[feedback][clinical-trial-ipd-sim]Visit timing has zero natural variation

[lengning]

What an independent reviewer sees: Every single visit record lands on the exact nominal protocol day (e.g., Day 29, 57, 85...) with literally zero variance across subjects. SD of actual-minus-nominal day = 0.0 at every time point.

Why it matters: This is the single most obvious synthetic-data fingerprint. Real trials have ±3–7 day scheduling jitter due to patient availability, clinic hours, holidays, and travel. Without this, any method depending on visit-window assignment (LOCF, MMRM windowing, exposure calculations) cannot be realistically tested.

Where in the spec to fix: The forward simulation loop (Step 5) processes visits at exact nominal days. The spec should mandate a jitter step in dag_state.R or equivalent:

actual_day <- nominal_day + round(rnorm(1, mean = 0, sd = visit_jitter_sd))
actual_day <- clamp(actual_day, window_lower, window_upper)

Suggested spec addition (Step 5 or dag_state.R contract):

"All assessment dates must include per-subject, per-visit scheduling jitter (default SD = 3 days) constrained within the protocol visit window. This applies to labs, vitals, efficacy assessments, and PRO collection independently."

[lengning]

Similarly -

Dropout/discontinuation restricted to scheduled visit days

What an independent reviewer sees: All 57 discontinued subjects have treatment durations mapping to one of exactly 12 values (29, 57, 85... 309 days) — each a scheduled visit day. No subject discontinues between visits.

Root cause: The forward loop evaluates discontinuation hazard only at visit time points. When the hazard fires, the discontinuation is time-stamped at the visit day.

Recommended spec clarification (Step 5 or scm_spec_template §6.6):

"Discontinuation is a continuous-time event. At each visit, evaluate whether discontinuation occurred during the preceding inter-visit interval. If yes, sample the exact day within the interval using the cumulative hazard. Do not restrict event times to visit-day values."

# At visit t: did discontinuation occur in [t-1, t]?
interval_days <- visit_day[t] - visit_day[t-1]
cum_haz <- lambda_disc * interval_days
if (runif(1) < 1 - exp(-cum_haz)) {
  disc_day <- visit_day[t-1] + ceiling(rexp(1, rate = lambda_disc))
  disc_day <- min(disc_day, visit_day[t])
}

[efchea1]

Reading through #179-184 together, there's a clean unifying pattern worth naming explicitly: #182 (AE onset), #183 (visit timing), and your discontinuation-timing follow-up here are all the same root cause expressed three ways, the forward simulation evaluates state only at discrete visit timepoints, when the underlying clinical process (AE onset, visit attendance, discontinuation) is continuous-time. The fix pattern is identical across all three: sample the exact event time within the inter-visit interval using the relevant hazard/distribution, rather than snapping to the visit day. Might be worth a single shared spec addition (a generic "continuous-time event sampling within visit windows" pattern in scm_spec_template.md) rather than three separate patches, since #182's onset-day formula and your discontinuation-day formula here are structurally the same rexp()/rgeom() sampling-within-interval approach. This connects to a pattern I've seen on the GSD/ADTTE side too (#166, #171), discrete-grid artifacts in continuous-time clinical processes are a recurring silent-failure class across this whole project, not just here.