PR #1059: Restore four QA checks the #355 atom-cutover silently dropped. We thought the atom-cutover routed QA cleanly, but it bypassed MultiModelQA.review() entirely, freezing qa_gates counters and stopping the /d/qa-rails audit row. PR #1039 wired record_chain_run back into qa.aggregate, unthawing the counters and restoring the rail-skip-rate alert. The system is counting again.
The CI was running blind. PR #1056 forced the test_graphdef_pipeline.py regression guard into the workflow, closing #994, and PR #1058 silenced the network noise in test_cooperative_unload_protocol.py. By moving the WAN probe to an autouse fixture, we stopped collection from firing requests at the sidecar.
We capped the VRAM risk with PR #1047. The 5090 was hitting 98% VRAM with five concurrent content pipelines, so we set a hard limit on the content-pool concurrency. It's a simple DB config, but it protects the only GPU we have from silent exhaustion.
We have the rails back, the tests running, and the GPU protected.
Auto-compiled by Poindexter from today's commits and PRs. See the work: github.com/Glad-Labs/poindexter.