R devel - Jul 2021 - undefined subclass warning

A colleague recently submitted a paper to JSS and was advised to 
address the following warning which occurs when their package 
(https://CRAN.R-project.org/package=pcoxtime) is loaded:

Warning message:
In .recacheSubclasses(def at className, def, env) :
   undefined subclass "numericVector" of class "Mnumeric";
definition
not updated

After much digging I *think* I've concluded that this comes from the 
following import chain:

pcoxtime -> riskRegression -> rms -> quantreg -> MatrixModels

   that is, loading any of these packages throws the warning. 
MatrixModels Imports: *only* {stats, methods, Matrix} and loading these 
by themselves is warning-less.

    I assume there is some mismatch/incompatibility between MatrixModels 
(which was last updated 2021-03-01) and Matrix (2021-05-24), which has 
this NEWS item in the most release 1.3-3 
<https://cran.r-project.org/web/packages/Matrix/news.html>:

* removed the nowhere used (and unexported but still active) class union 
"Mnumeric" which actually trickled into many base classes properties. 
Notably would it break validity of factor with a proposed change in 
validity checking, as factors were also "Mnumeric" but did not fulfill
its validity method. Similarly removed (disabled) unused class union 
"numericVector".

    It seems that REINSTALLING the package from source solves the 
problem, which is nice, but I don't fully understand why; I guess ???? 
there are class structures that are evaluated at install time and stored 
in the package environment ...

   Any more explanations would be welcome.

   cheers
     Ben

>>>>> Ben Bolker 
>>>>>     on Wed, 30 Jun 2021 20:23:27 -0400 writes:
    >    A colleague recently submitted a paper to JSS and was
    > advised to address the following warning which occurs when
    > their package
    > (https://CRAN.R-project.org/package=pcoxtime) is loaded:

    > Warning message: In .recacheSubclasses(def at className, def,
    > env) : undefined subclass "numericVector" of class
    > "Mnumeric"; definition not updated

    > After much digging I *think* I've concluded that this
    > comes from the following import chain:

    > pcoxtime -> riskRegression -> rms -> quantreg ->
    > MatrixModels

    >    that is, loading any of these packages throws the
    > warning.  MatrixModels Imports: *only* {stats, methods,
    > Matrix} and loading these by themselves is warning-less.

    >     I assume there is some mismatch/incompatibility
    > between MatrixModels (which was last updated 2021-03-01)
    > and Matrix (2021-05-24), which has this NEWS item in the
    > most release 1.3-3
    > <https://cran.r-project.org/web/packages/Matrix/news.html>:

    > * removed the nowhere used (and unexported but still
    > active) class union "Mnumeric" which actually trickled
    > into many base classes properties.  Notably would it break
    > validity of factor with a proposed change in validity
    > checking, as factors were also "Mnumeric" but did not
    > fulfill its validity method. Similarly removed (disabled)
    > unused class union "numericVector".

    >     It seems that REINSTALLING the package from source
    > solves the problem, which is nice, but I don't fully
    > understand why; I guess ????  there are class structures
    > that are evaluated at install time and stored in the
    > package environment ...

    >    Any more explanations would be welcome.
    >    cheers Ben

Yes, Ben,
you are right on spot and very close with your final guess.

Installation and even building of packages using S4 classes (their own *or* of
other packages where they import from) does store the class
definitions already in the binary "dump" of all the R code.

So yes, Matrix cleanup (dropping unused classes actually
helping/improving the class hiearchy by making it slightly
simpler) does need to re-installation of direct Matrix
dependencies in order to avoid the above warning --- which
otherwise has zero consequences.

So the referees of your colleague's paper / package should really learn
that they are wrong in their requirement of getting rid of that
warning.

and as you've suggested in another thread, I should alleviate
the problem by uploading a new version of 'MatrixModels' models
to CRAN {solving another small unrelated buglet} so the warnings
will go away for everyone who updates their installed packages.

Martin