Workaround overflow errors with Numpy ≥ 2.

[emollier]

This change works around OverflowError regressions caught by the test suite with Numpy 2 and later, independently reported in issue #127 and in Debian bug #1095090, by restoring the former (potentially buggy) behavior with uncaught overflows. Issues manifested like:

  File "/home/shuaiw/miniconda3/envs/methy/lib/python3.12/site-packages/pbcore/io/align/_BamSupport.py", line 66, in rgAsInt
    return np.int32(int(rgIdString.split("/")[0], 16))
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OverflowError: Python integer 3367457666 out of bounds for int32

and like:

    def _gapify(self, data, orientation, gapOp):
        if self.isUnmapped:
            return data

        # Precondition: data must already be *in* the specified orientation
        if data.dtype == np.int8:
            gapCode = ord("-")
        else:
>           gapCode = data.dtype.type(-1)
E           OverflowError: Python integer -1 out of bounds for uint8

While being at it, the change includes a fix proposal for situations where identifiers include hyphenations, fixing:

    return np.int32(int(rgIdString.split("/")[0], 16))
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: invalid literal for int() with base 16: 'cb4d472d-100C60F6'

[natechols]

Thanks for taking a stab at fixing this - I'll investigate it on our end. The identifier bug is fixed upstream, we're just slow to update this mirror. hold on wrong bug, this one is new. I'll take a look at that too.

[natechols]

@emollier Actually I'm a little surprised that you were able to trigger this bug, since setup.py is pinned to a very narrow range of numpy versions (to keep our builds stable). Is there some packaging or installation step in your setup that is allowing the module dependencies to be ignored?

[emollier]

Hi Nat, Thanks for the status about upcoming fixes! I must admit I was reluctant to send the patch at all given the nature of my fix, and I'm glad there exists something probably more appropriate in the pipeline. About how I hit the bug despite the numpy version specification: this was initially Debian bug #1095090[1] affecting the pbcore module. Since we ship the operating system as a consistent whole, there is only ever one version of a given Python module that is available under the form of a Debian package, and that includes numpy. Thus, when we moved to numpy 2 in unstable, this triggered a campaign of porting all existing packaged Python modules depending on numpy to make them compatible with the newer version that introduced breaking changes. At no point in the packaging step is involved pip, which would have otherwise handled the multiple modules versions. In hope this clarifies things, Have a nice day, :) Étienne. [1]: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1095090

[emollier]

Oh, nevermind that:

Thanks for the status about upcoming fixes! I must admit I was reluctant to send the patch at all given the nature of my fix, and I'm glad there exists something probably more appropriate in the pipeline.

I just saw your edited comment about the fix.

[natechols]

I don't think the modulus will give the intended result, which in the case of overflow will be a negative number. Based on my quick experimenting with one of the failing unit tests, if it overflows I think you need to take the modulus and then subtract np.iinfo(np.int32).max+1 again.

[emollier]

I had to scratch my head, but I think I see your point. Legacy behavior is: the overflow wraps around at np.iinfo(np.int32).min. Meanwhile my implementation is: wrap around at 0 straight. To be fully legacy compatible, my modulus wraparound is incorrect and I need to compensate so the index wraps between np.iinfo(np.int32).min and np.iinfo(np.int32).max instead of 0 to np.iinfo(np.int32).max. Do I understand correctly?

[emollier]

I have pushed a commit, this may clear things. Looks like I botched my first attempt. I pushed a hotfix but am still running tests.

[emollier]

Sorry for my mess, the commit should be in the state I intend now. I have verified the test suite passed.