chore: sync upstream PageBroker staging cap and MIG restore guard
Merged
RS
5d ago+650+650+012345678901234567890123456789-33-33-01234567890123456789
Cherry-picks ai-dynamo/snapshot 67664df (#640: keep
the memory-backed PageBroker staging volume below the agent and PageBroker memory limits) and 13dc5a7
(#318: refuse a checkpoint or restore on a MIG slice). Our 64Gi/64Gi defaults let a large capture
OOM-kill the agent; dende runs 2g.48gb MIG slices.
4c9f7ae addresses the review on top of the unchanged upstream commits: the MIG check fails closed when
nvidia-smi -L cannot list slices; the staging cap parses every Kubernetes quantity form (64e9, Pi, Ei,
milli) and rejects non-quantities instead of leaving the volume uncapped; the README, API reference and
storage.md give the 56Gi default and correct wording. Both fixes are candidates for upstream.
Unit and chart tests pass (74/74).