[mikessh]

Behavior

Gap-prior preparation previously evaluated every hypothetical length through the longest input. A long no-hit query therefore created millions of Python callbacks even when the axes had only two distinct lengths.

Prepare the existing dense native cube only for input length pairs. Every used cell retains the original prior value; scalar scores, grouped counts, shared priors and lane overrides keep their geometry. Empty/equal-length axes avoid unused callbacks. Callbacks are pure functions; unused length values are no longer evaluated or validated. Negative used-cell values remain errors. Dense cube allocation remains cubic in the maximum length.

Verification

• Added eight input-length/scalar-oracle and empty/equal-length checks; the four main cases failed before the fix.
• Complete seqtree suite: 1,045 passed, four existing optional skips; 7.36 s whole invocation. Strict Sphinx passes.
• Unchanged downstream VDJmatch suite: 1,522 passed in one 79.17 s invocation under its 120 s limit; 0.611 GB peak whole-tree RSS. This is a local candidate check, not a new release.
• Targeted complete-calibration profile: 48,722,400 prior callbacks become 468. Native binary, scoring costs, model populations and scientific outputs remain unchanged.

Prepared unpublished 1.3.1 metadata and built its wheel/sdist. A clean wheel install verifies exact scalar/matrix scores and native consumers with NumPy absent; the packaged Python source matches the tested tree. All nine upstream checks pass on the final 1.3.1 candidate head, including Windows C++, Python 3.10/3.12, benchmark/oracle tables and the appendix. Publication remains subject to separate author approval after final checks.