ci: free disk space before embeddings-all builds to stop intermittent "No space left on device"

[mosuka]

Summary

The ubuntu-latest x86_64 CI jobs that build --features embeddings-all (which pulls in candle-core / candle-transformers / tokenizers / hf-hub / image / reqwest) intermittently die with:

System.IO.IOException: No space left on device : '/home/runner/actions-runner/cached/.../Worker_...log'

The "Run test" step never completes (no failed step is recorded; the job log truncates right after the cache-restore step). The cause is only visible via the check-run annotations:

gh api repos/mosuka/laurus/check-runs/<job_id>/annotations

Observed on PRs #772 and #774 (the Test laurus (ubuntu-latest, x86_64-unknown-linux-gnu, stable) job failed at ~4m25s twice in a row before a later re-run happened to land on a runner with enough free space). The ubuntu-24.04-arm job is unaffected because it sets skip_embedding_features: true; Windows is unaffected (different disk layout).

This is an environment/disk-margin problem, not a code or test problem. The restored Rust build cache (~790 MB compressed, much larger expanded) plus the embeddings-all build artifacts push the runner's root disk over its limit. It is intermittent because the margin is razor-thin.

Root cause

GitHub-hosted ubuntu-latest runners ship with several large preinstalled toolchains the build does not use (.NET, Android SDK, GHC/Haskell, CodeQL bundle, cached Docker images). None of the workflows free any of it, so the heavy embeddings-all compile occasionally runs out of space.

Proposed fix

Add a reusable composite action .github/actions/free-disk-space that removes the unused preinstalled toolchains (reclaims ~25-30 GB) and prints df -h / before/after, then reference it (gated if: runner.os == 'Linux') right after checkout in the heavy embeddings-all jobs:

• regression.yml (PR-blocking) — clippy, test-laurus, test-server, test-mcp, test-cli, test-python, test-nodejs, test-ruby, test-php
• periodic.yml and release.yml — the equivalent clippy / test-* jobs

A composite action keeps the removal list in one place instead of pasting it into ~25 jobs. Because pull_request runs use the PR's own workflow definitions, the change self-validates on its own CI run.

The removal targets only safe, unused paths (/usr/share/dotnet, /usr/local/lib/android, /opt/ghc, /usr/local/.ghcup, /opt/hostedtoolcache/CodeQL, dangling Docker images) — it does NOT touch the Rust / Python / Node toolchains the jobs rely on.

Acceptance criteria

• .github/actions/free-disk-space/action.yml composite action added.
• Referenced (Linux-gated, after checkout) in the heavy embeddings-all jobs of regression.yml, periodic.yml, release.yml.
• The action prints free space before/after so future disk regressions are diagnosable from the log.
• CI is green on the fix PR (validates the action runs and the builds still pass).
• actionlint (if available) reports no errors on the edited workflows.

Out of scope (possible follow-up)

• release.yml build-* / publish-* artifact jobs (release-gated, not currently failing).

References

• Failing job (annotation shows the disk error): https://github.com/mosuka/laurus/actions/runs/26875741104/job/79262647643