Affected Repositories

• autolens_profiling (primary; workspace-type repo, docs only)
• PyAutoMind (ledger only: active.md registration, prompt move to active/)

Branch Survey

Suggested branch: claude/profiling-wiki-phase-2-b8vtjm (the session's designated branch, kept identical across repos; a remote-container clone, no task worktree).

Work Classification: Workspace (docs only).
Worktree root: none — remote container clone at /home/user/autolens_profiling.

Implementation Steps

1. Release verification (git tag --contains <merge> on blobless clones of PyAutoArray / PyAutoGalaxy / PyAutoLens / PyAutoFit): PyAutoArray #505 #516 #517 #518 #519 #520 #531 #533 #537 #541 #553 #554 #555 #557 #567 #582 (+ the no-note PRs for issues #538 #543 #581); PyAutoGalaxy #595 #597 #599 #602 #605 #629; PyAutoLens #718 #720 #750; PyAutoFit #1430. Merge SHAs come from the Mind records where recorded, else from the merge commit on main.
2. Backfill the 18 pages in wiki/campaigns/ from _template.md, one group per shared source set:
   • fixed_lens_light.md — records complete/2026/09/fixed-light-{verdict,hardware,library-path,draws,scaling}.md, fixed-lens-light-source-only.md, epic complete/archive/epics/fixed_lens_light_profiling_epic.md; issues behind Fixed lens light source-only inversion: A100 verdict (#248) #250 Fixed lens light: library-path timing of unconstrained and certified solvers (#251) #252 Fixed lens light: library-path rows on CPU and the laptop RTX 2060 (#253) #254 Fixed lens light: certified active-set cost on low-likelihood draws (phase 3) #256 Fixed lens light: source-pixel scaling of the certified solver per hardware (phase 4) #258 Fixed lens light: whole-likelihood verdict on HST + Euclid × N × hardware (phase 5) #260.
   • fixed_light_numba_cpu.md — fixed-lens-light-numba-cpu.md + fixed-light-numba-{phase1,solver,levers,s4,s4b,s5,s5b,s6}.md; profiling-contracts.md; issues research(fixed-light): phase 4 wave B — A-prime permute-active-last fnnls kernel A/B (numba CPU) #276 perf: measure numba CPU source-pixel scaling #282 fix: validate profiling experiments and clarify campaign results #284 and the earlier phase issues; caveat: bridge control job 343355 not in tree.
   • hst_gpu_residue.md — hst-gpu-non-solver-residue.md + hst-gpu-residue-p{1,2,3,4}.md; issues research: trace-based one-process decomposition of the certified Delaunay GPU call (residue phase 1) #268 profiling: HST GPU residue phase 2 — matched vmap-vs-jit A100 experiment #273 research: HST GPU residue phase 3 — PSF mapping-matrix cube convolution candidates on the A100 #295 research(profiling): HST GPU residue phase 4 — reuse the certified Cholesky for log det(F+λH) #303; caveat: phase-2 1e-9 gate failed on lanes (inconclusive), phase-3 fp32 rows diagnostic only.
   • certified_positive_solver.md — certified-positive-solver.md, certified-solver-phase-b.md, certified-solver-phase-c1-lane-rate.md, certified-solver-scalar-default-flip.md; issues PyAutoArray#566, research: certified solver phase B — production jit(vmap) solver policy on A100 #300, perf: certified solver phase C1 — uncertified-lane rate on real Nautilus batches #304; caveat: 1e-9 own-composition gate vs PDIP tolerance.
   • post_certified_breakdown.md, imaging_over_sampling.md, point_source_gpu_breakdown.md — drafts: Why + the pre-registered contract from the Mind draft prompt; status stays draft.
   • numba_interferometer_revisit.md — epic numba_interferometer_likelihood_revisit.md, numba-interferometer-pack.md, numba-interferometer-kernel-levers.md, numba-vs-jax-sparse.md; feat: numba w-tilde interferometer likelihood pack + breakdown scripts #225 research: numba vs JAX-CPU interferometer curvature — levers and verdict #226.
   • delaunay_nn_a100_series.md — delaunay-nn-{breakdown,constant-split-assembly,launch-latency}.md, delaunay-walk-early-exit.md, delaunay-adapt-split-regularization.md, sibson-single-concatenated-walk.md; library PRs PyAutoArray#531 #533 #537; profiling PRs perf: DelaunayNN likelihood breakdown on A100, like-for-like with delaunay.py + JIT-timed H row #221 A100 re-measure: JAX Delaunay walk early exit (PyAutoArray#531) #224 A100 re-measure: DelaunayNN Sibson launch latency, Phase A (PyAutoArray#533) #227 results: A100 A/B of the compacted ConstantSplit assembly (PyAutoArray#537, 12x) #231 profiling: Delaunay cells default to AdaptSplit (production regularization), per-scheme pins, A100 A/B #233.
   • a100_pixelized_baseline.md — a100-pixelized-baseline.md (perf(profiling): final A100 fp64 pixelized baseline, dense + sparse, for the matrix-free comparison #241 / perf(profiling): final A100 fp64 pixelized baseline — w-tilde-native sparse breakdown + 12-leg fresh-cache grid (#241) #242), xla-triton-gemm-off.md (results: XLA autotune vs Triton GEMM probe (job 342333) + doc corrections #230, PyAutoNerves#162).
   • matrix_free_pixelized.md — matrix-free-pixelized-likelihood.md (perf(profiling): matrix-free pixelized likelihood — CG + SLQ reference and N_src sweep #247 / perf(profiling): matrix-free pixelized likelihood — CG + SLQ reference and N_src sweep (#247) #249).
   • numpy_deflections_cpu.md — numpy-deflections-p{1,2,3}.md, epic numpy_deflections_cpu_speedup.md; PyAutoArray#516 #519, PyAutoGalaxy#595 #597 #599, PyAutoLens#718; lens: deflection-angle profiling axis + phase-1 before/after (PyAutoArray#514) #210 lens: phase-2 after-pass — scipy wofz + spherical MGE branch, re-pinned cells (PyAutoGalaxy#596) #212 lens/deflections: numpy deflections phase 3 — after-numbers, stellar re-pin, note (PyAutoGalaxy#598) #213.
   • gaussian_deflections_precompute.md — gaussian-precompute-p{1,2,3}.md, epic precompute_fixed_geometry_gaussian_deflections.md; PyAutoGalaxy#602 #605, PyAutoArray#520; lens/deflections: fixed-geometry deflection memo — phase 1 measurement of record (PyAutoGalaxy#601) #214 likelihood_runtime: JAX validation and before/after for the deflection memo trace-time fold (phase 2, PyAutoGalaxy#604) #216.
   • image_source_mappings.md — image-source-mappings-p{1,2,3}.md, epic image_source_mappings_epic.md; PyAutoArray#517 #518, PyAutoLens#720.
   • compile_time.md, preoptimization_baseline.md, reference_notes.md — the ledgers themselves (multistart_prodigy_compile_census.md, multiband_pyloop_productized.md, preopt_breakdown_baseline.md, the ten reference notes) plus the July/August records that cite issues maint: PreOptimizationTimes breakdown + README dashboard (polish 4) #59 and research: MultiStartProdigy JAX compile times — measure (MGE + meshes), make fast #93; reference_notes.md gets one verified summary line per note.
   • mass_field_profiling.md — mass-field-profiling-live.md (refactor: move live profiling builders to flat fields #287 / refactor: move live profiling builders to flat fields #288), epic mass_field_epic.md.
3. Discrepancies: PyAutoArray#582 F/D counts — read the PR diff / test and the ledger's phase-1 rows, state which count is right and why on interferometer_likelihood.md; every "not verified" release state replaced by the tag from step 1.
4. No-note speed-ups: journal entries on interferometer_likelihood.md (issues PyAutoArray#538, #539/#541, #543, #581), image_source_mappings.md (#515/#517), certified_positive_solver.md (scalar-default flip retired 2026-09-24), gaussian_deflections_precompute.md (phase 3 = autolens_workspace#530); plus a "Library speed-ups with no profiling note" table in wiki/index.md.
5. wiki/index.md: rename the second section to "Closed, parked and draft campaigns", refresh every cell that changed, remove "stub pages, backfill pending"; wiki/README.md no longer says stubs await backfill.
6. Validate: python3 scripts/misc/tooling/check_wiki.py --check, python3 -m pytest scripts/misc/test -q -k wiki, ruff check . / ruff format --check . (no Python touched), then commit, push, open the PR; merge stays human (/prm).

Key Files

• wiki/campaigns/*.md — the 18 stubs (only these and the four live pages' journals change)
• wiki/index.md, wiki/README.md — index rows and the phase-2 wording
• scripts/misc/tooling/check_wiki.py — unchanged; the gate
• results/notes/*.md — read only, never edited

Profiling research wiki — phase 2: backfill the closed campaigns

• Work type: docs
• Target: @autolens_profiling
• Epic: profiling-research-wiki
• Autonomy: human-required
• Filed: 2026-09-27
• Depends on: phase 1 (autolens_profiling#337) merged

Original prompt

"be sure we are building the research wiki of all this work so the profiling work we do is kept
and tracked" — phase 1 built wiki/index.md, the four live campaign pages and the drift check;
this phase fills the ~18 stub pages.

Scope

For every wiki/campaigns/*.md stub left by phase 1 (fixed lens light, fixed-light numba CPU,
HST GPU residue p1–4, certified positive solver, post-certified breakdown, numba interferometer
revisit, Delaunay/DelaunayNN A100 series, A100 pixelized baseline, matrix-free pixelized,
numpy deflections CPU, gaussian deflections precompute, image-source mappings, compile time,
PreOptimizationTimes baseline, reference notes, imaging over-sampling, mass-field profiling,
point-source GPU breakdown):

1. Move the "why" (question + pre-registered go/no-go rule) out of the GitHub issue body and the
   Mind complete/2026/09/*.md ## Original prompt / complete/archive/epics/* ledgers into the
   page. Cite the issue number; do not paste whole issue bodies.
2. Per-phase table (phase | dates | question | rule | result | RAL jobs | PRs) and the release
   table, verified with git tag --contains <merge-sha>.
3. Caveats section: provenance gaps (e.g. job 343355 control missing from the tree behind the
   fixed-light 1.80x), laptop-only rows, inconsistent numerical gates (1e-9 vs PDIP tolerance).
4. Record the ~7 library speed-ups that have no profiling note (PyAutoArray #515, #538, #539,
   #543, #581, the scalar-default flip autolens_profiling#302, gaussian-precompute p3) as rows or
   journal entries so the index is complete.
5. Update each index row's Status / Verdict / Last updated; check_wiki.py --check stays green.

Out of scope

Rewriting or moving the results/notes/ ledgers; the dashboard; any script change.