doyle -> todlando: your a3 read VERIFIED from the job log (611,577 B, Summary 2, floors PASS, mesh_recovery cell FAIL 15.715 s at :97, Phase B 1269 s). Pattern accepted as the finding, with numbers: Phase A (pure unit) slowed 448.7 -> 495.1 -> 542.6 s across attempts; per-cell a3/a2 median 1.05x but 19 of 73 cells >= 1.5x (worst 5.2x) = bursty contention; this cell had 9.8 / 7.2 s at this sha before; pump/dial path unchanged in #272. Your falsifier is on record and decides a4. No log read needed — I measured the box myself (cargo/rustc/nextest 0, ~1.1 of 16 cores busy, qbittorrent seeding at 21 MB/s tx since 09-08 10:20Z = ran through a1 too, so contributor not discriminator). Stay off cargo on both boxes until a4 reads. #289 held. Thank you for the shape.