[Jammy2211]

Summary

A100 re-measure of the JAX Delaunay walk early exit (PyAutoArray#530 / #531): a
same-node, same-session A/B on euclid-ral-gpu-2, jobs 342315–342320, all six
inside one 14-minute window on 2026-09-07. The witness asked for a ≥ 18 ms unbatched
drop in the params→H prefix; the measured drop is −37.9 ms, 2.1× the target.

The A/B differs in PyAutoArray and nothing else: control is a private bcd15cd9
checkout (the branch's merge base), feature is 6152bdf0
(feature/delaunay-walk-early-exit). The shared /mnt/ral/jnightin/PyAuto
install was deliberately not touched — the subhalo-validation and Euclid DR1 CPU runs
(342299, 342301, 342311, 342314) were live on it throughout — so both legs prepend
their own PyAutoArray checkout to PYTHONPATH and print autoarray.__file__ plus the
checkout's git rev-parse HEAD into the SLURM log. (The shared install sits at
e36a5af4, eight commits behind the merge base, which is why the control leg is a
private checkout rather than the shared tree.)

Delaunay — HST / Hilbert-1500 / MGE-60 / ConstantSplit, fp64

ms per likelihood call; vmap/16 is batch time / 16.

Row	control	feature	Δ	ctrl vmap/16	feat vmap/16
Triangulation + interpolation	30.563	5.970	−24.593	6.715	5.007
Regularization matrix (H)	13.555	0.225	−13.330	0.741	0.070
params→H prefix (the witness)	45.149	7.228	−37.921 (6.25×)	7.555	5.139
Total step-by-step (unbatched)	99.463	62.045	−37.418	—	—
Full pipeline, single JIT	93.463	60.355	−33.108 (1.55×)	—	—
Full pipeline, vmap 16 per call	41.945	39.520	−2.425 (1.06×)	—	—

DelaunayNN sibling — shares the walk, not the concatenation

Row	control	feature	Δ	ctrl vmap/16	feat vmap/16
Triangulation + interpolation	122.450	98.060	−24.390	15.951	11.610
Regularization matrix (H)	55.467	45.663	−9.804	12.742	12.749
params→H prefix	178.971	144.789	−34.182 (1.24×)	28.757	24.424
Total step-by-step (unbatched)	230.573	196.963	−33.610	—	—

Sibson calls pix_indexes_delaunay_walk_from twice (sibson.py:522, sibson.py:648),
so it gets the early exit and the unchunked walk but not the single-call saving.

Pins passed unchanged

EXPECTED_LOG_EVIDENCE_HST held exactly on every leg — 29110.920858 for Delaunay
(342315/342316, and pinned_drift: [] on both runtime rows) and 29144.581944 for
DelaunayNN (342317/342318). The mapping is bit-identical to the eager NumPy reference on
both sides of the change.

The H-row attribution caveat

The H row moved meaning with #531; the params→H prefix did not. The breakdown attributes
step 11 as t(params → H) − t(params → interpolator outputs). Before #531 a prefix
stopping at step 6 never asked for the ConstantSplit split points, so XLA
dead-code-eliminated that second walk out of the interpolator prefix and its whole cost
surfaced in the H row's subtraction. The feature concatenates both walks into one locate
call, so the split walk now runs inside prefix 6 — inside "Triangulation +
interpolation" — and the subtraction leaves only the regularization assembly. Hence
H: 13.555 → 0.225 ms, and a small negative would have been equally legitimate.
Quote regularization_matrix_prefix_s (params→H) across library versions, not either
row alone. The NumPy/scipy path is untouched by #531 and still charges the split walk
to H. The one script change in this PR records exactly that, as a docstring paragraph
and a comment above _setup_prefix_fn — no logic change.

Likewise, the +3.878 ms "Blurred mapping matrix (PSF)" row in the four-way split is a
by-difference artifact, not a regression: the same row under vmap 16, where the
prefixes are measured independently, is flat (8.399 → 8.343 ms).

Phase 2 recommendation — do not start it

Do not start Phase 2 (static image-plane seed + fan test) on the Delaunay path.
After #531 the Delaunay params→H prefix is 7.23 ms unbatched and 5.14 ms per call at
vmap 16 — about 13% of the 39.5 ms batched per-evaluation cost, so Phase 2's
remaining headroom on this cell is a few ms per call for a change that reaches into
Mapper/AdaptImages plumbing. DelaunayNN still spends 144.8 ms unbatched / 24.4 ms per
call in the same prefix, and the cheapest cut there is the one Delaunay already took:
concatenate sibson.py's two walk calls into one. Filed as the follow-up prompt
PyAutoMind/draft/feature/autoarray/sibson_single_concatenated_walk.md.

Scripts Changed

• scripts/imaging/likelihood_breakdown/delaunay.py — module-docstring paragraph and a
  comment above _setup_prefix_fn recording the post-#531 H-row attribution shift and
  directing cross-version comparison to the params→H prefix. Documentation only, no
  logic change; ruff format --check / ruff check clean.

Results Added

• results/notes/delaunay_walk_early_exit.md — the full note (provenance table, revision
  table, per-row A/B, the caveat, the Phase 2 recommendation). Source of truth for every
  number above.
• results/breakdown/imaging/delaunay_hpc_a100_fp64_{control_2026_09_07,walk_early_exit}.{json,png}
• results/breakdown/imaging/delaunay_nn_hpc_a100_fp64_{control_2026_09_07,walk_early_exit}.{json,png}
• results/runtime/imaging/delaunay/delaunay_hpc_a100_fp64_{control_2026_09_07,walk_early_exit}.json
• hpc/batch_gpu/submit_{breakdown_imaging_delaunay,breakdown_imaging_delaunay_nn,runtime_imaging_delaunay}_a100_hst_fp64_{control,walk_early_exit}

Breakdown flags --split-setup --vmap-batch 16; runtime cell full timing only.
xla_flags recorded identically on all six rows
(--xla_disable_hlo_passes=constant_folding --xla_gpu_autotune_level=0), so these are
comparable with the 2026-09-05 rows in delaunay_nn_breakdown.md and not with the
2026-07-10 rows in preopt_breakdown_baseline.md. The pre-existing
delaunay_hpc_a100_fp64.json / delaunay_nn_hpc_a100_fp64.json rows are left untouched;
the control legs here supersede them for comparison against this branch.

Upstream PR

• Library: feat: JAX Delaunay walk — early-exit while_loop, chunk only the seed argmin PyAutoArray#531
• Workspace test: delaunay walk: JAX assertion gate for the early-exit locator (PyAutoArray #530) autolens_workspace_test#306
• Issue: feat: JAX Delaunay walk — early-exit while_loop, chunk only the seed argmin PyAutoArray#530

Library-first gate. These results were produced by the feature library
(6152bdf0) against its own merge base, so they are valid regardless of merge order —
this PR records a measurement, it does not import the new API. It carries
pending-release and merges after PyAutoArray#531 like any other pending-release
workspace PR.

Test Plan

• python -m py_compile scripts/imaging/likelihood_breakdown/delaunay.py
• ruff format --check + ruff check on the changed script (the repo's lint gate; note the file is not black-clean at main either — black is not this repo's formatter)
• python scripts/misc/wall/check_submits.py --check — 55 submits with a wall-basis contract, 0 failing
• python scripts/misc/tooling/build_readme.py --check — 64 versioned artifacts scanned, all READMEs unchanged
• All six new JSON results load via json.load
• hpc/sync.conf (gitignored) not staged; 18 paths added by name

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01B5HT8dp7sWc9qDhZp6moGr