Fix serial-path reproducibility for L'Ecuyer-CMRG (list) seeds

[mronkko]

Summary

set_seed() did not actually apply list-type (L'Ecuyer-CMRG) seeds on the serial (non-parallel) execution path, so seeded runs were not reproducible when parallel = FALSE.

The helper assigned the seed to its local function frame:

set_seed <- function(seed){
    if(is.list(seed)) .Random.seed <- seed[[1L]]   # local binding -> discarded on return
    else set.seed(seed)
    invisible(NULL)
}

R's RNG only ever consults globalenv()$.Random.seed, so this assignment had no effect and the L'Ecuyer-CMRG state was silently dropped. The serial loop then ran off the ambient global RNG state.

This affected:

• runArraySimulation(..., iseed = <int>, parallel = FALSE) — genSeeds() produces a length-1 list holding a 7-integer L'Ecuyer state, which was dropped.
• runSimulation(seed = genSeeds(design, iseed = ...), parallel = FALSE) — same list-seed path.

The parallel path was unaffected, because it installs the seed via clusterSetRNGSubStream() / clusterSetRNGStream(), which assign() into the workers' .GlobalEnv.

Fix

One-line change: assign into .GlobalEnv (matching the idiom already used at analysis.R:21 and util.R:711):

if(is.list(seed)) .GlobalEnv$.Random.seed <- seed[[1L]]

Setting a valid L'Ecuyer-CMRG vector as the global .Random.seed also switches the active generator automatically, since R derives the RNG kind from .Random.seed[1]. No explicit RNGkind() call is needed (the array path already sets it).

Verification

Before the fix, two serial runs with the same iseed produced different per-replication draws; after the fix they are byte-identical (and parallel runs remain reproducible as before):

Mode	Before	After
parallel = FALSE	not reproducible	reproducible
parallel = TRUE	reproducible	reproducible

A regression test asserting serial-path reproducibility of runArraySimulation(..., iseed, parallel = FALSE) is added to tests/tests/test-03-array.R.

🤖 Generated with Claude Code