[seanthimons]

Problem

generate_bootstrap_tasks() built one tibble::tibble() per replicate (about 0.7 ms each). For 10,000 replicates that is 7.8 s, and it runs on the calling R process before any worker starts. A Shiny app that starts a bootstrap freezes every session while this happens.

Change

Draw all replicates in a single stats::rbinom() call (a groups x replicates matrix) and build each sample with tibble::new_tibble().

rbinom() consumes the random stream in order, so the draws are identical to taking them one replicate at a time. Existing seeds keep producing exactly the same replicates; the output of the new function is identical() to the old one.

Measurements

Ward data, beta_poisson fit, this machine.

Task generation alone:

Replicates	Before	After
1,000	0.75 s	0.03 s
5,000	3.96 s	0.15 s
10,000	7.77 s	0.23 s

End to end, exponential + beta_poisson, 8 local workers (current dispatch):

Replicates per model	App blocked at start (before / after)	Total (before / after)
1,000	2.1 s / 0.1 s	6.1 s / 4.9 s
5,000	9.7 s / 0.4 s	18.3 s / 12.9 s
10,000	17.6 s / 1.0 s	32.7 s / 23.1 s

The "before" columns already include the separate dispatch fix in #28; this change is independent of it and touches a different function in the same file.

Tests

tests/testthat/test-bootstrap-tasks.R keeps the old algorithm as a reference and checks that the new output is identical to it:

• observed and fitted probabilities, with a seed;
• all three model types and a single replicate;
• well-formed, numbered samples of the original groups;
• seeded runs are reproducible and leave the global random stream alone;
• bootstrap replicates are unchanged end to end on the sequential backend.

Full suite: 463 passed, 0 failed, 3 skipped.

Not in this PR

The mirai dispatch fix (#28), any app change, and version or NEWS edits.