refactor: extract SearchUpdater class from NonLinearSearch

[Jammy2211]

Overview

NonLinearSearch.perform_update (~155 lines) and perform_visualization (~80 lines) in abstract_search.py mix multiple concerns: sample persistence, latent sample computation, visualization, profiling, timing, and summary output. Extract these into a new SearchUpdater class to achieve separation of concerns, where each output task is its own focused method. Search-type-specific behavior (MCMC vs nested sampling plots) uses composition via a callable.

Follow-up to #1003 (closed without implementation).

Plan

• Create SearchUpdater class in autofit/non_linear/search/updater.py with focused methods for each output concern
• Decompose perform_update into: _update_iteration_state, _save_samples, _compute_latent_samples, _profile_and_summarize
• Move perform_visualization into SearchUpdater.visualize
• Use composition for plot_results — passed as callable, no parallel inheritance hierarchy needed
• Keep thin delegators on NonLinearSearch for backward compatibility
• No changes to any search subclass (emcee, zeus, dynesty, nautilus, bfgs)

Detailed implementation plan

Affected Repositories

• PyAutoFit (primary)

Work Classification

Library

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	main	clean

Suggested branch: feature/search-update-refactor
Worktree root: ~/Code/PyAutoLabs-wt/search-update-refactor/ (created later by /start_library)

Implementation Steps

1. Create autofit/non_linear/search/updater.py with SearchUpdater class (~180 lines)

   • Constructor receives: paths, timer, logger, plot_results_func, samples_from_func, should_profile, disable_output, iterations_per_full_update
   • update() — main orchestrator (replaces perform_update body)
   • visualize() — visualization (replaces perform_visualization body)
   • _update_iteration_state() — increment counter, update timer
   • _save_samples() — sample persistence
   • _compute_latent_samples() — latent sample logic
   • _profile_and_summarize() — profiling + summary output
   • _log_process_state() — static method, moved from NonLinearSearch
   • iterations property (read/write)

2. Modify autofit/non_linear/search/abstract_search.py

   • Add lazy _updater property that constructs SearchUpdater on first access
   • Replace perform_update body (~155 lines) with 3-line delegator to self._updater.update(...)
   • Replace perform_visualization body (~80 lines) with 3-line delegator to self._updater.visualize(...)
   • Wire self.iterations through the updater object
   • Net reduction: ~220 lines

Key Files

• autofit/non_linear/search/abstract_search.py — main refactoring target
• autofit/non_linear/search/updater.py — new file
• autofit/non_linear/mock/mock_search.py — overrides perform_update, no changes needed
• autofit/non_linear/search/mcmc/abstract_mcmc.py — plot_results override, no changes needed
• autofit/non_linear/search/nest/abstract_nest.py — plot_results override, no changes needed
• autofit/non_linear/search/mle/abstract_mle.py — plot_results override, no changes needed

Edge Cases

• MockSearch.perform_update fully overrides — continues to work, never touches SearchUpdater
• plot_start_point calls perform_visualization(paths_override=...) — handled by delegator
• paths reference sharing — lazy construction mitigates; invalidate if paths reassigned

Original Prompt

Click to expand starting prompt

Issue Move search update to new class #1003 — Extract SearchUpdate class: perform_update and perform_visualization in abstract_search.py interact with many modules (saving samples, visualization, profiling, outputting results). Extracting to a SearchUpdate class would make each output task its own method.
#1003

Do an assessment of if this is a good idea and if so come up with an example of th eimplementtation. I think its clear
abstract_search needs to be simplified and have a better sense of separation of concerns.

The goal here is really to achieve seapration of concerns in abstract_search.py and move these different categories of output
results to dediciated modules. A benefit is that some of these, like samples, are tied to the type of
search (e.g. mcmc, nested sampling). So the type of output can mayb eexploit composition or stronger typing
to make the work more clear and explicit.

[Jammy2211]

Smoke Test Results — 2026-04-13

Workspace	Passed	Failed	Total
autofit_workspace	7	0	7
autogalaxy_workspace	4	2	6
autolens_workspace	6	1	7
autolens_workspace_test	3	5	8

All failures are pre-existing and unrelated to this change:

• Missing .fits dataset files (interferometer/imaging)
• JAX numerical tolerance mismatch (point source)
• probabilities do not sum to 1 in draw_randomly_via_pdf (pre-existing samples issue)

No regressions from the SearchUpdater refactor.

[Jammy2211]

Shipped — 2026-04-13

PR

• refactor: extract SearchUpdater class from NonLinearSearch #1207

Summary

Extracted perform_update and perform_visualization logic from NonLinearSearch into a new SearchUpdater class in autofit/non_linear/search/updater.py. Each output concern (sample persistence, latent variables, visualization, profiling, summary) is now its own focused method. The search class retains thin delegators for full backward compatibility — no subclass changes needed. All 1205 unit tests pass and all relevant smoke tests pass.

Session Notes

• SearchOutput name was avoided because it's already used by autofit.aggregator.search_output.SearchOutput.
• _log_process_state static method is duplicated in both NonLinearSearch and SearchUpdater since it's called independently from both (at fit start and during updates).