Release v3.2: The Tailor

[3.2.0] — The Tailor (2026-09-05)

Added

• Sequential Chunk Stitching (_boundary_mapping): Implemented an automated stitching stage following the bisection phase to track features seamlessly across chunk boundaries via mutual pixel-overlap matching between adjoining boundary slices.
• Lightweight Boundary Reader (_read_boundary_frame): Added a memory-mapped boundary frame loader (memmap=True) that reads only the boundary slices (2D) without loading entire 3D FITS cubes into RAM.
• Stale Output Cleanup: Added automatic purging of old FITS files in 03-assoc/ prior to execution to prevent leftover files from previous runs from contaminating results.

Changed

• RAM-Safe Capped Pairwise Bisection: Capped bisection rounds to max_iter - 1 (stopping at temp8 / 512 frames). This prevents parallel processes from generating or concatenating 1024+ frame cubes, eliminating out-of-memory (OOM) fatal errors on large datasets.
• Unbounded Feature Lifetimes: Decoupled maximum trackable feature lifetime from the bisection depth limit (formerly capped at 2^(max_iter) frames). Features persisting across chunk boundaries now maintain continuous identity indefinitely.
• Strict Peak RAM Bounding: Restructured the stitching stage so that at most one 3D chunk is held in memory at any time alongside two 2D boundary frames.
• Robust Temp Directory Resolution: Replaced static round-index tracking with dynamic detection of the highest populated temporary stage, ensuring compatibility with dataset lengths that are not powers of two.

Fixed

• Inter-Chunk ID Collisions: Fixed potential label collision where unmatched features in subsequent chunks could inherit IDs already assigned in earlier chunks. Unmatched features now offset strictly outside existing global positive and negative bounds while boundary-matched features correctly inherit their parent tracks.
• Dimensionality Consistency: Resolved handling of 2D vs. 3D cubes when carrying forward odd-numbered trailing files across intermediate bisection iterations.