ci: persist cargo caches on runner volume

[wavekat-eason]

What

Point CARGO_HOME and CARGO_TARGET_DIR at the runner's persistent volume (/home/runner/runner) in the GitHub Actions runner image, instead of the image layer (CARGO_HOME=/usr/local/cargo) and the default ./target inside the checkout.

Why

Workflows run actions/checkout, which git clean -ffdxs the checkout's target/ every run. So Rust CI had no warm on-disk artifacts and relied on Swatinem/rust-cache, which tars ~370 MB and round-trips it to GitHub's cloud cache on every run — minutes of upload at a few MB/s on these self-hosted runners (and the wavekat-voice parallel-jobs split doubled it to two caches).

With cargo's caches as siblings of _work on the same persistent volume, compiled artifacts (incl. the 5–10 min ONNX/sherpa build) and the crate registry survive both the git-clean and container recreation. That lets wavekat-voice drop rust-cache (companion PR), eliminating the upload entirely.

The toolchain binaries stay baked at /usr/local/cargo/bin (still first on PATH) and rustup resolves toolchains via RUSTUP_HOME, so only the registry/git cache and compiled artifacts relocate.

Rollout (volume-preserving — keeps registration)

On the runner host (aoc-m3l):

cd wavekat.com && git pull
sudo docker build -t wavekat/gha-runner:latest scripts/docker
sudo systemctl restart gha-runner@{1..4}

systemctl restart re-execs docker run --rm … wavekat/gha-runner:latest with the rebuilt image, so the new ENV applies while the gha-runner-N volume (and the runner's registration) persists. Do not re-run setup-gha-runners-docker.sh for this — it wipes volumes.

First CI run per runner after the restart is cold (fresh cargo-home/cargo-target on the volume); every run after is warm with zero cache upload.

Sequencing

Deploy this before merging the wavekat-voice cache: false PR — otherwise that PR's runs go fully cold every run (no on-disk target, no cloud cache).

Note

The entrypoint's chown -R runner:runner /home/runner/runner now also recurses over cargo-target on container (re)start — a one-time-ish cost on restart, not per job. Can be narrowed later if it bites.

🤖 Generated with Claude Code

[github-actions]

🔗 Preview: https://ci-persistent-cargo-target.wavekat-com.pages.dev