Invalidated diagnostics live forever

[brson]

Updating source follows this pattern:

            self.db.update_file(self.filename, module_source.text);

            let diagnostics = self.db.diagnostics(self.filename);
            for diagnostic in &diagnostics {
                dada_error_format::print_diagnostic(self.db, &diagnostic)?;
            }

These diagnostics are "accumulated" by salsa. From what can tell these diagnostics stick around forever. If the source text changes then the diagnostics remain, but they reference text that no longer exists.

I have run into this in the repl, which continuously updates the source text.
At the moment the repl doesn't "commit" steps that fail to compile or run. That is, if enter foo() and foo doesn't exist, the repl will update the source to include a function that tries to call foo(), see that it fails and "revert" the changes so that call is no longer part of the source text. The result is that subsequent steps continue to observe that failed diagnostic, but it prints incorrectly:

>>> foo()
Error: can't find anything named `foo`
   ╭─[<repl-input>:2:17]
   │
 2 │ __repl_result = foo()
   ·                 ─┬─
   ·                  ╰─── here
───╯
compilation failed
>>> 4
Error: can't find anything named `foo`
   ╭─[<repl-input>:2:17]
   │
 2 │ ╭─▶ __repl_result = 4
 3 │ ├─▶ __repl_expr_1(__repl_result).await
   · │
   · ╰──────────────────────────────────────── here
───╯
compilation failed

Here the second step produces the same diagnostic but it is now pointing to different text.

I think I can mostly work around this behavior by ignoring previously-seen diagnostics, but this seems like a bug - the fixed diagnostic should disappear or otherwise be identifiable as invalidated.

[nikomatsakis]

Yeah, the salsa branch we are based off of has some bugs =)

I've been putting off fixing them because I wanted to make progress on dada but ...