Expand description
On-disk cache of parsed schema records.
Parsing a large SDL file dominates the fuzzy path (~28ms of a ~45ms query on a 48k-record schema), and the records depend only on the SDL text — so cache them keyed by a hash of it. Same idea and same house format rules as the vector cache: a small length-prefixed binary layout, dependency-free. A miss (schema edited, format bumped) simply re-parses and overwrites.