The following items should be addressed in simBinomialSeasonalExact()

1. Trial continuation after bound crossing.
   The simulation should stop a trial once an efficacy or futility bound is
   crossed at any analysis. Currently, mean looks used can exceed k,
   suggesting some simulated trials continue past a crossing. For k = 2,
   mean looks should be ≤ 2. This affects the reported mean looks and mean
   total events/enrolled. Note that Type I error (efficacy crossing
   probability) is reported under the non-binding futility convention:
   futility crossings do not prevent later efficacy crossings from being
   counted, but the stopping time should still reflect stopping at either
   bound.

2. Bounds adaptation to observed events.
   Total events at each seasonal analysis are random because all enrolled
   patients complete their season's follow-up (event counts are binomial
   given enrollment). The simulation should call toBinomialExact() within
   each simulated trial to update bounds based on the actual observed total
   events, rather than using fixed planned event counts. This also affects
   reported enrollment totals, which currently can be much larger than the
   design's planned enrollment for the fixed design. For the fixed design,
   per-season enrollment should match the design's planned enrollment.

3. toInteger() failure with ratio = 1 and rare events.
   toInteger() calls gsnSurv() with minfup = NULL, which can fail
   when the event rate is very low and equal allocation is used because
   the rounded-up integer event count is not achievable under the
   maximum-follow-up enrollment model. The proposed fix: toInteger()
   should manage to an integer sample size (rounding enrollment to achieve
   the desired randomization ratio) but round down for event counts.
   When the spending fraction at a look exceeds 1 because the integer
   event count is below the continuous design target, the spending
   function evaluates to the full cumulative alpha, which is the correct
   behavior. A warning should be issued when events are not achievable
   with the given enrollment model so the user is aware. As a current
   workaround, use ratio = 3 (or another unequal ratio) to avoid the
   error.

Notes and extensions

• The adaptive pathway shown here is blinded because it uses total event counts
  only; treatment-group differences are not used to update enrollment.
• Spending fractions are kept fixed at 1/k, 2/k, ..., 1.
• For modified intention-to-treat analyses, additional exclusion/dropout
  mechanisms can be layered into the simulation by reducing at-risk counts
  before each seasonal event/dropout draw.
• Simple superiority and non-inferiority variants can be studied by changing
  hr0/hr1 and rebuilding the same workflow.