[azlekov]

Problem

prune_caches ran lru_cap "$runner/bun", which LRU-evicts single files by atime across the whole /root/.bun volume. bun treats install/cache/<pkg>@<ver>@@@1/ as complete whenever the folder exists, so an evicted file silently goes missing from every later node_modules.

0-byte files go first, because copying them never reads them and their atime never moves. On the hetzner-cx43-runner-* runners (cap 8 GiB, pruner logged evicting least recently used files from /caches/3/bun on 10-16 and 10-17), next@16.3.8 lost dist/compiled/server-only/empty.js. After that, every Next build failed with Module not found: Can't resolve 'server-only' (dodi-smart/nis-suite#142). The same eviction can also delete /root/.bun/bin/bun.

Fix

• New lru_cap_bun, modelled on lru_cap_mise_installs. It evicts whole package folders, oldest package.json atime first (dir mtime when there is no package.json), until the whole volume fits under the cap (excess_bytes on the volume). This covers both cache/<pkg>@<ver>@@@1 and scoped cache/@scope/<pkg>@<ver>@@@1. It also removes the matching cache/<pkg>/<ver>@@@1 symlink and any empty name/scope dirs.
• It never touches bin/ or the @t@ manifest piles. It clears *.tmp / *.part leftovers and cache/.tmp/* entries older than 60 min; the age guard keeps it away from installs that are still running.
• prune_caches now calls lru_cap_bun "$runner/bun".
• hk stays on lru_cap: its volume holds configs/<hash>.json and pklr/packages/<hash>.{zip,url}. Each of these is an independent whole file that gets regenerated or re-downloaded when it is missing. There are no per-tool/version trees, and the volume is ~1 MB, far below the cap.
• pruner.sh returns early when it is sourced, so tests can call its functions.
• Updated the header comment and the README "Nightly" section.

Test

tests/pruner-bun.sh (plain bash, needs GNU find/stat/du) builds a fake bun volume. It holds packages with 0-byte files that have very old atimes, a scoped package, the version symlinks, @t@ piles, bin/bun/bunx, and stale and live temp entries. It then runs prune_caches with a cap that requires dropping exactly the three oldest packages. A new scripts job in the PR build runs shellcheck and tests/*.sh, and the gate job now requires it.

$ docker run --rm -v "$PWD":/w:ro ubuntu:24.04 bash /w/tests/pruner-bun.sh
[pruner] ... /tmp/.../caches/1/bun is 1002394B (cap 744173B)
[pruner] ... evicting least recently used packages from /tmp/.../caches/1/bun/install/cache
[pruner] ... /tmp/.../caches/1/bun/install/cache freed ~300039B
PASS: bun cache evicted by whole package (volume 702228B)

Against main's pruner.sh, the same test fails 16 checks, among them next lost server-only/empty.js and bin/bun removed. shellcheck -x pruner.sh tests/pruner-bun.sh is clean.

Separately, CACHE_CAP_BYTES on the cx43 host was raised from 8 GiB to 12 GiB (62 GB free on that disk; done outside this PR).

🤖 Generated with Claude Code