prompt: park jax-gradient-optimizer-benchmark (#95) at ship sign-off; log calibration row

Author: Jammy2211

active.md

@@ -3,10 +3,11 @@
 
 ## jax-gradient-optimizer-benchmark
 - issue: https://github.com/PyAutoLabs/autolens_workspace_developer/issues/95
-- status: workspace-dev
+- status: awaiting-input — all 5 candidates ran + written up; PARKED at ship sign-off (supervised). NOT committed/pushed. Resume: run ship_workspace gate → commit/push → PR on feature/jax-gradient-optimizer-benchmark (workspace-only, 6 scripts + findings.md + comparison.txt; per-run summaries gitignored).
+- question: https://github.com/PyAutoLabs/autolens_workspace_developer/issues/95#issuecomment-4961983910
 - worktree: ~/Code/PyAutoLabs-wt/jax-gradient-optimizer-benchmark
 - autonomy: supervised (--auto launched 2026-07-13 in-session; plan approved in-conversation incl. MAP objective via "do MAP, and go --auto"; no heart-ack given)
-- note: experiment/workspaces; searches_minimal JAX gradient optimizers on the MGE likelihood. Phase 1 = optax Adam/ADABelief + jaxopt L-BFGS (MAP, cold-start); Phase 2 = numpyro SVI + jaxopt Levenberg-Marquardt. Workspace-only, no PyAutoFit source. Compile-probe first; park with partial results if grad-path compile prohibitive on CPU. Ship sign-off parks (supervised).
+- result: single cold-start ALL fail wrong-basin (Adam r_E4.89 / ADABelief 5.01 / L-BFGS 4.42 / SVI 3.54±0.07); multi-start Adam 12× WINS (r_E1.600, logL+31788, 2/12 basin) = GIGA-Lens recipe. LM deferred (single-start→same basin; needs residual-vector). Follow-up filed: experiment/workspaces/next_wave_population_gradient_samplers_on_the.md (SVGD/flowMC/GGNS/SMC-HMC + converged Nautilus).
 - repos:
   - autolens_workspace_developer: feature/jax-gradient-optimizer-benchmark
 

autonomy_log.md

@@ -174,3 +174,4 @@ Outcome ∈ `merged-unchanged` / `amended` / `rejected` / `parked`.
 | 2026-07-13 | matern-tfp-jax-compat (PyAutoArray#385) | supervised | tests matern unit 2p (numpy path; full suite deferred to ship) / smoke n/a-at-park (new autogalaxy_workspace_test jax_assertions/matern_regularization.py passes: np-vs-jax parity nu=0.5+1.7, jitted, class entry point) / review deferred to ship / heart not-consulted-at-park | parked at ship sign-off (supervised bug cap; repro jax0.10.2+tfp0.25.0 crash confirmed→tfp-nightly==0.26.0.dev20260713 fixes; dep-pin + kv_xp docstring only, no API change; NOT committed; question #385 comment-4960930740; follow-up matern_vendor_bessel_kve filed) |
 | 2026-07-13 | jax-pytree-leaf-registration (PyAutoFit#1365) | supervised | tests new autofit_workspace_test jax_assertions/pytree_leaf_registration.py passes (leaf types not registered, tree_flatten roundtrip) + existing pytrees.py/enable_pytrees.py no-regression; helper unit-sane; full test_autofit deferred to ship / smoke n/a-at-park (jax_assertions not smoke-listed, matches siblings) / review deferred to ship / heart not-consulted-at-park | parked at ship sign-off (supervised bug cap; root cause = register_model registered builtins.int from af.Model(int) shapelet n/m → _partition vars(int) crash; fix = _walk skips no-__dict__ cls; producer-side, no silent guard; shapelet repro now flattens; NOT committed; question #1365 comment-4961041130) |
 | 2026-07-13 | verify-install-check-f-fix (PyAutoHeart#70) | supervised | tests bash -n OK + extracted F_driver py_compile OK + tests/test_verify_install_script.py 5p; cosmos_web_ring fits confirmed bundled / smoke n/a (harness shell only) / review deferred to ship / heart not-consulted-at-park | parked at ship sign-off (supervised bug cap; harness-only: F_driver loads bundled cosmos_web_ring pixel_scales=0.06 not stale simple + pin all 5 PyAuto pkgs to $TARGET_VERSION for coherent dev stack; no API/library change; full F PASS needs TestPyPI rehearsal re-run out-of-band; NOT committed; question #70 comment-4961101996) |
+| 2026-07-13 | jax-gradient-optimizer-benchmark (autolens_workspace_developer#95) | supervised | tests n/a (workspace experiment, no pytest/API change) / smoke n/a-at-park (searches_minimal scripts run directly on the real HST MGE likelihood; all 5 candidates completed) / review deferred to ship / heart not-consulted-at-park | parked at ship sign-off (supervised experiment cap; finding = single cold-start Adam/ADABelief/L-BFGS/SVI all wrong-basin r_E 3.5-5.0 logL~-158000, multi-start Adam 12x recovers truth r_E1.600 logL+31788 2/12 = GIGA-Lens recipe; LM deferred single-start; NOT committed; 6 scripts+findings.md+comparison.txt on feature/jax-gradient-optimizer-benchmark; question #95 comment-4961983910; follow-up next_wave_population_gradient_samplers filed) |