Why: Ruling 6 requires real-model before/after evidence. The after run (fixed
build b68833f, same golden tasks) proves the repairs landed: lint now actually
executes (real layout/lead findings appear vs silent empties before), the
pm-tool readability dead-loop is gone (4/0 -> 0), plan-followup reads original
materials 0 -> 49 times. Remaining failures are the documented P5/P2 scope
(rename-revision loops via scorer semantics, unverified numbers, verifier
parse hiccups), not regressions.
Evidence: SELF - report .task/phase01-eval/after-phase01-b68833f.json +
evidence row (1749s run); comparison table in the card.
Impact: PR #277 now carries the complete before/after pair for review.
Why: The backend branch gained the user-scope model capture fix, and the
live eval produced the before-baseline report on remote main (1ebd46e) that
the task card and the improvement-plan record reference as evidence.
Evidence: SELF - report .task/phase01-eval/before-main-1ebd46e.json (model
MiniMax-M2.7-highspeed, 4 golden tasks, 17 min run): reproduces the diagnosed
failure chain on the pre-fix build - readability dead-loop x4 on two tasks,
9 full regens with 0 original-material reads in plan-followup, 8 unverified
percentages, forged-marker false positive, internal check text leaked to the
user.
Impact: Report committed for the after-comparison once this PR merges; gitlink
points at the harness-inclusive backend head.
weichen