Impact

Open PR #297 (fix/research-roadmap-hotpath) is now stale against current origin/main. If merged as-is or resolved mechanically, it would roll back current ROADMAP truth by deleting the merged LeakyCLIP CLIP-inversion boundary gate and by removing the ReDiffuse STL-10 SimA-style score-norm weak result from the hot path.

This is separate from #299 / #301 / #304: it is not the Tracing row-six regression, the feature-packet consumer-boundary issue, or the already-closed LeakyCLIP boundary decision itself. It is an open-PR stale merge risk.

Evidence

• gh pr list reports PR Clarify superseded ReDiffuse preflight in roadmap #297 as BEHIND.
• git diff origin/main..origin/fix/research-roadmap-hotpath -- ROADMAP.md shows the branch would delete the current top ROADMAP section 2026-05-25 LeakyCLIP CLIP-inversion 边界门控.
• The same diff changes 2026-05-25 ReDiffuse STL-10 bounded scout 与 score-norm 结果 back to bounded scout 结果, removes the score-norm metrics (AUC = 0.5052947998046875, ASR = 0.525390625, TPR@1%FPR = 0.03125, TPR@0.1%FPR = 0.01953125), and removes the rediffuse-stl10-sima-score-norm-20260525.md evidence link.
• A three-way merge preview also touches the same ROADMAP hot path, so this is a concrete stale-branch risk, not only cosmetic drift.

Suggested fix

Close #297 as superseded or rebuild it on current main preserving:

• the LeakyCLIP CLIP-inversion boundary gate as the latest Lane A update;
• the ReDiffuse STL-10 denoising-loss plus score-norm weak-result closure;
• the current active_gpu_question = none, next_gpu_candidate = none, and no sidecar/GPU release state.

If the original #297 intent is still useful, carry forward only the safe historical note that the older ReDiffuse STL-10 preflight/download/training language is historical, without deleting later main facts.

Notify: Researcher Agent