perf: vectorise the over-sampled grid builder; skip the mask-edge overlay under fast plots

[Jammy2211]

Summary

grid_2d_slim_over_sampled_via_mask_from built every unmasked pixel's sub-pixel block in a Python loop (a linspace, a meshgrid and a stack per pixel). On a 2000x100 CTI frame that is ~180k iterations and ~12s per call, and FitDataset.__init__ forces the dataset's border relocator, which calls it once per fit — four calls, ~50s of the slowest smoke script on the Heart board (autocti_workspace imaging_ci/modeling/start_here.py, 89.5s on its slowest leg), for a relocator a CTI fit never uses.

The offsets for a given sub-size are the same for every pixel, so they are now built once per distinct sub-size and broadcast onto that sub-size's pixel centres; non-uniform sub-sizes keep the pixel-ordered layout through a per-pixel block start (cumsum of block sizes). The output is bit-identical to the loop: the old construction is kept verbatim as the oracle in a new test over random masks, uniform and mixed sub-sizes (including zeros) and a non-zero origin.

plot_array also derived the mask's edge grid on every call for the overlay. With PYAUTO_FAST_PLOTS=1 the figure is closed before it is drawn or saved, so the overlay is never seen; it is now skipped under that flag (~0.4s per figure on the CTI frame, ~100 figures per bypassed fit).

Measured on the autocti script, cold (dataset simulated inside the run, libraries from source): 117.3s → 58.5s with this change alone → 25.9s with the smoke profile's PYAUTO_FAST_PLOTS=1 as well (PyAutoLabs/autocti_workspace, same branch name).

Found and measured by PyAutoBrain hygiene ci + /ci_speedup.

API Changes

No public API change. grid_2d_slim_over_sampled_via_mask_from returns the same values in the same order; a private helper _sub_pixel_offsets_from is added beside it. plot_array behaves identically unless PYAUTO_FAST_PLOTS=1, where it no longer computes the (never-drawn) mask-edge overlay.
See full details below.

Test Plan

• test_autoarray — 1460 passed locally (libraries from source)
• new oracle test: vectorised routine == per-pixel reference, bit-exact, uniform + mixed + zero sub-sizes
• CI green on this PR
• autocti_workspace smoke gate, once its profile PR is up, shows imaging_ci/modeling/start_here.py well under its 89.5s baseline

Full API Changes (for automation & release notes)

Removed

• nothing

Added

• autoarray.operators.over_sampling.over_sample_util._sub_pixel_offsets_from(sy, sx, sub_size) — private; the (sub_size², 2) offset block one pixel's sub-pixels take, in meshgrid(..., indexing="ij") order

Migration

• none

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_015J8ZLmc55ZT1nSRntZMz9K

---

Generated by Claude Code