Procedural GM bank: pilot A/B, round 4 (after "376, 405, 293-alt end differ")

Scores now use a loudness-weighted (A-weighted) comparison with a separate onset term, so they are lower than in round 3 but track what you hear. Rate each pair good / ok / bad, and say what is off.

376 Standard kit kick (key 36): broadband noise + end fade matched (round 9) score 0.7091 model alone 0.384 d_spec 1.4671 d_onset 1.5802 d_env 0.5392 d_loop 0.0	original generated
161 Standard kit snare (key 38): end fade matched (round 9) score 0.7408 model alone 0.218 d_spec 1.2975 d_onset 1.1718 d_env 0.5653 d_loop 0.0	original generated
83 Closed hi-hat (key 42): refit with current pipeline (round 9) score 0.7183 model alone 0.3793 d_spec 1.8357 d_onset 1.8033 d_env 0.1654 d_loop 0.0	original generated
240 Fingered Bass (prog 33) score 0.8887 model alone 0.6741 d_spec 0.2979 d_onset 1.0518 d_env 0.0332 d_loop 0.0001	original generated
405 Strings (prog 48): noise calibrated by periodicity, no hiss (round 10) score 0.7974 model alone 0.7755 d_spec 0.7869 d_onset 0.8012 d_env 0.2678 d_loop 0.5932 pitch_off	original generated
293 Piano 1, wave 293: stiff-string attack + stretched loop, refit (round 9) score 0.7573 model alone 0.5529 d_spec 1.2089 d_onset 1.5637 d_env 0.1189 d_loop 0.3255	original generated