chore: switch Dockerfile to slim Python + uv lockfile

[cailmdaley]

Summary

Replaces the skaha/astroml base (which brings conda → double-installs every Python dep, ~5–10 GB image) with python:3.12-slim-bookworm + uv driven by a committed lockfile.

	Before	After
Base	images.canfar.net/skaha/astroml:latest (conda Python)	python:3.12-slim-bookworm
Deps	hand-pinned in Dockerfile, drifts from pyproject.toml	pyproject.toml is SSOT (what), uv.lock is the pinned manifest (which versions)
WeightWatcher	built from source with patched headers	Debian's weightwatcher package
CI	builds image, runs shapepipe_run -c …	also invokes source-extractor --version, weightwatcher --version, psfex --version

Why

Two problems with the v2.0 image:

1. Conda + pip double install: skaha/astroml ships conda; we then pip install everything on top. Two copies of every scientific package, multi-GB bloat, slow builds.
2. Runtime regressions slip through CI: the original failure that prompted this PR (a tool that builds inside the image but can't actually run on canfar) wasn't caught because the deploy workflow only invoked shapepipe_run.

Lockfile workflow

# regenerate after editing pyproject.toml dependencies
uv lock

# add a new dep
uv add 'somepackage>=1.2'    # updates pyproject and uv.lock

# bump pinned versions (deliberate)
uv lock --upgrade

uv sync --frozen in the Dockerfile fails the build if pyproject.toml and uv.lock are out of sync — so a stale lockfile cannot ship.

Test plan

• CI image build succeeds on this branch
• Test — binaries step passes (sextractor, weightwatcher, psfex)
• Test — shapepipe entry point passes
• Pull built image locally on canfar and run a real ShapePipe job (Cail to do, post-CI green)

Out of scope / follow-ups

• Optional: separate Dockerfile.canfar building on skaha/astroml if there's a deployment reason to keep that base image around (Martin mentioned this could be cleaner — currently conjectural).
• Dockerfile.jupyter is unchanged; it inherits from this base.

🤖 Generated with Claude Code

[cailmdaley]

@martinkilbinger ready for your eyes when you have a moment. Two things worth flagging:

Image works end-to-end on the CI runner. Build is green in 3m31s (uv installs 238 packages in 322ms — everything's a prebuilt wheel). The added Test — binaries step runs source-extractor --version, weightwatcher --version, psfex --version inside the built image and they all answer. shapepipe_run -c /app/example/config.ini also boots cleanly. Since that's the same failure mode you hit on canfar (a binary that the image carries but couldn't run), this should work on canfar too — but it'd be good to pull the image and try a real job to confirm before we merge.

Modernized package versions in the same PR (commit b62e9dbb). The lockfile now resolves to numpy 2.4, astropy 7.2, pandas 3.0, galsim 2.8, mpi4py 4.1, etc. (everything at current latest). Pyproject minimums bumped to the new major versions. The only line that's still pinned the old way is ngmix — that stays on Axel's stable_version branch until we do the upstream-ngmix migration. Worth knowing in case you hit a numpy-2 or pandas-3 compatibility surprise in real jobs; rolling back is just uv lock with a constraint.

(Disclosure: PR drafted by Claude in collaboration with me; I haven't run a full pipeline through the new image yet — that's the canfar test I want to do.)

[cailmdaley]

@martinkilbinger one more thing worth flagging since it explains a Dockerfile shape change you'll notice — why we dropped the WeightWatcher source build:

The previous Dockerfile (v1.x) built WeightWatcher from source with six `sed` patches to `prefs.h`, `globals.h`, `fitscat.h`, `prefs.c`, `main.c`, and `fitscat.c`. The new Dockerfile just `apt install`s `weightwatcher` instead.

The patches were converting tentative definitions like `prefstruct prefs;` in headers into proper `extern` declarations + a single definition in the corresponding `.c` file. They were a workaround for GCC 10 (April 2020) flipping its default to `-fno-common`, which broke WeightWatcher 1.12's (2014) use of common-symbol globals — every translation unit started getting its own definition, the linker complained.

Debian's `weightwatcher 1.12+dfsg-3` package on bookworm carries an equivalent patch upstream of the build (the `+dfsg-3` suffix indicates Debian-applied patches). So `apt install` gives us the same WeightWatcher 1.12 binary, just with Debian carrying the GCC-compatibility fix instead of us doing it inline.

Tradeoff for completeness:

	Source build (old)	apt (new)
Version control	We pin `WW_VERSION=1.12`	Tied to bookworm
Patch maintenance	Ours, in plain sight	Debian's, transparent
Build time	~30s + autotools deps	~1s
Risk if GCC changes again	We patch	Debian handles it

The apt route is probably the better default — Debian has stronger Astromatic build infrastructure than we do. If we ever needed a version bookworm doesn't ship, we'd swap back to source-build.

[martinkilbinger]

When running the tests from within the docker of this branch, I get errors:
For uv sync --extra test:
error: failed to remove file /app/.venv/lib/python3.12/site-packages/appdirs-1.4.4.dist-info/INSTALLER: Read-only file system (os error 30)
And
uv run pytest:
error: Failed to spawn: pytest Caused by: No such file or directory (os error 2)

And the slim bookworm image does not have some useful programs, e.g. vi, which should be added.

[martinkilbinger]

When running the tests from within the docker of this branch, I get errors:
For uv sync --extra test:
error: failed to remove file /app/.venv/lib/python3.12/site-packages/appdirs-1.4.4.dist-info/INSTALLER: Read-only file system (os error 30)
And
uv run pytest:
error: Failed to spawn: pytest Caused by: No such file or directory (os error 2)

And the slim bookworm image does not have some useful programs, e.g. vi, which should be added.

And cloning the branch and using that version, I also get an error message for `uv run pytest:
WARNING: Failed to generate report: No data to report.

[martinkilbinger]

Some of the issues are fixed, but I think /app is still a read-only file system and I cannot run the pytests.

I use apptainer pull shapepipe_uv.sif docker://ghcr.io/cosmostat/shapepipe:chore-uv-lockfile to get the docker image, then apptainer shell shapepipe_uv.sif and cd /app; uv run pytest.