[feat] Bucket layer for sync

[St4NNi]

Closes #262

Adds the shard-based sync partition layer on top of placement.

• Documents are grouped into shards (the sync partition unit; "shard" avoids collision with S3 buckets), each with a per-(strategy, shard) topic, manifest, and digest-based anti-entropy.
• Shard-topic genesis is single-origin by rank uniqueness: only the resolver's rank-0 holder of a (strategy, epoch, shard) creates the topic genesis; other nodes join-only. Outbox records whose shard-topic genesis hasn't arrived are deferred until it does, avoiding split-brain from concurrent first-writers.
• Bucket assignment and topic derivation are wired through document sync, onboarding, metadata, and the admin outbox; node-info and onboarding-placement invariants from the placement core are preserved.

Depends on #261 (merged).