feat(survey): evaluate LHS points in greedy-nearest order

[vsbuffalo]

camdl survey now evaluates its LHS points in a greedy-nearest order — seed a coarse spread, then each round evaluate the unevaluated points closest (in transform-normalized parameter space) to the current best-loglik point, a batch at a time so every core stays busy.

Why

The progress bar's best-loglik metric climbs fast and plateaus early when the box actually holds a basin — an at-a-glance read on whether the survey bounds are placed well — and the good region is the part that fills in first. Easier to watch, and quick to notice a mis-placed box. (No early-stop — deliberately. Every point is still evaluated; a survey's job is the global picture, and best-first + early-stop could bias toward the first basin and miss a distant better mode.)

Byte-identical output

This reorders when points run, never which:

• every point is still evaluated;
• each point's result is keyed on (seed, point_id) (derive_point_seed / mix_cell_seed), not a shared RNG stream, so it's independent of evaluation order;
• the final sort by (loglik desc, point_id asc) is order-independent.

So landscape.tsv is byte-identical to the old index-order sweep — no golden / expected-output / CAS-identity impact. The greedy distance metric reuses EstimatedParam::to_transformed (log/logit-aware) and only orders evaluation; it never touches a loglik.

Test

survey_greedy_order_invariant.rs pins it from outside: the same survey run at --parallel 1 vs --parallel 4 walks two different greedy orders and must produce byte-identical landscape.tsv. Goes red if a future change makes a point's result order-dependent.

Green locally: the new invariant test, plus survey_roundtrip and fit_survey_denominator (no regression).

[vsbuffalo]

Closing — wrong target. This applied greedy ordering to survey, but the intent was the 2D profile likelihood grid. Redoing against profile.rs.