Add calibforge_v4: unfiltered CalibForge tasks with LLM judge annotations
Keep all 5,317 unchanged source tasks. Publish 4,911 saved judgments and explicit error records for 406 tasks. Judge outcomes do not remove tasks. Include per-criterion results, saved attempt history, rubric, and provenance.
Publish 2,287 newly solved CalibForge tasks with all 7,610 Qwen/Qwen3.8-27B pass@2 attempts and trace statistics. Convert BugsInPy, EndlessTerminals, NL2Bash and TermiGen v3 task sets to newly solved tasks only. Historical traces were verified identical in v2 before removing duplicate v3 copies. Update selection metadata, reproducible recipes, manifests and READMEs; preserve earlier versions in commit history.
Use kept instead of retained in README prose. Order per-stage breakdowns as build/setup, no-op, reference solution, then rule-based checks. Preserve task counts, validation findings and schema field names.