feat: put the clipper in the search identifier (MLE searches only)

[Jammy2211]

Overview

The prior-support Clipper (#1477) does not enter the search identifier, so two runs differing only in prior-support enforcement — which demonstrably changes the answer — share one output directory, and with the .completed short-circuit the later run silently returns the earlier one's numbers (this bit the phase-2 validation campaign, which worked around it with unique per-arm names). Decision recorded in PyAutoMind (2026-08-18, human): the clipper enters the identifier, scoped to the searches that consume one; nested-sampler and MCMC identifiers must stay byte-for-byte identical.

Plan

• Declare the clipper an identifier field on the two clipper-consuming search families (multi-start gradient and BFGS), so runs with different clippers get different output directories.
• Leave the nested samplers, MCMC searches, and Drawer completely untouched — their identifiers must not re-key.
• Pin the nested-sampler identifiers with regression tests captured on main before the change, plus a structural tripwire so a future refactor cannot silently put the clipper within their reach.
• State plainly in the release notes that existing multi-start / (L)BFGS output directories re-key (orphaned on disk, not deleted).

Detailed implementation plan

Affected Repositories

• PyAutoFit (primary, only)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoFit	main @ c302f51	clean

Suggested branch: claude/clipper-search-identifier-4cowi2

Mechanism (verified at c302f51)

• The identifier hashes [search, model] (+ unique_tag) — autofit/non_linear/paths/abstract.py:279-283.
• An object declaring __identifier_fields__ contributes only those attributes plus its class name — autofit/mapper/identifier.py:113-121; an empty tuple contributes the class name only.
• NonLinearSearch.__identifier_fields__ = tuple() (abstract_search.py:318) and neither AbstractMultiStartGradient nor AbstractBFGS overrides it — so today only their class name hashes, which is why the clipper (and everything else on them) never entered.
• clipper is resolved on AbstractMLE.__init__ (mle/abstract_mle.py:16-22); the nested samplers and MCMC searches contain zero clipper references and declare their own explicit __identifier_fields__ tuples, so they are unchanged by construction.

Implementation Steps

1. autofit/non_linear/search/mle/multi_start_gradient/search.py — add __identifier_fields__ = ("clipper",) to AbstractMultiStartGradient (covers MultiStartAdam/ADABelief/Lion/Prodigy).
2. autofit/non_linear/search/mle/bfgs/search.py — add __identifier_fields__ = ("clipper",) to AbstractBFGS (covers BFGS/LBFGS).
3. Drawer untouched: it inherits the clipper attribute but never uses it (zero references in mle/drawer/search.py) — a setting that cannot affect the result must not re-key it. Its __identifier_fields__ = ("total_draws",) stands.
4. No change under nest/ or mcmc/.

Tests

• test_autofit/database/identifier/test_identifiers.py::test_dynesty_static (line 389) pins DynestyStatic's exact hash list — must pass unchanged; it is the existing guard for the hard constraint.
• Add sibling hash-list pins for Nautilus and DynestyDynamic (ideally Emcee/Zeus/NUTS too), captured on main before the change and asserted after.
• Structural tripwire: nested samplers have no clipper attribute (assert not hasattr(af.Nautilus(), "clipper")) — trips if a refactor hoists clipper from AbstractMLE to NonLinearSearch.
• MLE side: MultiStartAdam() vs explicit clipper=ClipperNone() identical; vs ClipperPriorBox() different; two ClipperPriorBox margins different; same trio on LBFGS.

Key Files

• autofit/non_linear/search/mle/multi_start_gradient/search.py — first identifier-fields declaration
• autofit/non_linear/search/mle/bfgs/search.py — second identifier-fields declaration
• test_autofit/database/identifier/test_identifiers.py — pins + tripwire + MLE separation tests

Question for the human at PR time

Because these searches' __identifier_fields__ is empty today, n_starts, total_steps, learning_rate etc. also collide. This PR re-keys the MLE directories anyway, so widening the tuple now would pay the orphaning cost once instead of twice — but that widens scope beyond the recorded decision. Default is clipper-only; if widening is declined, it gets filed as its own prompt.

Accepted cost (for the release notes)

Every existing MultiStart* / BFGS / LBFGS output directory re-keys, including default-ClipperNone runs (the hash gains the clipper field itself). Stored results are orphaned on disk, not deleted; re-running recomputes into a fresh directory. Nested-sampler and MCMC results are unaffected.

Original Prompt

Click to expand starting prompt

PyAutoMind active/clipper_in_search_identifier.md (formerly draft/feature/autofit/) — the decided revision of the prompt, recorded 2026-08-18. See that file for the full decision record, the rejected options, and the grounding audit.

Decision, verbatim: "The clipper enters the search identifier — but only on the searches that consume one, and the nested samplers' identifiers must stay byte-for-byte identical."

[Jammy2211]

Library PR created: #1494 (branch claude/clipper-search-identifier-4cowi2, pending-release).

• Source change: __identifier_fields__ = ("clipper",) on AbstractMultiStartGradient and AbstractBFGS — nothing else touched.
• Verified live: nested samplers (Nautilus, DynestyStatic, DynestyDynamic), MCMC (Emcee, Zeus) and Drawer identifiers byte-identical to their pre-change captures; ClipperNone default does not fork; ClipperPriorBox and its margins do.
• Tests: identifier file 29/29; full suite 1864 passed, 1 failed = the documented environment-specific test_nautilus.py::test__single_core_builds_no_pool (reproduced identically on a clean tree).
• Workspace impact: none — no workspace script consumes identifier values; no API surface changed. No /start_workspace follow-up needed.
• Open question for the human on the PR: widen __identifier_fields__ beyond clipper in the same re-key event, or file separately (default).

Merge stays human.

---

Generated by Claude Code