[elkronos]

…d spatially balanced sampling

Wrong answers fixed (each verified by simulation before and after):

• One sampled row per stratum or interval reported a variance of exactly 0 and a zero-width interval. The analytic estimator now declines with a reason, as survey does by default.
• A stratum allocated no rows made its frame rows unreachable and biased a total by their whole contribution (84% in one example) without a word. draw() now warns, naming the strata.
• The jackknife deleted rows across strata, counting between-stratum gaps as sampling variance (64x overstated). It is now the stratified JKn and matches the analytic form exactly for stratified and temporal designs.
• One row per cluster understated the multistage variance by ~30%; it now falls back to the ultimate-cluster jackknife.
• Neyman allocation dealt over-allocation surplus round-robin; capped strata are now taken whole and the rest re-split by the rule (Cochran 5.9).
• ht_mean(estimator = "ht") printed as a Hajek mean.
• design_systematic(order_by) returned rows in sorted rather than frame order, contrary to the documented contract.

Research-based methods:

• Confidence intervals use t on the design's degrees of freedom (PSUs minus strata), matching survey::degf(); coverage at four clusters rises from ~87% to nominal. df = Inf restores the normal interval.
• Systematic PPS uses Deville's (1999) variance approximation, identical to sampling::varest() and within 0.2% of survey's Brewer option.
• Systematic samples get the successive-difference approximation (Wolter) instead of no variance at all, with its limits stated.
• ht_total()/ht_mean() gain by for domain estimates from the full sample; standard errors match survey::svyby() exactly.
• design_spread(): spatially or covariate-balanced sampling by the local pivotal method (Grafstrom, Lundstrom and Schelin 2012) with exact first-order probabilities and a local-mean variance estimator.
• plot(type = "map") shows how evenly a sample covers two dimensions.

survey handoff:

• Multistage designs are stated as two stages, so survey's standard error now agrees exactly rather than differing by 5-10%.
• A capped reservoir uses the rows the stream reached as its population.
• Systematic PPS and certainty-over-PPS map to pps = "brewer".

Housekeeping: graphics and utils declared in Imports; variance.R split so joint probabilities live in joint.R; multistage error messages no longer embed source-line breaks or repeat the simulate hint.