fix(ci/runner): the disk guard had been dead code, and polled too slowly to matter

arch has been failing with `as: BFD (GNU Binutils) 2.47 assertion fail`, which
reads like a toolchain regression and is not one. The line above it is `can't
write 10 bytes to section .text._ZN6Vulkan...: 'No space left on device'` — the
assembler handling ENOSPC badly. Both recent arch failures are the runner filling
its disk, and by the time anyone looks, df reports 37% used.

Three things were wrong with the hygiene that was supposed to prevent this.

The cache cap and the burst-clear were dead code. They looked up the runner as
`docker ps -f name=gitea-runner-runner`, which matches zero containers now that
the replicas are `gitea-runner-fleet-runner-N-1`, so $RUNNER was always empty and
both branches were skipped. The store also moved: the fleet runs a standalone
cache-server bind-mounting a HOST directory, so no docker exec is needed at all.

The routine prune reclaimed nothing. `--filter until=6h` on a runner that rebuilds
its CI images every push means every image is younger than the window — measured:
0B reclaimed while docker system df reported 22.76 GB reclaimable. Now until=2h.

The burst guard never fired. It polled every 30 minutes for >=80% used, but three
concurrent Rust builds fill the disk and drain it again well inside that window,
so the poll kept landing on a healthy df. Now every 10 minutes, and it triggers on
a free-space FLOOR too — 80% of 123 G still leaves only ~25 G, which three jobs
swallow before the next poll.

Deployed to home-runner-1 and exercised: shellcheck clean, timer active on the new
interval, one run reclaimed 551 MB. Honest limit: this improves the odds, it does
not fix the cause. The remaining 22 GB of idle images are the fedora-rpm bases the
next run wants back, so there is no free headroom to reclaim — three replicas
building this workspace share one 123 G disk. The lever is capacity (grow the LXC)
or concurrency (drop to two replicas), and that is a judgement call, not a script.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

scripts/ci/docker-prune.sh

@@ -9,34 +9,48 @@
 set -u
 export PATH=/usr/bin:/bin:/usr/local/bin:$PATH
 
-RUNNER=$(docker ps -q -f name=gitea-runner-runner | head -1)
-ACTCACHE=/root/.cache/actcache/cache    # path INSIDE the runner container (HOME=/root there)
-CAP_MB=20000                            # clear the actcache once its blob dir exceeds ~20 GB
-BURST_PCT=80                            # full clear once the disk is this % full
+# The cache-server's blob store is a HOST directory: the fleet runs a standalone cache-server
+# service (compose.yml) that bind-mounts this path to /data, and every replica points at it over
+# HTTP (`external_server`). It used to live inside a runner container's writable layer, which is
+# why this reached in with `docker exec` — that is no longer where it is, and the container name it
+# looked for (`gitea-runner-runner`) does not exist either now that the replicas are named
+# `gitea-runner-fleet-runner-N-1`. Both halves silently did nothing: the filter matched zero
+# containers, so the cap and the burst-clear below were dead code. A plain host path needs neither.
+CACHE_DIR=${CACHE_DIR:-/home/runner/gitea-runner-fleet/cache}
+CAP_MB=${CAP_MB:-20000}                 # clear the cache store once it exceeds ~20 GB
+BURST_PCT=${BURST_PCT:-80}              # full clear once the disk is this % full
+MIN_FREE_GB=${MIN_FREE_GB:-45}          # ...or this little is left, whichever trips first
 
 # 1) Routine: trim aged images / build cache / stopped containers. sha-<commit> tags aren't
-#    dangling, so -a is required; until=6h keeps very recent ones for quick re-runs.
-docker image prune     -af --filter until=6h || true
-docker builder prune   -af --filter until=6h || true
-docker buildx prune    -af --filter until=6h || true
-docker container prune  -f --filter until=6h || true
+#    dangling, so -a is required. until=2h, not 6h: on a busy day every image is younger than six
+#    hours, so the filter matched nothing and a run reclaimed 0B while `docker system df` was
+#    reporting 20+ GB reclaimable. Two hours still protects a re-run of the push being worked on.
+docker image prune     -af --filter until=2h || true
+docker builder prune   -af --filter until=2h || true
+docker buildx prune    -af --filter until=2h || true
+docker container prune  -f --filter until=2h || true
 
-# 2) Cap the act_runner cache server store (the real disk filler). Clearing the blobs is safe —
-#    act_runner repopulates it and cache keys are content-hashed, so this only drops stale entries.
-if [ -n "$RUNNER" ]; then
-  SZ=$(docker exec "$RUNNER" du -sm "$ACTCACHE" 2>/dev/null | cut -f1)
+# 2) Cap the cache-server store. Clearing the blobs is safe — act_runner repopulates it and cache
+#    keys are content-hashed, so this only drops stale entries.
+if [ -d "$CACHE_DIR" ]; then
+  SZ=$(du -sm "$CACHE_DIR" 2>/dev/null | cut -f1)
   if [ -n "${SZ:-}" ] && [ "$SZ" -ge "$CAP_MB" ]; then
-    docker exec "$RUNNER" sh -c "rm -rf $ACTCACHE/*" && echo "actcache cleared (was ${SZ} MB)"
+    rm -rf "${CACHE_DIR:?}"/* && echo "cache-server store cleared (was ${SZ} MB)"
   fi
 fi
 
-# 3) Burst guard: a push-storm can fill the disk within one interval. Once >=BURST_PCT% full, prune
-#    ALL idle images/cache AND clear the actcache, regardless of age. In-use images are protected.
+# 3) Burst guard: a push-storm fills the disk WITHIN one interval — three concurrent Rust builds,
+#    each with a multi-GB target/, on top of a ~40 GB containerd image baseline. Trigger on a free
+#    -space FLOOR as well as a percentage: the percentage alone is the wrong instrument here,
+#    because 80% of 123 G still leaves only ~25 G, which three jobs can swallow before the next
+#    poll. In-use images are protected by the daemon, so this cannot pull the rug from a live job.
 PCT=$(df --output=pcent / | tr -dc '0-9')
-if [ -n "$PCT" ] && [ "$PCT" -ge "$BURST_PCT" ]; then
-  echo "disk ${PCT}% >= ${BURST_PCT}% — burst clear"
+FREE_GB=$(df --output=avail -BG / | tr -dc '0-9')
+if { [ -n "$PCT" ] && [ "$PCT" -ge "$BURST_PCT" ]; } ||
+   { [ -n "$FREE_GB" ] && [ "$FREE_GB" -lt "$MIN_FREE_GB" ]; }; then
+  echo "disk ${PCT}% used, ${FREE_GB}G free (thresholds ${BURST_PCT}% / ${MIN_FREE_GB}G) — burst clear"
   docker image prune -af   || true
   docker builder prune -af || true
   docker buildx prune -af  || true
-  [ -n "$RUNNER" ] && docker exec "$RUNNER" sh -c "rm -rf $ACTCACHE/*" || true
+  [ -d "$CACHE_DIR" ] && rm -rf "${CACHE_DIR:?}"/* || true
 fi