[lemire]

Copying a field or a value out of a simdjson_result (std::move(r).get(f)), or passing an sj::value by value, goes through the stack: the compiler writes the fields of the value iterator with an 8-byte and a 4-byte store, then reads them back with a 16-byte load. That load cannot be forwarded from the two stores, so it waits for them (see simdjson/simdjson#3052 and simdjson/simdjson#3061).

We now use the field or the value in place inside its result, after checking the error, and validate() takes its value by reference. No behavior change.

bench_ndjson, one thread, Xeon Gold 6548N, best of 3 interleaved rounds, GB/s (simdjson master):

case	GCC 16	clang 21
gharchive, gh schema, validate_skipped	1.933 → 1.938	1.335 → 1.944 (+46%)
gharchive, gh schema, no validation	3.120 → 3.098	3.163 → 3.152
flat, validation	0.731 → 0.742	0.776 → 0.763
gharchive, inferred schema	0.343 → 0.343	0.351 → 0.350

The gh schema reads six fields and validates all the others, which it skips; validate() took its value by value, so every skipped value paid for the stall with clang. With simdjson#3061 alone, clang gets the same gain but GCC 16 loses 4% on this case (1.859 GB/s); with this change on top, GCC 16 is at 1.961 GB/s.

Tests pass with the pinned simdjson 5.0.7 (Apple clang) and against simdjson master, with and without #3061 (GCC 16 and clang 21).