[Rd] Loading package namespace in data scripts vs. vignette check

[Ivan Krylov]

Hello,

I've been experimenting with a package providing its own vignette
engine and stumbled upon a problem similar to PR19020 [*]: the vignette
engine didn't see any vignettes under R CMD check. This was mostly
self-inflicted: contrary to WRE 1.1.6, the script in data/*.R loads the
package namespace. Since R CMD check calls tools:::list_data_in_pkg()
on the source package directory, this loaded the old installed version
of my package, the one without the vignette engine.

A popular idiom on CRAN avoids this problem by using delayedAssign():
list_data_in_pkg() sees the names of the datasets without causing the
side effects from forcing the promises, and all works well. Only a few
packages reference their own namespace in their data scripts without
delayedAssign (albatross, diptest, fastbeta, PhotoGEA).

I'll fix my data script so that it doesn't load the namespace. Does
this warrant any hardening in R CMD check? Perhaps the data check could
use a temporary library path, or vignette processing could be moved
into a child process (solving the workspace persistence problem as
well).

-- 
Best regards,
Ivan

[*] https://bugs.r-project.org/show_bug.cgi?id=19020