record: blackjax-16-upgrade-smoke shipped (autolens_profiling#143)

active.md

@@ -1,18 +1,5 @@
 # Active Tasks
 
-## blackjax-16-upgrade-smoke
-- issue: https://github.com/PyAutoLabs/autolens_profiling/issues/142
-- session: Claude Code cloud/mobile (no local worktree)
-- status: library-dev
-- worktree: none — cloud session works on the GitHub clone directly
-- repos:
-  - autolens_profiling: feature/blackjax-16-upgrade-smoke
-- prompt: active/blackjax_16_upgrade_smoke.md
-- note: Phase 0(b) / CP-2 environment half, CPU-only (human-directed): blackjax
-  >=1.6.2 install validation, blackjax.nss 2D-toy smoke with analytic logZ check,
-  af.BlackJAXNUTS 1.6 API check + PyAutoFit blackjax-cap check. GPU MGE smoke and
-  local/RAL venv upgrades stay laptop work.
-
 ## stored-sample-reconstruction-guard
 - issue: https://github.com/PyAutoLabs/PyAutoFit/issues/1486
 - status: library-dev — WORKSPACE HALF SHIPPED; the PyAutoFit hardening (#1486) is what remains

complete/2026/08/blackjax-16-upgrade-smoke.md

@@ -0,0 +1,68 @@
+Validated the blackjax ≥1.6.2 environment for the inference programme —
+Phase 0(b), the CPU half of critical-path item CP-2 — entirely from a
+mobile/cloud session (human-directed scope: install/API validation, no GPU).
+
+- **PR:** autolens_profiling#143 (merged `7a7eeaf`, closes #142).
+- **All three checks PASS:** (1) blackjax 1.6.2 + jax 0.10.2 (CPU) +
+  released autofit 2026.7.29.1 install cleanly together (`pip check`
+  clean); (2) mainline `blackjax.nss` 2D toy smoke — logZ −4.609 ± 0.069 vs
+  analytic −4.605, posterior recovered, full mainline API exercised
+  (native-space logprior_fn, num_delete, dlogz termination via
+  `state.integrator.logZ_live − logZ`, `ns.utils.finalise` +
+  `log_weights` ensemble); (3) `af.BlackJAXNUTS` runs on 1.6.2
+  **unmodified** — 0 divergences, acceptance 0.92 — no PyAutoFit
+  compatibility fix needed.
+- **Dependency findings:** PyAutoFit pins `blackjax>=1.2.0` (floor, no cap,
+  `optional` extra) — the Gate-A floor bump is a one-liner; blackjax is not
+  a hard autofit dependency, so every environment upgrade must install it
+  explicitly.
+- **Provenance upgrade for H2.1:** the installed 1.6.2 docstring states the
+  `num_inner_steps ≥ max(5, 2·d)` rule and the upward logZ-bias direction
+  verbatim — the programme's sharpest Phase-2 hypothesis is now cited from
+  installed source, not release notes.
+- **Deliverables:** `scripts/misc/searches/nss_smoke.py` (rerunnable,
+  exits non-zero on FAIL — run in each environment after its upgrade) and
+  `results/notes/inference/phase_00_unblocking/RESULTS.md`; PROGRAMME phase
+  table 0(b) → validated (cloud CPU).
+- **Remaining (laptop/RAL):** apply the upgrade to the local venvs and the
+  RAL stack and re-run the smoke there; the GPU MGE smoke half of CP-2.
+  Phase 0(c) (RAL harvest) is now the only untouched Phase 0 item.
+
+## Original prompt
+
+# Phase 0(b): validate blackjax ≥1.6.2 + mainline NSS smoke + BlackJAXNUTS 1.6 API check
+
+Type: maintenance
+Target: autolens_profiling
+Repos:
+- @autolens_profiling
+Difficulty: small
+Autonomy: supervised
+Priority: high
+
+Phase 0(b) of the inference programme
+(`autolens_profiling/results/notes/inference/PROGRAMME.md`) — the
+environment-validation half of critical-path item CP-2, runnable on CPU in a
+cloud session (human-directed 2026-08-18: "no need for laptop as not actually
+doing gpu runs for profiling just getting install type stuff working"). The
+GPU MGE smoke half of CP-2 stays laptop work.
+
+Task:
+- Install blackjax ≥1.6.2 with CPU JAX in a clean venv; record the resolved
+  versions and any dependency friction.
+- Smoke `blackjax.nss` on a 2D toy with an analytic evidence check (Gaussian
+  likelihood × uniform prior → known logZ), using the mainline API
+  (native-space `logprior_fn`, `num_delete`, dlogz termination, ensemble
+  logZ error via `utils.finalise`) — the API shape Phase 2's
+  profiling-local runner will wire.
+- Install released `autofit` alongside blackjax 1.6 and check
+  `af.BlackJAXNUTS` against the 1.6 API (import + tiny CPU fit on a toy
+  likelihood). Check whether PyAutoFit's requirements cap blackjax below
+  1.6 — a cap is a source finding for the §7 change list.
+- Record findings in
+  `results/notes/inference/phase_00_unblocking/RESULTS.md` (created by this
+  task) and update the PROGRAMME phase-state table: 0(b) validated on cloud
+  CPU, local venv + RAL stack upgrades remaining.
+
+No profiling conclusions from any timing observed here (wrong tier, CPU toy
+scale); this is install/API validation only.

complete/index.md

@@ -6,7 +6,7 @@ Token-light navigation over the finished-work records (schema:
 only then grep a dated bucket. Curators: edit the band between the CURATED
 markers; everything below GENERATED is rebuilt.
 
-1012 records across 7 buckets.
+1013 records across 7 buckets.
 
 <!-- CURATED:START -->
 ## Highlights
@@ -29,6 +29,7 @@ _(curate hard-won records here — survives regeneration.)_
 - [autohands-firewall-allowlist](2026/08/autohands-firewall-allowlist.md) — auto-closed by the merge
 - [autolens-multiplane-redshift-warning](2026/08/autolens-multiplane-redshift-warning.md)
 - [autolens-tracer-validation-guards](2026/08/autolens-tracer-validation-guards.md)
+- [blackjax-16-upgrade-smoke](2026/08/blackjax-16-upgrade-smoke.md)
 - [border-relocator-backend-parity](2026/08/border-relocator-backend-parity.md) — Isolated the native NumPy/JAX likelihood gap to mathematically non-unique PCA axes for a near-isotropic border…
 - [border-relocator-degenerate-pca](2026/08/border-relocator-degenerate-pca.md) — Stabilized near-isotropic border PCA axes with a deterministic axis-aligned branch at a scale-aware eigenvalue…
 - [circular-ell-comps-image-gradient](2026/08/circular-ell-comps-image-gradient.md) — Replaced the Sersic image path's Cartesian-to-polar ellipticity conversion with an algebraically equivalent Ca…