v0.9.5Research Radar publication-boundary repair
Research Radar publication-boundary repair
GIDS now revalidates autopilot-published Research Radar records against hard public-boundary rules, keeps model summaries flowing when a field is returned as prose, and avoids slow-route/export-memory stalls during publication.
Fixed
- Autopilot-managed articles with animal-only, basic-research, or plant-only classification evidence can no longer remain in the public Research Radar catalogue.
- Automatically published summaries now return to review or archive when their parent article is removed from the public boundary.
- Summary enrichment now accepts scalar/list model fields as conservative evidence objects, reducing avoidable failures from otherwise usable model responses.
- Chronically failing Model Center routes are no longer used as active summary candidates until a structured health check proves recovery.
- Provider 403/no-access responses, including Chinese no-access messages, now mark the affected model route unavailable instead of retrying it for every summary.
Improved
- The repaired Research Radar export passes public-release validation again, with 964 public articles and 334 weekly briefs after recovery.
- Site-data export now scales partitioned download workers to the service memory guardrail, preventing releases from stalling under cgroup memory pressure.
- The dashboard worker now keeps site releases from overlapping with memory-heavy Research Radar and AI jobs.
- Queued site releases now hold off new Research Radar catch-up work so publication is not starved by continuous summary batches.
- Release commands now drain verbose Astro/Node output by byte chunks so builds cannot stall on a full stdout pipe.