Allow deprecated syntax for specified versions of specified packages

[andrjohns]

The main blocker for the rstan -> CRAN transition is the number of packages that still have not updated their deprecated syntax, despite plenty of notice.

To workaround this, the current PR adds an 'exception list', where specified packages at specified versions (currently only the broken reverse-dependencies) have their Stan code canonicalised/updated at installation-time for compatibility with stan 2.36+.

If a version for one of these packages changes, then the automatic canonicalisation does not occur and the author would be required to update their Stan syntax for their package to build.

This allows for a compromise where abandoned/unmaintained packages are no longer blocking rstan updates, while any packages still receiving active development will be forced to update their syntax once a new version of rstan is published.

As part of testing this, I've created an rstan branch at the 2.39 release point here: rstan@v2.39.0 and added handling for the RNG changes (ecuyer->mixmax).

With these rstantools changes and that rstan branch, I'm able to build all CRAN reverse-dependencies under both combinations (2.32 rstan + 2.39 StanHeaders) and (2.39 rstan + 2.39 StanHeaders) except for two (tmbstan and rmdcev)

[andrjohns]

Pinging @bgoodri for the rstan release perspective, and @WardBrian on the canonicalisation approach (thanks!)

[WardBrian]

The overall approach seems prudent and sound.

Two questions:

• Do any of the affected packages use #includes? Are those already processed by the time the canonicalization would happen?
• Why 2.32? I believe 2.33 would still work, and had a few other improvements that might benefit pretty printing (though it's been a while, maybe they're the same in that regard)

[andrjohns]

Do any of the affected packages use #includes? Are those already processed by the time the canonicalization would happen?

Yeah, it uses the rstan::stanc_process function to first replace all #include chunks with the respective code so the canonicalisation is applied to the whole model

Why 2.32? I believe 2.33 would still work, and had a few other improvements that might benefit pretty printing (though it's been a while, maybe they're the same in that regard)

Pretty much just because I just copied the stanc.js file from the current CRAN rstan and it worked, so I didn't look anywhere else 😅

[WardBrian]

Sounds good to me

[andrjohns]

Reverse-dependencies check with this rstantools branch plus CRAN rstan and StanHeaders 2.39.0 is here: https://github.com/andrjohns/rstan/actions/runs/29497863526#summary-87637709672

2 install failures: tmbstan (PR opened) and rxode2ll), and ~20 test/vignette/example errors - mostly issues with optimizing and unconstraining parameters

[jgabry]

Thanks @andrjohns, looks good to me. @bgoodri?

[andrjohns]

The majority of the revdep test failures were caused by Stan's optimize now returning an extra converged__ col that rstan wasn't expecting.

Patching that out of the StanHeaders includes brings the revdep failures down to only 5: https://github.com/andrjohns/rstan/actions/runs/29560600034#summary-87834072879