Ships the hard A/B evaluation used for the August 2026 core-toolset
performance batch (#77056) as a reusable harness: 9 error-inducing trap
tasks derived from measured production waste classes, two-arm
PYTHONPATH-only comparison, ATOF-trace-based scoring, resume-safe
batteries.
Hardened from the original one-off: paths de-hardcoded (ABEVAL_ROOT /
ABEVAL_HOME), encoding= on all file IO, startup crashes retry on resume
instead of polluting cells, post-hoc grading fix for err_inline_script
baked in. Live-smoked end to end (baseline arm, qwen3-coder-30b,
err_multi_dir: exit 0, correct on-disk verification, resume record
written).
Teknium