profiling: phase-2 instrumentation and the measured result (#507)

[Jammy2211]

Summary

The measurement half of PyAutoLabs/PyAutoArray#507 phase 2: instrument first,
then record the result.

Step 0 split the MGE operated mapping matrix row into its three pieces plus a
residual (the same residual scheme the phase-1 F rows use, so the four rows sum
to the un-split step) and recorded the sparse operator's model-independent
geometry into every result JSON. That checkpoint is what chose the phase's two
levers: profile evaluation is 78 % of the MGE row, not the PSF convolution,
and the mapper x mapper kernel does exactly 1.773e8 inner accumulations at
the HST fiducial (the plan had estimated "~2e8").

Step 4 appends the phase-2 section to
results/notes/numba_curvature_matrix_f_split.md — the step-0 split, the
per-step paired tables, the measured two-stage crossover sweep, the pins, the
pool ratio, and what was deliberately not paired — and re-runs the four cells on
the final branch.

cell	base (1b89404b)	after steps 1-2	final	whole phase
hst rectangular, step total	0.6214 s	0.4324 s	0.3334 s	1.86x
euclid rectangular	0.2226 s	0.1870 s	0.1417 s	1.57x
hst Delaunay-1250	0.6331 s	0.6076 s	0.4884 s	1.30x
hst: F mapper x mapper	0.2581 s	0.0819 s	0.0825 s	3.13x
hst: MGE total	0.1970 s	0.1937 s	0.0844 s	2.33x

Three arms, paired B/A/B/A in one session, three rounds per cell in rotating arm
order with the first discarded (numba cache: true recompiles on an arm
switch), mean of the last two.

Scripts Changed

• scripts/imaging/likelihood_breakdown/pixelization_numba.py — step 0: split
  the MGE row into MGE: image_2d_from(grid) x60, MGE: image_2d_from(blurring_grid) x60 and MGE: batched PSF convolution, with the
  pre-existing row becoming the residual; mge_split_reproduces_step records a
  one-off bit-identical check that the three pieces reconstruct the step's
  output; _geometry_constants writes the PSF shape, the sparse operator's
  stored-pair count and row lengths, and per mapper its params,
  data_to_pix_unique shape and pix_lengths into the JSON, plus the exact
  inner-operation count of both kernel forms. Step 4: _mge_blurring_stack
  now calls the same path operated_mapping_matrix_override calls instead of
  hand-rolling the per-profile loop — after perf(linear): share one eccentric-radius grid across an MGE basis (PyAutoArray#507) PyAutoGalaxy#590 the
  hand-rolled version would have kept timing the old code on the new library
  and mis-attributed the win to the residual row. It dispatches on the attribute
  so both libraries stay measurable and records which branch it took as
  mge_blurring_stack_path.
• scripts/imaging/likelihood_breakdown/delaunay_numba.py — the same two
  changes, kept in sync as STEP_ACCESSORS has been throughout.
• results/notes/numba_curvature_matrix_f_split.md — new "Phase 2 (#507)"
  section.
• results/breakdown/imaging/{pixelization_numba_breakdown_hst, pixelization_numba_breakdown_hst_rtu, pixelization_numba_breakdown_euclid, delaunay_numba_breakdown_hst}_v2026.8.17.1.{json,png} — re-run on the final
  branch.
• README.md, scripts/misc/likelihood_breakdown/README.md — auto-tables
  regenerated by build_readme.py; only this task's three rows move.
• A final commit runs ruff format on the two harnesses: the step-0 commit had
  left them failing this repo's ruff format --check lint gate (the branch had
  not been PR'd, so nothing had run it). origin/main was and is clean.

Upstream PR

• perf(numba): the mapper x mapper block and the MGE operated matrix (#507) PyAutoArray#508 — the mapper x mapper kernel and the OverSampler
  divisor cache.
• perf(linear): share one eccentric-radius grid across an MGE basis (PyAutoArray#507) PyAutoGalaxy#590 — the shared eccentric-radius grid for an MGE
  basis.

Merge order: PyAutoArray#508 -> PyAutoGalaxy#590 -> this PR. The note and the
artifacts here describe library behaviour that must already be on main.

Test Plan

• ruff check . — All checks passed.
• ruff format --check . — 181 files already formatted (was 2 failing
  before the format commit).
• python scripts/misc/tooling/build_readme.py --check — clean, i.e.
  idempotent after the regeneration.
• AUTOLENS_PROFILING_SMOKE=1 on both changed harnesses — imports + module
  setup OK, which is what the lint job runs.
• Both harnesses run for real on all four cells (that is where the committed
  artifacts come from); mge_split_reproduces_step is True on every one, and
  mge_blurring_stack_path reads shared_geometry on the final artifacts
  and per_profile_loop on the base arms.
• Pinned log-likelihood checks at the explicit rtol=1e-6 PASSED on all
  three pinned cells and are unchanged to every recorded digit.

Heart

Heart is RED at ship time for a pre-existing reason unrelated to this task.
Verbatim from pyauto-heart readiness --json (2026-08-28T21:34:42Z, score 45):

• red_reasons: "release validation FAILED (stage integrate)"
• yellow_reasons: "workspace validation not passing (2 failed, cloud#33179766004: autolens_test scripts/imaging/rectangular_mge.py, autolens_test scripts/imaging/rectangular_mge_rtu.py)",
  "manifest drift: session-start hooks (generated) — 32 mismatch(es) vs PyAutoMind/repos.yaml"

Human authorisation to ship over it, given 2026-08-28, verbatim:

prm and then kick off phase 2, I authorize the heart RED thing

Generated by the PyAutoLabs agent workflow.