Refactor theta_prior into extensible Prior objects

[blankjul]

What

Lifts the Dace MAP length-scale prior from a bare (mean, lam) tuple (hardcoded as a Gaussian ridge inside DaceProblem) into a small, extensible Prior abstraction:

• Prior — penalty(Z) + grad(Z) in the encoded log10 search space; the extension point.
• GaussianPrior(mean, lam) — the ridge lam·Σ(log10 θ − mean)², reproducing the tuple byte-for-byte.
• resolve_prior — None → None, (mean, lam) → GaussianPrior, a Prior → itself.

DaceProblem now delegates the penalty and its gradient to the Prior. theta_prior (on Dace, and via the MAP selection) accepts either the tuple or a Prior object.

Why

theta_prior was the one un-abstracted knob in the otherwise-object-based Selection framework. This makes it extensible and byte-identical.

The backstory (honest)

This branch started as a SAAS sparse prior (SAASBO-style). I built it and benchmarked it, and found — and verified directly — that a MAP SAAS is not competitive: learned-τ MAP is degenerate (it prefers dense solutions — the classic MAP-of-a-hierarchical-model pathology, exactly why SAASBO uses full-Bayesian NUTS), and fixed-τ prefers sparse but its accuracy benefit is seed-dependent. Rather than ship an uncompetitive feature, I dropped SAASPrior/SAAS and kept only the clean, byte-identical Prior refactor. A genuine sparse prior returns with the planned fully-Bayesian GP work (HMC/NUTS + Bayesian model averaging), where marginalization — not optimization — is what makes it work; the Prior base is the seam it will plug into.

Safety

Pure refactor: theta_prior=(mean, lam) behaves identically, pure-MLE untouched. Golden 19/19 byte-identical, full pyclawd check green (523 tests).

Targets the docs integration branch.

🤖 Generated with Claude Code