== Background ==
Previously, QUnit.onError and QUnit.onUnhandledRejection could report
global errors by synthesizing a new test, even after a run has ended.
This is problematic when an errors ocurrs after all modules (and their
hooks) have finished, and the overall test run has ended.
The most immediate problem is that hooks having finished already,
means it is illegal for a new test to start since "after" has already
run. To protect against such illegal calls, the hooks object is
emptied internally, and this new test causes an internal error:
```
TypeError: Cannot read property 'length' of undefined
```
This is not underlying problem though, but rather our internal
safeguard working as intended. The higher-level problem is that there
is no appropiate way to report a late error as a test since the run
has already ended. The `QUnit.done()` callbacks have run, and
the `runEnd` event has been emitted.
== Approach ==
Instead of trying to report (late) errors as a test, only print them
to `console.warn()`, which goes to stderr in Node.js. For the CLI, also
remember that uncaught errors were found and use that to make sure we
don't change exitCode back to zero (e.g. in case we have an uncaught
error after the last test but before our `runEnd` callback is called).
== Changes ==
* Generalise `QUnit.onUnhandledRejection` and re-use it for
`window.onerror` (browser), and uncaught exceptions (CLI).
* Fix broken use of `QUnit.onError` in `process.on( "uncaughtException" )`.
This was passing the wrong parameters. Use the new onUncaughtException
method instead.
* Clarify that `QUnit.onError` is only for `window.onerror`. For now,
keep its strange non-standard signature as-is (with the custom object
parameter), but document this and its return value.
* Remove the unused "..args" from `QUnit.onError`. This was only ever
passed from one of our unit tests to give one extra argument (a
string of "actual"), which then ended up passed as "actual" parameter
to `pushFailure()`. We never used this in the actual onError binding,
so remove this odd variadic construct for now.
* Change `ProcessingQueue#done`, which is in charge of reporting
the "No tests were run" error, to no longer rely on the way that
`QUnit.onError` previously queued a late test.
The first part of this function may run twice (same as before, once
after an empty test run, and one more time after the synthetic
test has finished and the queue is empty again). Change this so that
we no longer assign `finished = true` in that first part. This means
we will still support queueing of this one late test. But, since the
quueue is empty, we do need to call `advance()` manually as otherwise
it'd never get processed.
Previously, `finished = true` was assigned first, which meant that
`QUnit.onError` was adding a test under that condition. But this
worked anyway because `Test#queue` internally had manual advancing
exactly for this use case, which is also where we now emit a
deprecation warning (to become an error in QUnit 3). Note that using
this for anything other than the "No tests run" error was already
unreliable since generally runEnd would have been emitted already.
The "No tests run" test was exactly done from the one sweet spot
where it was (and remains) safe because that threw an error and thus
prevented runEnd from being emitted.
Fixes qunitjs#1377.
Ref qunitjs#1322.
Ref qunitjs#1446.