Retrofit claim assertions into the existing vignettes (vignettes are integration tests) #62

Principle

MockData's vignettes run the package's own code on realistic metadata, and R CMD check rebuilds every one of them in CI. That makes them integration tests, but only if a vignette fails when the code stops doing what its prose says. CONTRIBUTING.md now states the rule and the pattern (added in #61): compute every factual claim with inline R, and guard the claims with a hidden chunk of named stopifnot() conditions, so a regression fails the render with a message naming the broken claim.

Current state (2026-09-26)

The three survival vignettes follow the pattern (17 assertions between them), and a mutation check in #61 showed they catch regressions: flipping the censoring comparison failed two of them, and making it non-strict failed the third. The other 14 vignettes carry 2 assertions in total (tutorial-dates):

The tutorials with many inline values are the quickest wins: the claims are already computed, and only the checks are missing. The how-to and explanation articles with no inline values need their prose claims identified first.

Done when

• Each vignette's factual claims are computed, not typed, and asserted in hidden stopifnot() chunks with named conditions.
• Known issues stated in prose (for example Gompertz survival parameters produce degenerate follow-up times #54, Administrative censoring date is drawn per person, not as a fixed study end date #55, Minimal example: height garbage range (2.1;inf] produces NaN values #60) are asserted with the issue number in the message, so fixing the issue fails the render until the docs are updated.
• For each vignette, one deliberate code mutation in a scratch copy is shown to fail its render.

Relates to #61.