Fix enrollment pooling that dropped per-stratum periods in simulation inputs

[lhunt13]

Fixes #3.

Problem

get_dist_tite() and get_info_tite() pool arm-level enrollment for ahr_dd()/gsDesign2::ahr() with group_by(rate, duration) |> slice_head(n = 1), which collapses enrollment periods that have identical values in different strata into a single row keeping only the first stratum's label. In multi-stratum designs where strata share period values — e.g. the shared zero-rate lead-in periods of a sub-study that opens late — all but the first stratum lose those periods and are treated as enrolling from time 0. Expected events, information, and the simulated z-statistic means are inflated, so simulated power (Tables 6a/6b) for TTE hypotheses can exceed the full-alpha local power (Table 5). Full analysis in #3.

get_boundaries() already pools correctly, so the design boundaries and local power were unaffected.

Fix

Pool per stratum and enrollment period in both functions, mirroring get_boundaries():

enroll_rate |>
  dplyr::group_by(.data$stratum, .data$rate, .data$duration, .data$index) |>
  dplyr::summarise(rate = sum(.data$arm_rate), .groups = "drop") |>
  dplyr::arrange(.data$index)

With this change, ahr_dd()/ahr() reproduce gs_power_ahr's reference events/information/effect size to machine precision for every TTE hypothesis in the bundled platform_study (e.g. H8 first look: 104.77 events as designed, instead of 142.2). Re-running platform_study (same seed), simulated power changes only for hypotheses in late-opening sub-studies (H3–H11; H8 first-look power 0.55 → 0.43); H1, H2, and H12 are bit-for-bit unchanged. Single-stratum configs are unaffected (the pooling is a no-op there).

Tests

New tests/testthat/test-enroll-pooling.R: a two-stratum sub-study opening at month 14 with 2:1 allocation; asserts get_info_tite() matches gsDesign2::ahr() (independent reference) and get_dist_tite() matches ahr_dd() when both are fed the explicitly correct stratified enrollment. All three assertions fail on main and pass with this change. Full suite passes (98/98); R CMD check clean apart from the two notes already present on main.

Note: a smaller residual difference between simulated power and Table 5 remains after this fix, traced in #3 to the information scale used for the simulated z-statistic means (info vs info0) — left out of this PR pending your input on the intended convention.

🤖 Generated with Claude Code