Re-arm regularizers when the parameter layout changes (saturated GoF was unwalled)

[lucalavezzo]

Separate from #147 — that one is about the external-term constant, this one is about regularizers reading the wrong parameters.

The bug

A regularizer that resolves a parameter position is only correct for the layout it resolved against — and the layout is not stable within a fit session. The saturated goodness-of-fit path swaps the model for CompositeParamModel([original, SaturatedProjectModel]), whose [poi(m1), poi(m2), ..., pou(m1), pou(m2), ...] ordering inserts one POI per projected bin ahead of the original model's block. In a 2D ptll × yll SCETlib fit that displaces all 7 λ by 39 positions.

Nothing tells the regularizers. rabbit_fit.py restores regularizers after the composite re-init but never re-arms them, and the arming loop lives in Fitter.defaultassign() — which that path doesn't call. It calls xdefaultassign(), deliberately, so the saturated fit starts from the parameter defaults.

For WRemnants' NPDampingWall, which caches positions at construction, the effect was:

wall asks for      reads index   which now holds           value
lambda2_nu              0        saturated_ch0_ptll0      +0.935
lambda4_nu              1        saturated_ch0_ptll1      +0.940
lambda_inf_nu           2        saturated_ch0_ptll2      +0.939
...

Per-bin factors near 1.0, far above the wall's 0.005 margin — so every condition was trivially satisfied and the penalty was identically zero. The saturated reference ran unwalled while the nominal fit it is compared against was walled: a likelihood ratio whose two halves had different objectives. The reference duly parked at lambda2_nu = -0.231, a point that costs 1225 nll units once the wall is applied.

Impact on results

Measured on 2D real-data fits — same card, same minimiser, same start, only the wall now readable:

fit	chi²/39 before	after	p before	p after
MSHT20 prior-free	47.58	62.11	16.28%	1.07%
MSHT20 lattice cov	33.44	65.56	72.10%	0.49%
CT18Z prior-free	57.19	56.29	3.01%	3.60%

Directly verified: the broken reference's postfit lambda2_nu = -0.23081 (wall would charge 1224.8); the fixed one lands at +0.00488, resting on the 0.005 margin, with the wall charging 0.000.

There was already a guard

set_expectations has a width check whose comment names this exact scenario ("a wrapping/composite param model (e.g. the saturated goodness-of-fit path) reorders/resizes the block in a way this wall's flat indexing cannot follow"). It never fired, because it sits on a code path the saturated fit doesn't take.

Related: 7bd0d81 fixed the x0 permutation for this same layout; the line directly below it re-points regularizers without rebinding them.

Changes

• Regularizer.set_expectations gains parms (names of the current vector), defaulted to None so existing implementations keep working. Docstring now states the layout can be reordered and resized mid-session and that positions must not be cached — it previously said "called once per fit configuration", which reads as an invitation to cache.
• Regularizer.resolve_indices(parms, names) resolves by name and raises on a missing name, so the easy path is the correct one.
• Fitter.arm_regularizers() passes self.parms; defaultassign() delegates to it; rabbit_fit.py calls it after the composite re-init.
• Staleness tripwire: init_fit_parms marks regularizers stale; _compute_nll_components refuses to evaluate a penalty while they are. A future path that changes the layout and forgets to re-arm crashes instead of silently applying each regularizer to whatever sits at its old positions.

Also included: persisting the saturated fit

Under results["mappings"][<mapping>]["saturated_fit"], using the same key names the primary fit uses at top level (parms / nllvalreduced / edmval / cov / epoch_loss / minimizer_status), so a reader can treat the sub-group exactly like the main results dict.

Previously that fit's entire state was discarded and only its chi² survived, which made ordinary questions unanswerable from the files: where the reference put its parameters (the parms beside the chi² are the nominal ones echoed back by the outer --noFit pass); which bins the model misses (the saturated POIs are one per projected bin, so their postfit values are a per-bin residual map); whether the reference converged (this step runs --noHessian, so no edmval, and scipy's outcome was logger.debug-only); and what nll_sat is (recoverable only by inverting the chi² definition).

A few tens of kB against a ~100 MB cov. It's in this PR because it is what made the bug visible.

Notes for reviewers

• This changes published goodness-of-fit numbers for any walled fit. That is the fix, but it will look like a regression to anyone diffing.
• svd.py needs no logic change (it holds no per-parameter positions); its signature is widened for interface compatibility.
• Tests: tests/test_regularizer_layout.py covers name resolution across two layouts, refusal to guess on a missing name or absent parms, and the backward-compatible signature.
• Lint: black / isort --profile black / flake8 with the CI flags, run in the CI container image, clean repo-wide.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CnJ9YKK8c1q1sCDouM6y1c