A bare float32 score conflates "this row has no score, it came from
a scalar predicate" with "this row's vector similarity score is
0.0", which is a legal value: zero L2 distance and zero dot product
both arise in real queries. Mixing scored and unscored results in
one response (e.g. hybrid scalar+vector) becomes ambiguous.
Add has_score (proto field 2, sandwiched between Ref and Score) and
the corresponding HasScore field in the design's Go sketch. Update
the prose paragraph to call out the new flag's purpose.
8fe0b68b7c