No source change here. The phone's hpp is the file:../ sibling vendored into worklet.bundle at pack time, so hyper-proto-peer 0.3.119 — which retries a byte fetch that failed instead of treating it as done — reaches the handset only in a new favela-android build. The servers picked 0.3.119 up on their own (release-ready respawn, lastExitCode 75 on all three) and never had the fault: they author locally, so record bytes exist before the firehose frame that references them. The phone is the node that RECEIVES a co-owned record, hits the ~1s gap between pointer and bytes, and was the one measured losing a post for 13 minutes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>