A few follow-ups on the eval loop, plus Open Responses support for the server judge.

Run pages now show token, cost, and latency totals live while a run is still in progress, instead of only after it completes. The totals are read from the items on the fly for running and evaluating runs and from the materialized columns once the run is terminal.

Adds a review queue: a page and endpoint that list run items still needing a human verdict (never annotated, or marked unsure), so a reviewer can work through one queue instead of opening runs one by one. It can be scoped to a project, and annotating an item drops it from the queue. The query is backed by a Postgres test.

Adds Open Responses support to the server judge. Open Responses is a vendor-neutral, multi-provider API standard; the judge now speaks it by default and falls back to Chat Completions per connection. Each LLM connection carries a protocol field (Responses or Chat Completions); new connections default to Responses, and existing connections are backfilled to Chat Completions so nothing currently working changes. The connection form gets an API selector. Covered by unit tests for both request and response shapes and a live integration test against the real endpoint.

Adds live-HTTP integration tests for the judge against a real OpenAI-compatible endpoint (Chat Completions and Responses), gated on OPENAI_API_KEY and excluded from the normal test run. Points failsafe at the exploded classes so these run in the server module, whose Spring Boot repackage would otherwise hide the classes.

Also hardens the reporter integration test to locate items by input rather than position, since a batch flush leaves their order unspecified.