mind: complete PowerLaw omega convergence research

active.md

@@ -2,6 +2,7 @@
 
 ## pix-prodigy-gpu-compat
 - issue: https://github.com/PyAutoLabs/autolens_workspace_developer/issues/125
+- prompt: active/pixelized_prodigy_laptop_gpu_phase_1_compatibility.md
 - pr: https://github.com/PyAutoLabs/autolens_workspace_developer/pull/126 (OPEN, mergeable)
 - session: codex (phase 1) -> claude 2026-08-11 evening (n_starts control + phase 2)
 - status: workspace-dev — phase 1 and phase 2 COMPLETE 2026-08-13, all 13 cells landed, PR #126 ready for review
@@ -60,14 +61,3 @@
 - artifacts-are-laptop-only: Actions artifact downloads are blocked from cloud/mobile sessions (egress policy 403s `productionresultssa2.blob.core.windows.net` on CONNECT) — this is what stopped the cloud session finishing the ingest. Both wiki drift reports were captured to `~/.pyauto-heart/release_20260807_wiki_drift/` while on the laptop.
 - do-not: do NOT use the nightly driver for a manual release — AUTONOMY.md forbids converting a manual release into the scheduled-nightly exception.
 - repos-none-claimed: this entry claims NO repos — deliberately on one line, NOT as 2-space `  - Repo` bullets, because `worktree_check_conflict` treats any such bullet as a live claim.
-
-## power-law-omega-convergence
-- issue: https://github.com/PyAutoLabs/autolens_profiling/issues/125
-- prompt: active/power_law_omega_convergence.md
-- session: chatgpt-work
-- status: research
-- validation: initial term-count convergence sweep complete; full accuracy/cost/prior/likelihood grid pending
-- finding: `component.power-law.series-vs-hyp2f1-divergence`
-- repos:
-  - autolens_profiling: research/power-law-omega-convergence
-

complete/2026/08/power-law-omega-convergence.md

@@ -0,0 +1,33 @@
+## power-law-omega-convergence
+- issue: https://github.com/PyAutoLabs/autolens_profiling/issues/125
+- completed: 2026-08-14
+- workspace-pr: https://github.com/PyAutoLabs/autolens_profiling/pull/126
+- merge-commit: ee8a34c51d1fc3fc8f1913964fff9c73eeff0b5a
+- summary: Bounded the fixed 20-term JAX PowerLaw omega recurrence across the packaged slope and ellipticity priors, measured its JAX transformation and CPU cost, and carried the backend difference through a complete 7x7 `FitImaging` fixture. The stable finding remains persistent profiling evidence.
+- validation: GitHub Actions lint run 31812775745 succeeded at the exact merged PR head across ruff lint/format, README idempotence, 41 tests, links, and all section smoke tests. The one-shot evidence run 31811929415 also succeeded with JAX 0.10.2 on CPU and the full PyAuto source stack.
+- evidence: The 20-term angular series exceeds `1e-4` relative error over 0.05736 absolute default-prior mass. Covering the live 0.999 ellipticity clamp at that tolerance needs 10,240 terms. A static binned `lax.switch` policy retains reverse-mode differentiation and `vmap`, but is 125.3x slower than 20 terms at the clamp; the bounded complete-likelihood grid moved by at most 0.005575 log-likelihood units.
+- decision: No PyAutoGalaxy source issue was opened. The tested candidate did not meet the prompt's requirement to improve accuracy without imposing worst-case cost on ordinary shapes, and the bounded likelihood probe did not establish materiality sufficient to justify that cost.
+- release: not performed; this merged profiling research changes no packaged library behavior.
+
+## Original prompt
+
+# Bound PowerLaw omega-series accuracy and cost
+
+Type: research
+Target: autolens_profiling
+Repos:
+- autolens_profiling
+Difficulty: medium
+Autonomy: supervised
+Priority: high
+Status: issued
+Source: `component.power-law.series-vs-hyp2f1-divergence`
+Issue: https://github.com/PyAutoLabs/autolens_profiling/issues/125
+
+The JAX PowerLaw deflection path uses a fixed 20-term Tessore–Metcalf omega series while NumPy uses SciPy hyp2f1. The stable detector measures a reachable, science-affecting relative deflection error of 0.297 at factor 0.99.
+
+Map accuracy and cost over public slope, ellipticity factor, angular coordinates, term policy, and actual prior reachability. Exercise eager, jit, reverse-mode grad, and vmap. Include cold-compile and warm-runtime cost plus at least one complete likelihood sensitivity probe.
+
+A constant term-count bump is not an adequate answer: the initial sweep shows ordinary factors converge quickly but factor 0.99 can need more than 1280 terms, depending on slope. Evaluate statically binned lax.cond/lax.switch scans, and reject dynamic-loop candidates that lose reverse-mode differentiation.
+
+Retain the stable finding ID. Do not modify PyAutoGalaxy in this task. Open a bounded source issue only if the evidence identifies a defensible policy that improves accuracy without imposing worst-case cost on ordinary galaxy shapes.