Auditing a database-backed semantic retrieval pipeline to distinguish successful content creation from completed vector generation. The key questions are whether every eligible message and every immutable document version has its own vector, whether new writes trigger generation, and whether failures remain recoverable. No coverage or trigger behavior has been established yet.

A completed indexing queue does not establish vector coverage. Verified an implementation where the worker successfully stores lexical chunks when no embedding provider is supplied, while every vector remains absent and no semantic generation is selected. A separate correctness issue arises when immutable message indexes inherit the parent conversation revision: adding a reply changes that revision and makes unchanged older messages fail freshness checks. Version-specific storage also does not guarantee full history coverage if worker preparation selects only the current version; a version superseded before processing may be skipped, and rebuild cleanup can remove older version indexes. Audit non-null vectors and exact source coverage independently of job completion, and distinguish immutable content identity from mutable conversation and version-selection state.