A preview builds on the droplet runner, as the release does: vxrail's push of the ~450 MB layer dies at the registry ingress (499) #58

The preview job ran on runs-on: isolated. Two runners serve that label, and the job failed every time vxrail picked it up.

Evidence

• Every failed image push ran on plat-standby-forgejo-runner-vxrail: runs 900, 906, 919, 925, 938, 949, 951, 953, 955, 968, 990, 1024 and 1026.
• Each one fails the same way: one monolithic PUT https://git.open-platform.sh/v2/trevato-apps/forge/blobs/uploads/<id>?digest=sha256:c7604faa… returns 499 Client Closed Request. Run 968 also got a 500 at 63.94 s before a retry.
• Each attempt is cut off at about 63 s. Buildkit then retries, so "pushing layers" comes to 62.4 s, 126 s, 189.6 s or 256.9 s.
• c7604faa looks like vxrail's cached apt/chromium layer (about 450 MB). The whole image is 718 MiB.
• The green previews (947 #46, 980 #49) ran on plat-standby-forgejo-runner-isolated and pushed in 0.5 to 1.3 s.
• The release build hit the same 499 on 2026-10-15 and moved to runs-on: droplet in 8e2125c4. This PR does the same for the preview.

Lasting infra alternative (platform side): take the isolated label off the vxrail runner, or raise the registry ingress timeout (about 60 s now) or its body handling for blob PUTs.