[canonical 4.2 harness] session=C3
[canonical 4.2 harness] protocol: 10 warmup + 100 continuous timed iters
[canonical 4.2 harness] inter-shape settle: 5.0s (NOT section-4 cooldown)
  lcsa_small_seq4k                 d=0.239 V2 p50= 1.143ms SDPA p50= 2.573ms ratio= 2.25x
  lcsa_small_seq4k_sparse          d=0.067 V2 p50= 0.695ms SDPA p50= 2.548ms ratio= 3.66x
  lcsa_mid_seq8k                   d=0.119 V2 p50= 1.177ms SDPA p50= 6.453ms ratio= 5.48x
  lcsa_mid_seq8k_sparse            d=0.030 V2 p50= 0.629ms SDPA p50= 6.504ms ratio=10.35x
  lcsa_large_seq16k                d=0.120 V2 p50= 2.019ms SDPA p50=12.786ms ratio= 6.33x
  lcsa_large_seq16k_sparse         d=0.030 V2 p50= 0.905ms SDPA p50=12.956ms ratio=14.31x
  lcsa_mid_seq8k_very_sparse       d=0.011 V2 p50= 0.518ms SDPA p50= 6.523ms ratio=12.59x

[canonical 4.2 harness] session 'C3' -> docs/methodology/canonical-bench-data.json
