[v34-investig] correctness smoke...
  smoke: rmse=4.4228e-08 maxerr=3.8147e-06 NaN=0 Inf=0 -> PASS
[v34-investig] initial cooldown 180.0s

[v34-investig] PROBE 1/5: B+C+E_aggregate_predecessor_vs_v34
  V34 vs predecessor path: includes hypotheses B (cross-SG sync elim), C (simd_shuffle_xor vs MPP reduce), and E (Apple defaults). Aggregate measurement.
  baseline_env={'MFA_V6_USE_V34': '1'} alt_env={'MFA_V6_USE_V34': '0'}
  v34_small_d64          ALT=  0.504ms BASE=  0.496ms ALT/BASE= 1.01× drift= 4.0%
  v34_small_d128         ALT=  0.924ms BASE=  0.823ms ALT/BASE= 1.12× drift= 5.9%
  v34_mid_d128           ALT=  3.632ms BASE=  3.045ms ALT/BASE= 1.19× drift= 1.6%
  v34_large_d128         ALT= 13.141ms BASE= 11.176ms ALT/BASE= 1.18× drift= 0.5%

[v34-investig] PROBE 2/5: A_tgp_low_sg2
  Hyp A: lower TGP occupancy. MFA_V6_EXEC_SG=2 (default=4).
  baseline_env={'MFA_V6_USE_V34': '1'} alt_env={'MFA_V6_USE_V34': '1', 'MFA_V6_EXEC_SG': '2'}
  v34_small_d64          ALT=  0.607ms BASE=  0.481ms ALT/BASE= 1.26× drift= 7.6%
  v34_small_d128         ALT=  0.904ms BASE=  0.886ms ALT/BASE= 1.02× drift=10.2%
  v34_mid_d128           ALT=  3.075ms BASE=  3.264ms ALT/BASE= 0.94× drift= 0.1%
  v34_large_d128         ALT= 11.164ms BASE= 11.165ms ALT/BASE= 1.00× drift= 0.1%

[v34-investig] PROBE 3/5: A_tgp_high_sg8
  Hyp A: higher TGP occupancy. MFA_V6_EXEC_SG=8.
  baseline_env={'MFA_V6_USE_V34': '1'} alt_env={'MFA_V6_USE_V34': '1', 'MFA_V6_EXEC_SG': '8'}
  v34_small_d64          ALT=  0.516ms BASE=  0.553ms ALT/BASE= 0.93× drift=14.9%
  v34_small_d128         ALT=  0.819ms BASE=  0.852ms ALT/BASE= 0.96× drift=42.1%
  v34_mid_d128           ALT=  3.065ms BASE=  4.533ms ALT/BASE= 0.68× drift= 2.2%
  v34_large_d128         ALT= 11.141ms BASE= 11.172ms ALT/BASE= 1.00× drift= 1.2%

[v34-investig] PROBE 4/5: D_block_r_64
  Hyp D: larger tile = more register pressure. MFA_V6_BLOCK_R=64 (default 32).
  baseline_env={'MFA_V6_USE_V34': '1'} alt_env={'MFA_V6_USE_V34': '1', 'MFA_V6_BLOCK_R': '64'}
  v34_small_d64          ALT=  0.612ms BASE=  0.583ms ALT/BASE= 1.05× drift=10.2%
  v34_small_d128         ALT=  0.858ms BASE=  1.285ms ALT/BASE= 0.67× drift= 6.3%
  v34_mid_d128           ALT=  3.025ms BASE=  3.083ms ALT/BASE= 0.98× drift= 1.2%
  v34_large_d128         ALT= 11.148ms BASE= 11.185ms ALT/BASE= 1.00× drift= 0.5%

[v34-investig] PROBE 5/5: D_block_c_64
  Hyp D companion: larger K-tile. MFA_V6_BLOCK_C=64 (default 32).
  baseline_env={'MFA_V6_USE_V34': '1'} alt_env={'MFA_V6_USE_V34': '1', 'MFA_V6_BLOCK_C': '64'}
  v34_small_d64          ALT=  0.591ms BASE=  0.990ms ALT/BASE= 0.60× drift= 1.6%
  v34_small_d128         ALT=  0.825ms BASE=  0.952ms ALT/BASE= 0.87× drift= 3.0%
  v34_mid_d128           ALT=  3.166ms BASE=  3.080ms ALT/BASE= 1.03× drift= 3.9%
  v34_large_d128         ALT= 11.441ms BASE= 11.265ms ALT/BASE= 1.02× drift= 3.6%

[v34-investig] wrote: docs/v6-nax/v34-forward-investigation-data.json
