Fix flaky test

[pepeiborra]

"add missing modules (non workspace)" became flaky after the recent changes to snapshot VFS data.

The test calls createDoc to open a module B imports A where A does not exist, then waits for the missing module error and calls createDoc to open a module A. Importantly, both B and A live outside the workspace and thus are not subject to FileWatched notifications.

Unfortunately, the test wasn't testing the right thing because createDoc sends a FileWatched notification right before sending a DidOpen notification! This was causing a race condition, where the FileWatched notification triggers a build that may or may not run the FileExists ModuleA key:

• If it does, then the key will resolve to False since ModuleA exists neither in the file system nor in VFS, and the test will fail since the key will not get rebuilt again (nothing dirties it).
• If the build gets aborted before it does, then the key will get computed in the next build (for the DidOpen notification) and will return true because the file now exists in VFS.

I imagine that the test was passing before since the VFS state would get updated concurrently, and the race condition would not get triggered in practice.

To fix the test, we simply tell lsp-test not to send FileWatched notifications.

I also found an old comment stating that "VFS existence is not enough, a file must exist in disk". This is clearly not true anymore, so I have deleted it.

Fixes #2831

[wz1000]

Thanks for looking into this! I ran the testsuite a couple of times but could not reproduce this failure locally at all.

[michaelpj]

To fix the test, we simply tell lsp-test not to send FileWatched notifications.

I'm confused, that doesn't look like what the change does? And I don't see the comment you mentioned being deleted?

[pepeiborra]

I'm confused, that doesn't look like what the change does? And I don't see the comment you mentioned being deleted?

Fixed now, thanks for noticing this. I must have forgotten to git push -f