BLD: Switch CI tests to use pixi for environments

[ColmTalbot]

This is an attempt to remove our need to generate containers in scheduled jobs and have them occasionally just fail because of upstream changes and instead use pixi (conda+pypi+rust) with a lock file.
The lockfile can be modified on a schedule using pull requests so we know updating our test environment won't break things.

[Copilot]

Pull request overview

This PR updates the project’s CI and supporting test/docs tooling to use Pixi-managed environments (with a lockfile) instead of prebuilt container images, aiming to reduce flakiness from upstream container changes and make environment updates more controlled.

Changes:

• Replace container-based GitHub Actions workflows with prefix-dev/setup-pixi and Pixi environments (unit tests, array-backend tests, docs build, latest-dynesty).
• Add Pixi workspace configuration (pixi.toml) and remove container build infrastructure (Dockerfile, conda env template, scheduled container build workflow).
• Adjust tests to work with the new CI setup (ROQ data path discovery + CI ROQ download/cache; reduce JAX dependency in a prior test; increase signal-delay in sampler interrupt test).

Reviewed changes

Copilot reviewed 11 out of 13 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
test/integration/sampler_run_test.py	Increases SIGINT trigger delay to reduce interrupt-test flakiness in CI.
test/gw/source_test.py	Makes ROQ basis directory discovery configurable (incl. env var) rather than hardcoding /roq_basis.
test/core/prior/base_test.py	Removes dependency on jax.numpy for a namespace assertion in a warning test.
pixi.toml	Defines Pixi environments/features for CI, docs, GW deps, array-backends, and dynesty variants.
docs/Makefile	Runs Sphinx via Pixi documentation environment.
containers/environment.yml	Removed legacy conda environment template used for container builds.
containers/Dockerfile	Removed legacy CI container build (including ROQ data baked into image).
.gitignore	Ignores Pixi and venv working directories.
.github/workflows/unit-tests.yml	Switches unit tests / sampler tests / array-backend tests to Pixi; adds ROQ download+cache job.
.github/workflows/pages.yml	Switches docs/notebook execution to Pixi documentation environment.
.github/workflows/latest-dynesty.yml	Switches dynesty-on-master testing to Pixi environment.
.github/workflows/build-containers.yml	Removes scheduled container build/push pipeline.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[lucascolley]

I was randomly browsing the project after noticing that you use array-api-extra, and saw this too — very cool! Feel free to drop me a ping if you ever have questions about using Pixi or array-api-extra :)

[mj-will]

Overall, LGTM. Just a couple of minor comments and questions.

[mj-will]

nessai-bilby can be installed via conda

[ColmTalbot]

This introduces a circular dependency, which made it very unhappy.

$ pixi lock
Error:   × failed to solve the pypi requirements of environment 'documentation' for platform 'osx-arm64'
  ├─▶ failed to resolve pypi dependencies
  ╰─▶ Because there is no version of bilby==2.8.2 and you require bilby==2.8.2, we can conclude that your requirements are unsatisfiable.
  help: The following PyPI packages have been pinned by the conda solve, and this version may be causing a conflict:
        bilby==2.8.2
        See https://pixi.sh/latest/concepts/conda_pypi/#pinned-package-conflicts for more information.

[mj-will]

Weird, I dug into this at it seems that the issue is that nessai-bilby depends on bilby and if one installs that from conda then it pins the versions to bilby==2.8.2. The editable pip install of bilby for the tests then fails since it doesn't provided 2.8.2.

From some digging there may be ways around this but I'm not familiar enough with pixi to know if they make sense. For example:

[workspace.conda-pypi-map]
conda-forge = { mapping = { bilby = false } }

apparently tells pixi to not pass the conda version to pip.

[ColmTalbot]

Can we defer this to a follow-up PR?

[mj-will]

Noting that this also closes #1123 since the base py3.x envs don't have jax in them.

[adivijaykumar]

Thanks @ColmTalbot! I am very much in favour of this.

The lockfile can be modified on a schedule using pull requests

Unless I completely misunderstood what you are suggesting, don't we need a separate workflow yml for this?

[adivijaykumar]

(unrelated, but I'm reading up about pixi, and it could be better than conda for managing the igwn-* environments on the LDG clusters)

[ColmTalbot]

Thanks @ColmTalbot! I am very much in favour of this.

The lockfile can be modified on a schedule using pull requests

Unless I completely misunderstood what you are suggesting, don't we need a separate workflow yml for this?

Yeah, my plan was to implement this after this PR is merged. It would need to run based on an existing lock file on main.

[adivijaykumar]

Looks good to me!