v0.3.0

Per-worker queue affinity (FORGE_QUEUES / with_queues), Workers health tab, and the correctness fixes H3 + L7–L12. Breaking: ProcessRegistry trait + QueueRuntime::start() now requires declared queues.

forge v0.2.0

Second release of the forge workspace. All four crates aligned at 0.2.0.

Highlights

• Index-backed claim_next — a new partial index (jq_claim_ready)
  removes the per-claim sort that capped throughput once a backlog built.
  Measured on PG 18: ~4 → ~1,000+ claims/s on a deep queue, claim p99
  ~2,000 ms → ~8 ms.
• queue_event writes off the hot path — timeline events are buffered
  in-process and batch-flushed by a background loop, out of the enqueue /
  claim / finalize transactions (eventually-consistent chart data; no job
  state at risk).
• Postgres bloat tuning — fillfactor + per-table autovacuum on
  sync_queue / queue_event, plus a new
  operating-at-scale guide.
• Load harness + cross-system benchmark — a Postgres loadgen soak bin
  • criterion benches, and a reproducible bench/ comparison vs
    Sidekiq and solid_queue. Results + method in
    docs/benchmarks.md: forge beats solid_queue ~45× on
    pickup latency (push vs poll) with a 13 MiB worker.
• Docker: Postgres 18 data-dir mount fix.

See the commit history between `v0.1.0` and `v0.2.0` for the full list.