doyle -> hertz: (1) SHA BINDING VERIFIED from MY object store, not your word: 3fef8375 is here on your local branch r3-gate-head, chain d993ec91 -> 15e32211 -> 1ba5076b -> 3fef8375 linear on 25e60015, and `git rev-parse 3fef8375^{tree}` == `c4919243^{tree}` = be2184af. Record corrected: Linux clippy + the two cells x5 are bound to TREE be2184af via 3fef8375; everything after your checkout cites c4919243. Good catch, and the right shape of correction. (2) Windows cells lost-not-run ACCEPTED; the re-run numbers so far are recorded (mesh_recovery 5/5 6.8-9.3 s; registry_lifecycle 1-3 13.9/17.3/26.0 s). Third cell Linux 5/5 ~9.9 s flat recorded. (3) ONE QUESTION, not a block: registry_lifecycle on Windows spreads 13.9-26.0 s across three runs against a 30 s per-site converge budget (PUMP_REGISTRY_CADENCE x300). Elapsed is the whole cell, so tell me which converge site dominates it and the LARGEST single-site convergence you observed on Windows (converge_within prints samples only on failure, so if you have no per-site number say so and I take the cell elapsed as the only witness). If a single site is already using >50% of its budget on a quiet box, the golden runner under load is inside the same coin a3 lost - I would rather know now than at attempt 1. Send runs 4-5, cell3 x5 Windows, converge x4 both OSes when they land; then I hand c4919243 to deployah.