Both documents described the redesigned mirror check as "two precise questions" and recorded one residual -- the false negative, an undeclared mirror on a host that declares none. Measured otherwise. Gating the survivor scan on `mirrors` being non-empty exempts host #1 by construction, but the second question is still the same "holds the same value implies is a copy of the leaf" inference the retraction above correctly names as the root error. So on a host that DOES declare mirrors -- host #2, this slice's whole subject -- an unrelated sibling coincidentally equal to the leaf is still refused, at the preflight, as a blocking stop with no audit entry. That is the failure that broke host #1, surviving one host over. The root error is narrowed, not removed, and the difference is the whole distance between a guarantee and a trade-off. Two further limits stated rather than assumed away: the scan walks only the clone of `outputs.within`, so a duplicate kept outside that container is invisible however it is declared; and an undeclared mirror on a host declaring none is not detectable at all. This branch has now retracted two claims that were repeated rather than re-measured. This one is corrected before it earned a third. Slice: #6 Claude-Session: https://claude.ai/code/session_01BNUFdwMp4k1bxcGsANgXzo

V5: Second host, zero AGT changes #6

Description

Amendments

Activity

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions