plat 545

PLAT-545: old release copies were never pruned and filled the RTS disk

State: fixed on main (scripts), not yet run on RTS. P1 while the disk is 99% full.

Found: 2026-10-05, RTS deploy exited 1 with gateway: initialize project SQL tables: database or disk is full. The root disk was 99% full (95 of 96 GB): 27 release copies, 34 GB, under /var/lib/video-studio/video-studio/releases, plus 12 GB of Go build cache and 5.7 GB of npm cache. The failed deploy left the previous release serving.

Cause: the shared pruner (deploy/common/prune-releases.py) aborted with release cleanup skipped: Permission denied: '/proc/<pid>/root' since the slot accounts were introduced (it resolves paths found in process data, and another account's /proc/<pid>/root cannot be resolved). The deploy only prints that as a warning, so nothing was pruned after any deploy.

Fix:

• The pruner skips an unresolvable path instead of aborting, and always keeps the newest releases (--keep-newest, default 2: the live one and the one before it) on top of the active one and anything a readable process still uses, because other accounts' processes cannot be fully inspected.
• New --only-if-free-below-gb N: a deploy first makes room, only when space is short (15 GB), before it delivers or builds anything. Wired into all three deployers: RTS (deploy.sh), the rootless-Linux products (Confida, Excellence, SparkQuill; deploy/rootless-linux/build-and-activate.sh) and Dominion (deploy-dominion.sh). A failure here only warns.
• One test pins the abort and the newest-kept rule.

Left: run the next RTS deploy (it prunes first; about 28 GB come back); check the other servers' disk and releases (Excellence, Confida, SparkQuill, Dominion); decide on the Go build cache (12 GB) and npm cache (5.7 GB) on RTS.