ci-build: clean up orphaned per-build upload staging dirs

[jmandel]

Problem

publish / publish-ig receive each build into a staging dir
~/uploading/<RANDOM>[_org_repo] and only mv it into www/ on success. If a
run is interrupted before that final mv — Spot preemption of the build node, a
ci-build pod restart, a failed tar, or an mv error — the staging dir is left
stranded at the root of ~/uploading, and nothing ever reclaims it (the
reindex autoclean only operates inside www/ig).

In production these have accumulated into hundreds of dirs / hundreds of GB on
the ci-build persistent disk (observed ~763 G used, 81% full), dominated by
frequently-built IGs (e.g. many *_HL7_fhir-sdoh-clinicalcare). www/ is the
served site and is fine; everything else at the root of ~/uploading is orphaned
build cruft.

Fix

• publish / publish-ig — the EXIT trap (which previously only killed the
  keepalive) now also removes the run's own staging dir and any half-rotated .old.
  Guarded with [ -n "$TARGET" ] so an unset variable can never expand to
  rm -rf ~/uploading. It's a no-op on the success path (the dir is already moved).
• reindex — adds an hourly GC sweep of ~/uploading/* (excluding www/ and
  lost+found) older than GC_AFTER_DAYS (default 2), as a backstop for runs
  that are SIGKILLed past the trap.

Safety

• The served site (www/) is never touched.
• -ctime +2 leaves in-flight uploads alone; -ctime matches the conventions
  already used by the reindex autoclean and publish branch-cleanup.
• The [ -n "$TARGET" ] guard prevents any rm -rf ~/uploading footgun.

Deploy notes

These scripts are baked into the ci-build image, so the fix takes effect after a
rebuild + redeploy. On first run the hourly GC will clear the existing backlog
(everything >2 days old) automatically — no manual cleanup required, though the
disk can be reclaimed sooner with a one-off
find ~/uploading -maxdepth 1 -mindepth 1 -type d ! -name www ! -name 'lost+found' -ctime +2 -exec rm -rf {} \;.

🤖 Generated with Claude Code