Add more well-formedness checks to the verifier

[zrlk]

We should add opt-in checks for things like

• All vnames should have a corpus set (possibly modulo usrs)
• No anchors are the targets of edges
• Only known names are used (for edges, fact names, fact values)
• ...

Some of these can be done during input ingestion (vnames/values); others can be datalog.

[zrlk]

Another check: no diagnostic nodes emitted. If a diagnostic is emitted, the error report should contain information about it (message, details, anchor location, and so on).

[shahms]

Capturing thoughts from chat:

I think it's reasonable to "warn" on diagnostic nodes which don't match another goal, but I worry that failing on such nodes will just discourage their use. I can certainly imagine a indexer author seeing the node in the graph, emitting a relevant diagnostic, being surprised when confronted with a bunch of unexpected test failures, and changing the diagnostic into a log message instead.

Diagnostics are generally more visible than log messages and, when suitable, preferable.

[salguarnieri]

I was thinking about the same idea (diagnostics sometimes being preferable to logs in the right situations) but then it was brought up that they end up in the graph so there is some cost to emitting them.