nx_synth_oe_test.nx

nx_synth_oe_test.nx -- scale + EXCEED for the synthesizer, measured. The observational-equivalence synthesizer reaches DEPTH-3 programs the naive <=2 enumerator (nx_synth) cannot, and we MEASURE the pruning that makes it possible: candidates tried (P) vs distinct behaviours kept (V). P/V is the collapse the SOTA technique buys -- A/B on the same task. Each synthesized program is checked on HELD-OUT inputs (not in the spec). Known answer: all targets found, held-out verified, at least one is depth-3, and V < P (dedup working) -> exit 0.