Shape agreed so far: - A dedicated co-hosted artifacts repo (e.g. bq2-artifacts.fossil) served beside the project repos — NOT uv in the primary bq2 repo (keeps primaries lean; contention isolated to artifact operations). - Fossil unversioned storage (uv) as the mechanism: digest-addressed, versioned, HTTP-served, replicates with `fossil sync --uv` (aarch64 backup node and other mirrors pick artifacts up automatically). - MEASURED baseline on this host: `fossil uv add` of a 500MB incompressible file = 16.7s local (mostly zlib attempting to compress incompressible data) — seconds, not minutes; no build-aside/swap needed for publishes at this scale. Earlier "minutes" concern was wrong. - Naming should carry arch + version (e.g. bq2-rootfs-<ver>-<arch>.tar.gz) so both architectures publish side by side. - dl.beagle.cc becomes a public face fed FROM the artifacts repo rather than a separate manual upload target.

Open design points: retention/GC policy for old artifacts; who may publish (daemon post-build vs human); whether the publish step joins the job pipeline as a bounded fossil child (see 36bc75236c envelope).