[JasonW404]

Memory configuration and failure isolation have runtime gaps: user-scoped Agent disable lists are loaded but not enforced in the current run path, embedding retries can wait 30/35/40 seconds, and default Dreaming model construction uses an observer class instead of an instance.

Real PostgreSQL diagnostics reproduce evidence truncation: 12,000 hits produce only 10,000 counted hits; 10,000 earlier hits for another Agent hide all 20 later target-Agent hits. Larger synthetic candidate diagnostics cover 100/1,000/10,000 candidates, and 38 existing targeted unit tests pass. These are baseline diagnostics, not post-fix acceptance.

This draft contains only an empty bootstrap commit. Analysis, functional and technical SPEC documents with acceptance criteria are prepared in the independent nexent-doc repository. Product implementation awaits the required SPEC approval. The user authorized local Docker for subsequent verification.