tomnomnom__gron task and task_cleanroom images use different Go toolchains

[kehao95]

Hi, I noticed a toolchain mismatch between the official images for
tomnomnom__gron.88a6234.

After pulling the current linux/amd64 images:

docker pull --platform linux/amd64 programbench/tomnomnom_1776_gron.88a6234:task
docker pull --platform linux/amd64 programbench/tomnomnom_1776_gron.88a6234:task_cleanroom

the two images report different Go versions:

docker run --rm --platform linux/amd64 \
  programbench/tomnomnom_1776_gron.88a6234:task_cleanroom \
  sh -lc 'go version'

docker run --rm --platform linux/amd64 \
  programbench/tomnomnom_1776_gron.88a6234:task \
  sh -lc 'go version'

Observed:

task_cleanroom: go version go1.21.0 linux/amd64
task:           go version go1.24.2 linux/amd64

Since inference uses :task_cleanroom while programbench eval defaults to
:task, this means a Go submission can be developed/compiled against one Go
toolchain but evaluated against another.

Is this toolchain difference intentional for this instance?

[klieret]

Hi, thanks for raising this issue. This isn't great, I agree. Pretty confident we'll close this loophole fast, I guess we could either change the images, but maybe it makes more sense to long term switch eval over to the task_cleanroom containers, which might mostly be a bit of a challenge with the gold validation (though I guess we could still run gold validation on task instead).

gold validation = checking that original repo passes all tests