[Jammy2211]

Summary

Phase A0b of the search-extensibility epic (PyAutoFit#1670): dead code out of the non-linear search layer, the latent bugs survey 01 §8 named, and the two repairs the A0a(i) conformance suite carried as strict xfails. Seven commits, the test-directory move last and move-only. Shipped under the human's 2026-10-08 --auto launch; tier judge, so merge is a human /prm.

• Removed make_sneakier_pool, SneakierPool and its FunctionCache helpers (no users anywhere in the workspace), the unread NonLinearSearch.iterations, Zeus's unused discard/thin/chain; _log_process_state is one module function.
• Drawer runs under NullPaths (timer guard) and honours its search_internal argument; Nautilus no longer mutates iterations_per_full_update; AbstractNest defaults number_of_cores=1.
• Unknown constructor kwargs log a WARNING (never raise); samples_from falls back only on (FileNotFoundError, NotImplementedError) and logs it.
• SearchUpdater.update converts samples once and passes them to visualize; Emcee computes autocorrelation once per conversion (unit test counts exactly one).
• BlackJAXNUTS and SMC accept the inverse_mass_matrix kind strings their own search.json writes ("none" → None; "array"/"result" load but raise at fit time with a clear message). ROUND_TRIP_XFAIL deleted; golden identifiers unchanged (BlackJAXNUTS 2a6f5217…, SMC 3baa494f…).
• Stale AbstractNest and apply_test_mode docstrings fixed.

Two things for the reviewer (the plan did not settle them; both are reversible by reverting one commit):

1. SettingsSearch.search_dict keeps its shape: about 223 workspace and pipeline call sites splat it into af.Nautilus(...), so filtering use_jax_vmap there would have changed a public property. A new search_dict_for(search_cls) filters by signature instead. Splatting the plain search_dict into a non-Nautilus search now warns about use_jax_vmap; no call site does that today.
2. Narrowing the samples_from except exposed that plot_start_point asked an empty Emcee backend for samples and relied on the swallowed AttributeError. The start-point visualisation now passes plot_search=False and outputs only the analysis visuals; at that point there were never samples to plot, so the output set for fresh runs is unchanged.

The new kwarg warning also caught real typos outside this repo, filed as a Mind follow-up: auto_correlations_settings= in autofit_workspace searches/mcmc.py and workspace_test Emcee.py/Zeus.py (settings silently ignored until now), nlive= to DynestyDynamic in searches/nest.py, and dynamic_delta/delta in test_autofit/graphical/hierarchical/test_optimise.py.

API Changes

Removed dead parallel helpers and the unread iterations attribute; added SettingsSearch.search_dict_for, updater.log_process_state, and optional arguments on SearchUpdater.visualize / Emcee.samples_info_from / Nautilus.iterations_from. Behaviour: unknown search kwargs warn; AttributeError inside samples_via_internal_from now propagates instead of silently falling back to samples.csv; AbstractNest.number_of_cores defaults to 1 (was None). No downstream library (PyAutoLens/PyAutoGalaxy/PyAutoArray/PyAutoCTI) references any removed symbol (grep over their packages is empty).
See full details below.

Test Plan

• pytest test_autofit -x: 3073 passed, 2 skipped, 3 xfailed (only the three CONFIG_MUTATION_XFAIL cases, A3)
• unittest-nojax emulated (meta_path finder hiding jax/jaxlib/blackjax/optax/prodigyopt): test_autofit/non_linear 915 passed, 37 skipped, 3 xfailed
• autofit_workspace searches/{mcmc,nest,mle}.py and autofit_workspace_test searches/{Emcee,Zeus,DynestyStatic,Nautilus,LBFGS,Drawer}.py under PYAUTO_TEST_MODE=1: all exit 0
• grep -rn make_sneakier_pool autofit/ empty; new test__fit_with_default_null_paths fails on main, passes here
• CI green on unittest 3.12 / 3.13 / nojax and docs

Full API Changes (for automation & release notes)

Removed

• NonLinearSearch.make_sneakier_pool — no callers
• autofit.non_linear.parallel.SneakierPool (and from parallel/__init__), parallel/sneaky.py FunctionCache, initializer, fitness_cache, prior_transform_cache — only used by the removed pool
• NonLinearSearch.iterations — never read; SearchUpdater owns _iterations
• NonLinearSearch._log_process_state, SearchUpdater._log_process_state — replaced by updater.log_process_state()

Added

• autofit.non_linear.search.updater.log_process_state()
• SettingsSearch.search_dict_for(search_cls) — settings filtered to what search_cls.__init__ accepts
• NonLinearSearch._known_kwargs — kwargs never reported as unknown (save_all_samples, initial_values, inplace)
• autofit.non_linear.search.mcmc.blackjax.chains.SERIALISED_KINDS, inverse_mass_matrix_spec_from
• Optional arguments: SearchUpdater.visualize(samples=, plot_search=), perform_visualization(plot_search=), Emcee.samples_info_from(auto_correlations=), Nautilus.iterations_from(iterations_per_full_update=)

Migration

• Before: search.make_sneakier_pool(fitness) → After: no replacement (dead path)
• Before: af.SomeSearch(**settings_search.search_dict) for a non-Nautilus search → After: af.SomeSearch(**settings_search.search_dict_for(af.SomeSearch)) to avoid the use_jax_vmap warning
• BlackJAXNUTS/SMC reloaded from search.json with inverse_mass_matrix="array"/"result" now raise at fit time naming the fix (pass the original seed)

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: safe (header: safe, cap: refactor → safe)
• Plan: on the issue (refactor: search hygiene and dead code (search-extensibility A0b) #1670), written at start, unmodified since
• Gate: tests 3073 pass / 2 skip / 3 xfail (+ nojax emulation 915 pass; downstream suites n/a, no removed symbol referenced outside PyAutoFit) · smoke afW searches/{mcmc,nest,mle} + afT searches/{Emcee,Zeus,DynestyStatic,Nautilus,LBFGS,Drawer} exit 0 · review CLEAN (Fable session over the Opus-written branch; claim "identifiers unchanged, ROUND_TRIP_XFAIL removed" basis-cited by the conformance run and the golden table test) · Heart STALE release validation incomplete: no rehearsal for current source
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)? Review items 1 and 2 above
• Human: merge, amend, or reject — then log the outcome

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code