Problem
Malformed cited entries (e.g. ["1", "two", 2.5, -1]) are silently dropped and never recorded, so eval scores hide model misbehavior.
Where
vaultrag/generate.py (citation index parsing vs dropped recording).
Acceptance
Add a failing test first asserting malformed entries surface in dropped_citations. Then fix.
Problem
Malformed
citedentries (e.g.["1", "two", 2.5, -1]) are silently dropped and never recorded, so eval scores hide model misbehavior.Where
vaultrag/generate.py(citation index parsing vsdroppedrecording).Acceptance
Add a failing test first asserting malformed entries surface in
dropped_citations. Then fix.