prompt: tail-burndown refined — multi-root-cause (A fixed; B/I perf-timeout; C viz-timing; D-H TBD)

Author: Jammy2211

active.md

@@ -5,7 +5,7 @@
 - issue: https://github.com/PyAutoLabs/PyAutoHeart/issues/72
 - status: in-progress (supervised --auto) — ROOT CAUSE = jax 0.9.2(dev)→0.10.2(release) numerical drift on pixelized JAX inversions (isolated: only jax varied; rectangular Δ2.6e-4, mge Δ1.8e-3 rel). Human approved strategy (ii) loosen golden + keep round-trip, AND "match dev to release". DONE: (1) bumped ~/venv/PyAuto jax→0.10.2 (whole tail now reproducible in dev); (2) cluster A fixed+validated (autolens_workspace_test multi/rectangular.py+rectangular_mge.py: golden→gross-guard rtol=1e-2, exact check=vmap==jit round-trip). Committed d1d3cfe LOCAL (not pushed). Flagged to user: mge 1.8e-3 (7×) as possible (iv)-investigate.
 - worktree: ~/Code/PyAutoLabs-wt/release-validation-tail-burndown (autolens_workspace_test on feature/release-validation-tail-burndown)
-- resume: triage remaining clusters ON jax 0.10.2 (now reproducible): jax_grad×2 (slow>2min, likely same golden pattern), viz_jit×3 (autogalaxy_workspace_test — add repo to worktree), database×6, aggregator×2, chaining/slam/group, interferometer, timeouts×5 (shapelets/modeling needs release-profile fast-mode flag). Non-JAX clusters NOT yet confirmed pass-on-0.9.2 → may be real bugs. Then commit/push → PR(s), re-validate mode=release. Inventory: active/release_validation_tail_burndown_2026_07.md. 3 blocker bugs DONE—don't reopen.
+- resume: MULTI-ROOT-CAUSE (not uniform jax-drift — refined #72 comment-4966797130). A=jax-drift FIXED (d1d3cfe). C viz_jit×3 = NOT config-gap (env_vars_release.yaml correct, TEST_MODE=0 real search) — fails on wheels 'no fit.png quick-update viz did not fire'; needs slow real-search release-env repro, likely timing/flake. B jax_grad×2 = autodiff-vs-FD + eager-vs-jit(rtol1e-10) gradient-audit asserts; repro VERY slow (FD >12min, didn't finish) — tolerance OR real gradient change TBD. D-H (database×6/aggregator×2/chaining/slam/interferometer/minimal_output) unclassified, maybe real bugs. I timeouts×5 perf. Best run as register_and_iterate cluster-by-cluster (several need slow real-search/FD repros), NOT one marathon. Inventory: active/release_validation_tail_burndown_2026_07.md. 3 blocker bugs DONE—don't reopen.
 - question: https://github.com/PyAutoLabs/PyAutoHeart/issues/72#issuecomment-4966720654 (mge-drift-investigate optional; else continue)
 - autonomy: supervised (--auto launched 2026-07-14 "continue and start --auto"; test cap; no heart-ack; ship parks for human sign-off)
 - ENV NOTE: ~/venv/PyAuto is now jax 0.10.2 (was 0.9.2) — affects concurrent sessions (next-wave gradient optimizers tuned on 0.9.2). Stale jax_cuda12_plugin 0.9.2 (GPU only) + jax-finufft<0.8 conflict (unused). See [[reference_laptop_gpu_jax_setup]].