other build of same version changed tools behavior (metilene)

[koriege]

Hi,

I stumbled across a wired issue and wanted to ask for your wisdom.
At the time of writing, the latest build for the tool metilene I am referring to is h7b50bb2_5 (https://anaconda.org/bioconda/metilene/files). And since h7b50bb2_4 its output heavily differs from all previous bioconda builds, although the version/source has never changed since 2018 (http://www.bioinf.uni-leipzig.de/Software/metilene/metilene_v02-8.tar.gz). I also see no indication for the behavioral change in the build.sh or meta.yaml file of the recipe. I compared a locally compiled version 0.2.8, with conda/mamba installed 0.2.8=hec16e2b_2, 0.2.8=h7b50bb2_5, 0.2.8=h7b50bb2_4 as well as conda build versions of the aforementioned hashes. Only 0.2.8=hec16e2b_2 and below returned the same results as the locally compiled version.
Is someone able to reproduce and explain this to me? Thank you!

command to reproduce: metilene -X 8 -Y 8 -t <threads> -m 8 -M 300 -a 2M -b 30M rates.tsv

input data: rates.tsv

[holtjma]

@koriege I ran into a rogue PR that changed an unrelated tool, see #59640

The same user has submitted PRs that change the tool you are referring to, see:
#57875

And a follow-up change here:
#59579

Sounds like you may be having a similar experience.

[koriege]

Dear @holtjma thanks for looking into this!
However, whereas in #59640, on good intention I guess, a main.cpp was injected into pb-StarPhase in order to squeeze in a usage upon calling the tool without CLI parameters, I don't see any modifications to the metilene source code. Also the builds I am referring to in the OP are not non-functional. I appears, that the output table is different (truncated, with different p-values). This is even true for your rebuild 0.2.8=h7b50bb2_6. The latest working build, that returns identical output compared to the manually compiled original source is from 2022 (0.2.8=hec16e2b_2). Later builds of version 0.2.8 are consistent in their output (h031d066_3 (2023) h7b50bb2_4 (2024) h7b50bb2_5 (2025) h7b50bb2_6 (2025)) but not to any previous 0.2.8 build, nor to the manually compiled source. In terms of reproducibility this is a problem.

[koriege]

fyi. I found the culprit in the source. Only GCC v12+ brought it to light. PR follows asap