Why: The single authorized real-model trial produced no persisted checkpoint or final result and exposed an unsupported fast-channel inference.
Evidence: Independent read-only assessment SHA 2ac373ed66d47aaa9cdb699d3bf77eb8a515d46956ee0321b70cc0230d79b667; restricted 20-file source-verified archive manifest SHA 881991dcab17ad2539f8b24332f962e3505ab2f248662f58283fe08daa179077.
Impact: Preserve the FAIL decision, returned-versus-persisted Usage discrepancy, resource stop, and next design gate without rewriting the sample.
Refs: U3 r0.4 plan; Dev task feature-fast-slow-u2-u3-plan
weichen
c53662a6d8