Properly report multiple root causes of wildness of the same pointer

[mattmccutchen-cci]

My recollection of the discussion at the 2020-10-20 status meeting (with the usual caveat that I'm still not solid on the subject matter):

A pointer may have multiple immediate root causes of wildness; the example mentioned was a void * in an extern. 3C should be able to report all of them. Currently 3C cannot because the pointer gets a constraint edge from the global "wild" node and each constraint edge can have only one associated reason. I guess one reason per edge has been sufficient for all other edges, but not for root-cause-of-wildness edges. Possible solutions:

1. Allow every edge to hold multiple reasons. May complicate all code that handles edge reasons and hurt performance.
2. Use additional nodes as needed to represent root causes of wildness so we have a separate edge for each reason.

Since solution 2 was my idea, Mike suggested that I might eventually implement it as my first subject-matter contribution. However, someone who knows 3C better than I do should think for a moment if any other case is likely to arise in which we might want multiple reasons on a single constraint edge and, if so, whether we'd be satisfied to use extra nodes for that too.

[mattmccutchen-cci]

I'm linking to John's comment explaining that if this issue is fixed, the current approach to constraining expression variables in unwritable files could lead to noise in the root cause analysis and mentioning some potential solutions. I don't think it's worth filing a separate issue for the expression variables at this time.

[kyleheadley]

Once #708 is merged Constraints will have a new addReason function that saves additional reasons to print out with the root cause as notes rather than warnings, and dumped to json. This was important for conflicting types, but could be used more broadly.

It doesn't affect the content of this issue, but it does add a tool for completing it, if that helps.

[mattmccutchen-cci]

Interesting point. If we are accepting the performance cost (whatever it is) of each edge having a vector of supplemental reasons for conflict reporting, then presumably we would also accept the cost of having a vector of independent primary reasons. However, as you point out, actually changing constrainToWild to accumulate independent primary reasons in a vector would be a separate functionality change.