Ideas queued, in priority order
- Split lock and workspace coordination into a separate redb database per repository if a Linux measurement confirms lock operations queueing behind acceptance (macOS shows lock p50 above 100 ms under load versus 4 ms idle).
- Online scrubber in the server: background re-hashing of stored objects at
a bounded rate, reporting through the same events as
doctor. - Parallel data connections in the client (sessions already bind data connections).
- Per-path state-at-change query so backward syncs stop being full reconciles.
- Whole-file reflink reuse during sync where the working copy already holds an identical file (needs a content-id index of the workspace).
- Never: garbage collection, history pruning, pack compaction, or a custom database engine in this line of work.
Implemented: optional immutable per-repository LZ4 storage and independently configurable workspace transfer compression; see client usage.