Follow-up from the #38 seed-contract whole-branch review.
The v0.5 seed contract (#38) makes .with_mock_seed() — save/restore .Random.seed + RNGkind, L'Ecuyer-CMRG sub-streams — the seeding primitive for the paths reachable through create_mock_data() (native, postprocess, simstudy, legacy loop). But the standalone exported generators, when called directly with a seed, still use a bare set.seed(seed) under the ambient RNG kind and clobber the caller's RNG stream:
create_cat_var(), create_con_var(), create_date_var()sample_with_proportions(), make_garbage(), apply_garbage() (R/mockdata_helpers.R)create_survival_dates(), create_wide_survival_data()
These are not reachable from create_mock_data() (the orchestrator passes seed = NULL to them and owns the seeding), so there is no isolation gap in the main entry point — this is about direct callers of the standalone functions. The v0.5 NEWS note is already scoped to seeded create_mock_data()/generate_mock_data_*()/postprocess_mock_data() calls to avoid over-promising.
Fix direction: route these through .with_mock_seed() (or a shared seeding helper) so a directly-supplied seed gets the same isolation + L'Ecuyer contract. Decide whether the survival generators warrant their own frozen stage index. Add isolation tests (caller .Random.seed/RNGkind unchanged) for each exported generator.
Sized S–M. Additive; but note it would change seeded output of direct standalone calls (MT→L'Ecuyer) — bundle with the v0.5 break or document as a further break.
Follow-up from the #38 seed-contract whole-branch review.
The v0.5 seed contract (#38) makes
.with_mock_seed()— save/restore.Random.seed+RNGkind, L'Ecuyer-CMRG sub-streams — the seeding primitive for the paths reachable throughcreate_mock_data()(native, postprocess, simstudy, legacy loop). But the standalone exported generators, when called directly with aseed, still use a bareset.seed(seed)under the ambient RNG kind and clobber the caller's RNG stream:create_cat_var(),create_con_var(),create_date_var()sample_with_proportions(),make_garbage(),apply_garbage()(R/mockdata_helpers.R)create_survival_dates(),create_wide_survival_data()These are not reachable from
create_mock_data()(the orchestrator passesseed = NULLto them and owns the seeding), so there is no isolation gap in the main entry point — this is about direct callers of the standalone functions. The v0.5 NEWS note is already scoped to seededcreate_mock_data()/generate_mock_data_*()/postprocess_mock_data()calls to avoid over-promising.Fix direction: route these through
.with_mock_seed()(or a shared seeding helper) so a directly-suppliedseedgets the same isolation + L'Ecuyer contract. Decide whether the survival generators warrant their own frozen stage index. Add isolation tests (caller.Random.seed/RNGkindunchanged) for each exported generator.Sized S–M. Additive; but note it would change seeded output of direct standalone calls (MT→L'Ecuyer) — bundle with the v0.5 break or document as a further break.