prompt: ep-hierarchical-regression (#1352) final decision — seed the stochastic test, A reverted

Author: Jammy2211

active.md

@@ -2,8 +2,8 @@
 
 ## ep-hierarchical-regression
 - issue: https://github.com/PyAutoLabs/PyAutoFit/issues/1352
-- status: library-dev — CORRECTED root cause (deep research): EP fits consume GLOBAL np.random via LaplaceOptimiser.n_refine=3 → refine_state → mean_field.sample → NormalMessage.sample (normal.py:274,278). Fit's fixed point (good vs sigma-collapse) depends on ambient np.random state at fit start; test seeds only before data gen, not the fit. #1351 added test_ep_statistics_fixes.py (Monte-Carlo KL tests) which shifts ambient RNG into a failing region. PROVEN: reseed-1 before fit = PASS 5/5; sweep pre-fit seed 0..7 = PASS×7, seed7 FAIL. Factor ordering ALREADY deterministic (my earlier hypothesis wrong). Fix candidates A(seed refine sampler)/B(deterministic refinement)/C(test seed) — awaiting user steer; approved "factor-schedule" approach SUPERSEDED
-- note: LATENT reproducibility bug beyond this test — all EP/Laplace fits non-reproducible across process states. Touches CORE optimiser (blast radius = every graphical/EP fit). isolation failure of test_full_hierachical is PRE-EXISTING at all commits (CI regression is the in-suite flip). Found during /health check green-light sweep 2026-07-10 (otherwise-green: 2474 lib tests + 47 workspace smoke). #1352 comments carry full evidence
+- status: library-dev — FIX = seed the inherently-stochastic test (user steer 2026-07-10: "it's inherent randomness, seeding is good if feasible"). test_full_hierachical fits a MARGINAL hierarchical EP model; Laplace refinement (n_refine=3) draws from global np.random (NormalMessage.sample), so fixed point (good vs sigma-collapse) depends on ambient RNG + variable-id ordering left by prior tests. #1351 (Monte-Carlo KL tests in test_ep_statistics_fixes.py) shifted state into a failing region. Fix: np.random.seed(0) right before the fit. Optimiser fix A (controlled refine RNG) BUILT then REVERTED — insufficient alone (id-order also flips; A+C still flips under id-offset sweep). Shipping test-seed only (1-file diff). Validated: graphical suite 216 pass. NEXT: ship_library
+- note: seed makes CI green; test stays inherently stochastic (residual id-order sensitivity documented, not chased per user "keep it feasible"). isolation failure PRE-EXISTING at all commits (CI regression is the in-suite flip). Found during /health check green-light sweep 2026-07-10 (otherwise-green: 2474 lib + 47 smoke). #1352 comments carry full RNG+id-order root-cause evidence
 - worktree: ~/Code/PyAutoLabs-wt/ep-hierarchical-regression
 - repos:
   - PyAutoFit: feature/ep-hierarchical-regression