observation above remains valid and is not reinterpreted as an end-to-end gain.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.