Fix course mismatches found in a full re-audit

[PeterSlijkhuis]

Before: a few exercise checks rejected a correct answer or misread a common one, a handful of explanations did not match what the exercise or data shows, and the normal-distribution simulation could show an area for a cut-off the slider no longer displayed.

After: every lesson code block (298) and every exercise solution, alternate and wrong answer (89 exercises) was re-run in native R against the current CSVs, including lme4, lmerTest, emmeans, BayesFactor, brms and lavaan. No quoted number was wrong. These are the fixes:

• m4-3-a: also accepts ggplot2 versions where theme_classic() blanks panel.grid.major/.minor.
• m10-3-b: uses numeric_columns(), so a correct aggregate() answer is no longer rejected.
• m12-3-a: no longer claims every non-Tukey adjustment is more conservative (Holm is not, always).
• m14-3-a: an ML fit (REML = FALSE, which the book recommends) now gets an explanation instead of a bare number mismatch.
• 12-1 quiz: lm() makes the same assumptions as the pooled t-test; Welch assumes less.
• 15-2: introduces exp() as an odds ratio before the exercise that asks for one.
• 16-3: the BayesFactor default prior can change the evidence label (remote: 3.28 vs 1.62).
• 17-1: fixed.x = FALSE so lavaan does not silently drop about a quarter of its bootstrap draws.
• 17-2: the pressure/support example now has the positive sign the data have.
• 08-3: "about a tenth of a point apart" (the gap is 0.103). m7-2-a: check comment numbers updated.
• Distribution sim: the cut-off is clamped to the slider range when the SD changes, with a test.

How: native R 4.3 from the Ubuntu archive, a harness that runs each check through the same wrapCheck preamble as src/r/checker.ts. Type check, static validation and all unit tests pass locally; the webR suite runs in CI. (A first attempt at an m3-1-b branch for range() inside summarise() was dropped: the site's dplyr already stops with an error there.)