prompt: consolidate #189 into main — one phase-1 record, follow-ups filed

[Jammy2211]

Two sessions independently wrote a completion record for the prior-support Clipper. #189 was opened first, by the session that actually shipped PyAutoFit#1477; #190 (mine) landed later because I never listed open PRs before starting. This consolidates them so main holds one record, not two.

Closes #189 in favour of this — nothing from it is dropped.

The record is now the union, and #189's half is the better half

complete/2026/08/prior-support-clipper.md keeps its slug (six things already reference it) but takes #189's body, which is materially richer than what I wrote:

• The design decision — why one relative margin * (upper - lower) is wrong in two separate silent ways, and the three-case table (two-sided finite / unbounded / half-open exclusive) that replaced it.
• Eight traps measured against a running install, including that the naive margin turns every unbounded prior into NaN — which would have made the feature actively harmful on exactly the models it targets, with a symptom indistinguishable from the bug being fixed, while every bit-identity test still passed.
• The verification log — bit-identity 10/10 on both searches, core promise 8/8, lane deaths 249 → 0 with 252 clips end-to-end, guards verified by inversion.
• The process lesson: a first commit shipped an undefined optimize in LBFGS._fit, and the full 1790-test suite passed against it, because nothing in the library suite ever executes an LBFGS fit. A green suite is not coverage.

My "What shipped after this record was first written" section is appended, since #189 predates #1478/#1479/#1480 and therefore lists follow-ups 1 and 2 as open when they are now fixed.

Phase-2 harness traps salvaged into the campaign prompt

These cost the phase-1 session real time and were about to cost the GPU session the same:

• The .completed marker short-circuits fit() — three successive versions of a resume test "passed" while testing nothing.
• fit() rebuilds search.paths, so instance-level monkeypatches are silently discarded. Patch at class level.
• search_internal is deleted on successful completion; capture it as it is written.
• Seed random and numpy before every fit, or bit-identity comparisons report spurious mismatches.
• A box containing the optimum never exercises the clipper — phase 1's first efficacy attempt measured 0 clips for exactly this reason.

Plus: arm 3 does not exist yet (phase 1 ships the mask, no momentum reset), and two phase-1 measurements to carry in as priors.

Follow-ups 3 and 4 filed as prompts

• loggaussian_prior_declares_own_support.md — LogGaussianPrior reports (-inf, inf) for a support of (0, inf). The clipper special-cases it; the prior should declare it. Flagged as needing care: the nested samplers map through the prior in unit-cube coordinates.
• clipper_in_search_identifier.md — Autonomy: human-required, deliberately. Leaving it out collides result-affecting runs; putting it in orphans every stored multi-start and BFGS output directory. Four options laid out, none pre-selected, and it is best decided before phase 3 so the re-baseline and any re-keying are not entangled.

One correction absorbed

test_nautilus.py::test__single_core_builds_no_pool passes in CI and fails only in local venvs — on 3.12 and 3.13, from two independent sessions. "Not caused by this work" is right; "pre-existing on clean main", the looser phrasing I used in #1479 and #1480's bodies, is wrong. Recorded in the consolidated file rather than by editing merged PRs.

Bookkeeping

lifecycle.py check OK, index --apply run, dashboard.md regenerated (138 → 140 drafts).

🤖 Generated with Claude Code

https://claude.ai/code/session_01FzF2XmKQaqZRWZfZMxvTNR

---

Generated by Claude Code