Refactor bundlePackages()

[hadley]

This turns the main strategy "inside out", so that we find all uninstalled packages, then all packages with missing repo, then collect the package metadata into a list.

I've switched the error and warning to use cli and spent a little time polishing the language. But I'm not sure it's that great yet, so I'd love your feedback :)

I also rewrote inferRPackageDependencies() to use a much simpler strategy since there can only be one appMode.

Fixes #659

[aronatkins]

Wordsmithing is unending.

[aronatkins]

It's going to take me time to grok that things like is_installed are coming from rlang rather than part of this package. All other external references use packagename::XX names.

[hadley]

Yeah, our philosophy is that rlang is the sort of missing stdlib, that we've decided that we can't live with out and use so frequently that :: is high friction.

[toph-allen]

Just a heads up that Tests are failing on R 3.6.3 and R 3.5.3. Not immediately obvious to me from the GH Actions output what's incompatible.

[hadley]

Ugh, that was a dumb and rather hard to track down mistake.

[hadley]

@aronatkins I think this is worth another review. In summary:

• We better handle my personal motivation: development versions of packages that I've installed locally (Warn if package versions are newer than those on CRAN #659). We do this by only matching CRAN if the locally-installed package name matches a CRAN package name and the local version is not ahead of the CRAN version.
• This gives me more confidence that we are detecting true failure states, so the new checkBundlePackages() function errors for both uninstalled packages and packages we don't know how to reinstall. If this causes the initial deployment to fail, you'll lose your work defining the destination, but I plan to fix that in saveDeployment() before upload? #677.
• Since the failure mode is hard to fully explain in an error message, I have fleshed out the docs in appDependencies() and point there from the error.

There is some mild risk that I have introduced bugs that are not caught by the unit tests because this code is hard to test. I think this is somewhat mitigated by the fact that I will be using the dev version of rsconnect, have a mix of CRAN + local + GitHub packages, and will be actively deploying projects over the next few weeks.

[toph-allen]

Ugh, that was a dumb and rather hard to track down mistake.

stringsAsFactors strikes again!

[aronatkins]

OK, with testing concerns about archived packages.

[aronatkins]

What defines package_version? Could you explain how this will cope with archived packages (e.g. Kmisc, from #508)?

[hadley]

package_version() is the base R function that takes a string and turns it into a version object so that 1.2.3 compares correctly to 1.10.4.

Good question about archived packages. I will investigate.

[aronatkins]

Thanks (on both). I'm more used to seeing packageVersion and forget about package_version.

[hadley]

Ok, I think I've fixed it but I need to figure out how to test it and that feels like a post-weekend problem 😄

[hadley]

I reverted the changes to the findPackageRepoAndSource() to preserve the existing algorithm, so this PR reverts to a pure refactoring. I plan to merge this shortly and then work on some more refactoring before re-tackling the motivating problem.