Declare the predictor prior scale

[beckyfisher]

What

Adds an explicit predictor_scale declaration to the default-prior routes. A user who supplies an already logged predictor can now select predictor_scale = "log", including when every value is non-negative. "concentration" makes the recorded-concentration interpretation explicit, and "auto" retains the existing sign-based choice.

Closes #317.

Why it matters

The nec and ec50 prior no longer has to depend on the units used to record an experiment. On the declared log scale, log(1) = 0 remains a tested predictor value rather than being discarded as a zero control.

Scientific decision

No property of a numeric column can distinguish precomputed logged values from recorded concentrations. The new declaration is therefore explicit. The default remains "auto" so existing calls and published analyses retain their priors.

"log" keeps every distinct supplied value and builds the existing normal prior on that scale. "concentration" omits zero controls, logs the positive concentrations, and builds the existing lognormal prior. It refuses negative values before any model or grouped level is fitted. The historical uninformative widths remain unchanged: 10 * sd(x) on an already logged predictor and the 95% coverage width on recorded concentrations. The "regularizing" set retains its 98% coverage rule.

Validation

• test-define_prior.R, including the 30-cell The nec/ec50 gamma prior cannot describe a log-spaced concentration series #302/The regularizing nec and ec50 prior is uniform where the predictor is supplied logged #314 regression gates
• test-fit_bayesnec.R, with brm() mocked to inspect the fitted prior
• test-bnec.R and test-bnec_group.R, including pre-loop refusal
• test-bnec_hurdle.R and test-bayesnechurdlefit-methods.R, including
  zero-component-call refusal for factorised fits and amendments
• installed-package call to get_priors(..., predictor_scale = "log")
• R CMD check --no-manual --no-build-vignettes --no-tests: two existing vignette warnings and two existing code/package-structure notes; no new finding

Implementation detail

The declaration is passed through bnec(), get_priors(), amend(), bnec_group(), factorised hurdle fits and joint hurdle/zero-inflated prior construction. validate_predictor_scale() centralises the accepted values and checks a declared concentration scale before model or group loops. predictor_prior() repeats the validation as a backstop for internal routes.

The internal vector previously named predictor_scale was renamed prior_predictor; it represents the whole predictor used to prime and bound both blocks of a joint fit, not the user's scale declaration.

[beckyfisher]

The PR body was updated to include the factorised hurdle preflight coverage and its zero-component-call regression tests.

[beckyfisher]

Independent review completed after two cycles.

One real problem was found in the original PR: factorised hurdle fits and amendments validated the growth and survival components sequentially. If negative predictor values occurred only among deaths, the growth component could begin sampling before the survival component rejected predictor_scale = "concentration".

This is fixed by validating the complete predictor before either component call and explicitly forwarding the validated declaration to both components. Regression tests now verify that invalid full predictors cause zero component calls. The reviewer independently reproduced the corrected behaviour and found no remaining defects on re-review.