Construct

Appears fixed in next build.

Cool.

Well yeah, my long, confusing paragraph before was trying to say that
thing about both suddenly touching when the line is bounding box!

Hello!

Sorry to butt in, but here's your problem:

http://files.getdropbox.com/u/529356/obstangle.cap

Yes, there seems to be a bug. But it appears to be a combination of
per-pixel collision and the angle to which the obstacle is set.

SOMETHING WEIRDER:

Change the obstacle's collisions to Bounding Box. BOTH of the blue boxes
turn transparent for the rotated obstacle! THE MYSTERY DEEPENS

Just an update: I made an entirely new box within my game and collided it
with a 90° upwards wall and it does the same one-pixel thing.

Something's gone wrong in my CAP, but it's not in events.

But this just points to inconcistency in the bug. In your CAP you've just
done exactly the same thing I did, and in mine it somehow touches the wall.
It can't possibly be my code then, can it?

In my second example I cloned the green box to make the blue one that
touches the wall and it shares no events to the green box, it just has the
transparency code. I tried changing the red lines I use to bounding box in
case they created some slightly larger collision box, but then it turns
both transparent which means at THAT point it gets a collision box larger
than the "box". Note that the red line in my example is 90° from its
actual image.

I even tried removing all events in my perpixel.cap except for the one
where the blue boxes test for the wall, and I get the same bug.

Either this has to do with angled sprites, certain images sizes, resized
sprites, just being random, or all of the above. When I remove everything
but that one event making the box transparent, I can't imagine it's my code
in any way.

I've uploaded a new file demonstrating there is no bug. I still think it's
a problem in your original .cap's events. What do you need to modify in my
file to make per-pixel collisions not work?

http://www.konjak.org/perpixel.cap

Here, further proof, and I made this test in seconds.

Two of the same sprite at the same X right up against a wall. If either
overlaps an event says to make it 50% transparent. The one sprite with "per
pixel", though it doesn't visibly overlap, turns transparent!

It's not my code.

But it's fixed by changing to bounding box... ?

As far as I know the per-pixel collisions are working fine. To prove that
it's not working you just need to put two objects in a layout and show that
an 'is overlapping' condition is either falsely running or not running.
Looking at this .cap, I'm inclined to think it's simply a mistake in your
events: either it's jumping more than one pixel a tick and is stopping a
pixel early, or there are rounding errors in floating-point calculations
and the result is off-by-one - either way, something to fix in your events.