Actual result is when leaving "badFilter".)

Can you please provide reproducing scripts that actually give the expected output with the "goodFilter" ?? And fyi, with non-blocking you should use fread() instead of fgets()..

See:

http://capricorn.physics.fsu.edu/~ddm05/working_demo.txt

Or were you asking for something else?

The whole fgets vs fread thing shouldn't matter for this bug, I've came across plenty of "you should use fgets instead of fread when not blocking" statements before I submitted this code. So I did test fread which had no changes with respect to the bug except more overhead to extract separate lines.)

We must have different idea of "working" or this is what you expect:

$ sapi/cli/php -n working_demo.txt
bool(true)
Loop!
xLoop!
x

Now, what should this output?

Yes, that is what I expect to see.

To restate:
I do not expect to see an infinite number of "x"es on any non-block read like you see in "bug_demo". I expect the non-block read to return immediately resulting in another "Loop!" statement. i.e. I expect the output of "bug_demo" with the "badFilter" to be the same as with the "goodFilter" and the number of "x"es to be similar to the number of "Loop!"s.)

Well, your test is inherently buggy. You never return PSFS_PASS_ON from the filter to begin with. And yes, to get this really working you HAVE to use fread(). 

My code was not intended to be "real world", it was intended to accent a bug. However, here's a slightly rewritten bug demonstration:
http://capricorn.physics.fsu.edu/~ddm05/bug_demo2.txt

With non-blocking I/O, I expect to see ~4 "Loop!" statements for each timestamp, since "Loop!" echos every .5 sec and the timestamp echos every 2 seconds.

P.S. The output should be the same as when the filter is removed.)