this is a quick sketch to protect against more repos disappearing while the main hubble project is still under development. scoping it to non-bluesky makes it a quick project (hopefully bluesky is doing a good job backing up repositories).

since the purpose is simple temporary archiving, we make some specific choices:

no identity resolution
- which means no signature verification -- any PDS can lie about repos it hosts
- but we keep all repos from every PDS, even if they have cross-PDS DID-conflicts
full re-fetch on repo update
- we retain the latest repo rev per (pds, repo) to avoid downloading exact archives we already have
- don't run too often, to avoid excessive egress from indie hosts
- offer Accept-Encoding headers

since this is part of hubble, we also make some choices that mostly only make sense for where we're headed:

record-level storage
- repositories are spread out over a rocksdb keyspace, with individual records keyed by pds, did, and record path
- repo deletes use rocks' range-delete functionality
- lexicographic key order means we can still recover a full STAR-lite archive efficiently, or a stream-ordered CAR with minimal overhead using the STAR-lite MST-recovery algorithm.
repo statistics info
- wire size, uncompressed car size, backlinks counts, blob ref counts by mime type, etc.