perf: rotation-matrix grid transform + one Grid2D per VectorYX2D (numpy deflections phase 3)

[Jammy2211]

Summary

Phase 3 of the numpy-deflections-cpu epic (PyAutoGalaxy#598) — the shared-geometry lever every @transform-decorated profile pays once per call.

• geometry_util.transform_grid_2d_to_reference_frame builds the clockwise rotation from one scalar cos/sin of the angle and applies it as two multiply-adds per coordinate ([y cos a − x sin a, x cos a + y sin a]), instead of the polar form's per-pixel sqrt + arctan2 + sin + cos. Algebraically identical (max abs difference 2e-15 on a 15k-point grid); 1.42 ms → 0.21 ms on that grid (~7×).
• VectorYX2D reuses the Grid2D it is handed when it is already paired with the same mask. @to_vector_yx passes the input grid straight through, and the constructor was re-converting its values and rebuilding a second Grid2D on every deflection evaluation. One Grid2D construction per call instead of two; other inputs take the unchanged conversion path.

Net on this box (OMP_NUM_THREADS=1, hst 15,361-point Grid2D): Isothermal.deflections_yx_2d_from 1.95 ms → 0.93 ms with the PyAutoGalaxy hoists from #598 on top. The measurement of record (before/after on hst + euclid for all nine profiles, pins at rtol 1e-6) lands in the autolens_profiling PR of this phase.

The rotation is exact where the polar form left round-off: an on-axis point at angle 0 now gives an x-component of exactly 0.0 where it was ~5e-17. That is what moved one pinned sample in the autolens_profiling stellar cell (relative shift 1.0 on a 1e-17 value) and is re-pinned there with provenance; every other pin held.

Merge order: this PR → PyAutoGalaxy#599 (profiles) → autolens_profiling (numbers).

API Changes

No public API change. transform_grid_2d_to_reference_frame returns the same coordinates; VectorYX2D(values, grid, mask) accepts the same inputs and, when grid is a Grid2D on mask, its .grid is now that object rather than a copy.
See full details below.

Test Plan

• test_autoarray — 1349 passed, 61 skipped (-n 4); geometry, grids, vectors and decorator tests re-run after each change
• ruff check / ruff format --check — no new findings in the changed files (both were already format-clean on main)
• Field-level A/B vs main on the hst grid for Isothermal / PowerLaw / NFW / NFWSph / gNFW: max relative difference ≤ 1e-8
• autolens_profiling scripts/lens/deflections/{total,dark,stellar}.py hst + euclid: every pin held at rtol 1e-6 except the exact-zero sample above

Full API Changes (for automation & release notes)

Removed

• none

Added

• none

Changed

• autoarray.util.geometry.transform_grid_2d_to_reference_frame(grid_2d, centre, angle, xp) — same signature and result; rotation-matrix evaluation (no per-pixel arctan2/sin/cos).
• autoarray.VectorYX2D.__init__(values, grid, mask, store_native=False) — when grid is a Grid2D whose .mask is mask and store_native is False, self.grid is grid itself instead of a re-converted copy.

Migration

• none required

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Re4JLhTCFnGYeez8oEyM6M

---

Generated by Claude Code