[beykyle]

Summary

Three fixes:

1. kernel(params=) is keyed by hyperparameter name. sklearn orders a kernel's hyperparameters alphabetically by attribute name, so a positional list for a kernel with two or more free hyperparameters could silently swap their priors.
2. An error model that isn't finite at θ has zero density instead of raising. An overflowing kernel amplitude used to escape log_posterior as a scipy ValueError and crash the sampler.
3. Faster marginal priors. norm, uniform and hypsecant use closed-form logpdf. A frozen scipy logpdf costs ~100 µs per call; a 10-parameter log_posterior drops from ~0.94 ms to ~0.25 ms.

Behaviour changes

params=

• Takes a mapping keyed by kernel_param_names(kernel), e.g. params={"length_scale": ell, "periodicity": per}.
• A positional list or tuple is accepted only for a kernel with a single free hyperparameter. Every existing caller is in that case.
• Unordered containers (set, generator, a dict as a multi-element value) raise TypeError.
• A key naming a fixed hyperparameter is reported as fixed.

Likelihood at an invalid θ (non-finite prediction or covariance): returns -inf, never raises.

• The factorisations run unchecked. CompiledConstraint._stats treats a non-finite d2 or logdet as zero density. The hot path does no per-term scanning.
• This also fixes a pre-existing Woodbury bug: an overflowing residual gave d2 = inf - inf = NaN, so log_posterior was nan and emcee raised.
• Overflow no longer emits a numpy RuntimeWarning on every evaluation.
• A rejected covariance containing NaN warns once per constraint and points to constraint.matrix(theta), which names the term.

Construction: constant terms are checked once, and a non-finite contribution raises ValueError naming the term. A NaN matrix term now gets that error instead of "not symmetric".

Predictives at an invalid θ: raise a public ValueError. Previously they returned NaN draws, which made the whole percentile band NaN.

• Affected: predictive_draws (including model_only and given=), grid_draws, gp_predictive_draws and matrix().
• The message names the sample row and its parameter values, and gives the likely causes: undiscarded burn-in or wrong column order; in a prior predictive, a prior covering a region where the error model is undefined.
• A finite but singular covariance is still drawn from, because term subsets are legitimately rank-deficient.
• heldout_log_predictive(given=) returns -inf at such a row, as it already did without given=.

Priors

• norm and uniform are bit-identical to scipy wherever scipy is finite. They return -inf instead of raising where z overflows.
• hypsecant stays finite in the tails, where scipy's overflows to -inf past |x| ≈ 710.
• Families are matched by exact type; anything else falls back to dist.logpdf.

Other: KernelTerm uses Term's short repr. kernel_param_names is added to the API reference.