fix: stop passing dataset=None to fit_cls when sensitivity Job is complete

[Jammy2211]

Summary

Fixes AttributeError: 'NoneType' object has no attribute 'data' on re-runs of sensitivity mapping. Job.perform() was producing dataset = None as a "skip simulation if job is complete" optimization (commit 41095a0, Oct 2024), then handing that None straight to base_fit_cls / perturb_fit_cls. Consumer Analysis classes that unpack dataset.data / dataset.noise_map in __init__ — including the user-facing autofit_workspace/scripts/features/sensitivity_mapping.py tutorial — crashed.

The fix is a 7-line change: always call simulate_cls. Re-runs still skip the expensive non-linear search via Search.fit's load-from-zip path (paths.restore() unzips, paths.is_complete returns True, result_via_completed_fit loads samples from disk). Only the typically-cheap simulator call is no longer optimized away.

The library's own test_perform_twice test in test_autofit/non_linear/grid/test_sensitivity/test_functionality.py masked the contract issue because the test conftest's Analysis.__init__ simply stores dataset without unpacking — real workspace code unpacks it eagerly.

API Changes

None — internal change only. Job.perform() is library-internal; no public symbols added, removed, renamed, or signature-changed. The behavioural change is "re-runs of sensitivity mapping no longer crash" — strictly a fix.

Test Plan

• pytest test_autofit/non_linear/grid/test_sensitivity/ passes 25/25 (incl. test_perform_twice).
• autofit_workspace_test/scripts/database/scrape/sensitivity.py reproduced the AttributeError on re-run pre-fix; passes on a clean two-run cycle post-fix (first run completes, second run hits is_complete=True for both cells via zip presence, Search.fit.restore() unzips, load path returns).

Known limitation (out of scope)

If a previous run was killed mid-fit (leaving e.g. [perturb].zip present but [base].zip missing), the re-run will hit a different failure on that cell: Search.fit finds no zip to restore for the missing side, paths.is_complete is False, the resume logic kicks in, and Fitness.check_log_likelihood raises SearchException because the new (non-deterministic) simulator output's FoM doesn't match the persisted partial-fit FoM. That's a separate bug from this one — flagged for follow-up. This PR fixes only the is_complete=True path that produced None.

🤖 Generated with Claude Code