When a witness recorder is active, run z3::optimize.minimize on the
SAT model before extracting bytes for fuzz seeds. Refines integer
params toward the smallest path-and-hazard-satisfying value, so
seeds emit `arg=1` instead of `arg=0xDEADBEEF`.
Measured on synthetic access_at_offset(p, offset) with the path
`0 < offset && offset < 100`: primary-witness median 33 -> 5
(85% smaller). divide_by_count(p, count) with `count > 0` was
already path-minimal so no change.
Only the primary witness is refined; recordAlternateWitnesses
still uses the original model so the corpus covers the
equivalence class.
Curl-scale validation deferred -- fuzz-prep on full curl crashes
pre-existing (verified Z5-OFF baseline exit 139).
