The first run flagged one unchanged float32 VJP coordinate: reference -3.966e-14, runtime 0. The float32 local derivative scale is about 0.0036, so this cancellation residue is far below one unit of roundoff at the working scale. The harness now includes 8*float32_epsilon*max_abs_reference for float32 arithmetic, in addition to relative tolerance. Both before and after had the same false flag. The production patch was not changed.
First compatibility process exited with status 152 after reaching the 30 CPU-second limit, without emitting a result. No compatibility pass is claimed for that invocation. Subsequent diagnosis uses phase logging and a 5 CPU-second limit. Process-list inspection was unavailable in the sandbox (sysmond service not found); no escalation was needed for it.
Empty-input isolation: original GroupNorm on shape (2,0,6) returned [[],[]]. Prototype v1 reached evaluation but was stopped by a 2 CPU-second limit. It introduced exposure to the existing fast layer_norm empty-reduction path. Final patch adds an x.size guard and keeps the original path for empty arrays. v1 source, patch, metadata, successful nonempty regressions and diagnostic logs are preserved.