Improve global env and knitr unwind

[lionel-]

Closes #136
Follow-up to #187

• Consistently use session scope inside knitr, no matter the evaluation environment.
• Do not use session scope with non-top-level evaluations in the global environment.

[jennybc]

LGTM

I made a few small suggestions.

I test drove this a bit and everything feels right.

Something that doesn't work the way you might expect (but also didn't work before) is this:

If you render foo.Rmd with rmarkdown::render("foo.Rmd") in your current R session, any defer()ed events remain deferred after the render. Which makes sense and yet feels semi-wrong. You can clear them with withr::deferred_clear() or run them with withr::deferred_run() or by restarting R. I'm just bringing this up to plant the idea that maybe one day we'd want a way to express that events scheduled during the course of rendering an .Rmd should be executed "when this render is done". Basically, a notion of .Rmd scope or file scope, I suppose. The obvious / low-tech way to do this is to simply insert a concluding chunk that contains withr::deferred_run().

But I definitely don't want to clog up this PR with a new matter. It's just something I thought about while testing this.

[lionel-]

If you render foo.Rmd with rmarkdown::render("foo.Rmd") in your current R session, any defer()ed events remain deferred after the render. Which makes sense and yet feels semi-wrong.

I agree. Also in notebooks, rendering chunks causes the deferred expressions to pile up in the session. I think the knitr support as it stands in this PR should mostly be regarded as a stopgap.

To implement chunk scope or document scope, we'd need a redesign of evaluate to use eval_bare() instead of eval() when looping over the lines of a chunk, as well as a modification to knitr. Hadley told me yesterday this wouldn't be worth it.

Worth noting that the "document scope" that you propose would devolve to "chunk scope" when knitting a single chunk in a notebook, IIUC. This seems to make sense.

Also I realised yesterday that in document-generation environments such as roxygen2, we can make something like this work easily, without any changes to the rmd stack:

local_foo(.frame = my_scope())

I look forward to being able to clean up state in Rmd docs using a similar pattern. Though IIRC @gaborcsardi told me this might not be straightforward because of the way Rmd pieces are assembled from multiple locations.

[hadley]

Won't you get the same problem as source()ing a file that uses local_*() at the top-level?

Would it be possible to look in the call stack for specific calls (like source() or render()) and attach the cleanup to that frame?