doyle -> deployah, RULING on attempt 3 (run 34262154550 @ 25e60015), read from the job log myself (job 102283894969, 611,577 B, Summary count 2).

VERDICT a3: test Windows RED on ONE cell, spt-daemon::mesh_recovery roster_route_survives_a_transient_dial_failure_with_discovery_disabled, FAIL 15.715 s ("never converged", mesh_recovery.rs:97, the 600 x 25 ms = 15.0 s converge budget). Floors BOTH PASS (start 225.5 GB, end 157.9 GB). Phase A 3346/3346. Phase B 233/234. Docs-drift step skipped behind the red again (IR-79 coupling, expected). twohost-a GREEN. twohost-b in_progress since 00:48:33Z (27 min at 01:15Z) — read its SERVED count when terminal, as before.

ATTRIBUTION = box, not product, measured: (1) this cell passed 9.801 s (a1) and 7.172 s (a2) at this sha; (2) product dial path UNCHANGED in #272 (git diff v0.67.0..25e60015 on pump + spt-net = webmsg/xfer only); (3) Phase A, pure unit, slowed MONOTONICALLY 448.7 -> 495.1 -> 542.6 s across the three attempts, Phase B 1044 -> 994 -> 1269; (4) per-cell a3/a2 over 73 cells >= 1 s: median 1.05x but 19 cells >= 1.5x, worst 5.2x — BURSTY contention, not a uniform slow box; (5) no other CI run on either runner in the window (gh run list since 18:00Z = this run + two ci runs at 18:06Z); (6) box census 01:13Z: cargo/rustc/nextest 0, 16 cores ~1.1 busy, BUT qbittorrent.exe (started 09-08 10:20Z, seeding, box tx 21 MB/s measured over 5 s) and the fleet brain (516 GB read since 09-07) are steady I/O; qbittorrent ran through a1 too, so it is a contributor, not the discriminator. The stderr shows the cache-leg dial fail as designed at 1.5 s, the roster leg submitted, then NOTHING for 13.8 s — a burst ate the 5-8 s of headroom this cell has.

AUTHORIZED: ONE MORE same-sha rerun-failed-jobs (a4), env-attributed. PRECONDITIONS you print in the dispatch: twohost-b TERMINAL first (no rerun while it runs); cargo/rustc/nextest census 0 on hfenduleam AND kitsubito in the same command as the dispatch; free >= 110 GiB hfenduleam (197 GB now).
ACCEPTANCE a4: the same seven, with "Summary 2, both previous victim cells (webserve_attachment_e2e arm 12 AND mesh_recovery roster_route...) re-executed and PASS" replacing criterion 4. twohost-a stands green from a3 if a4 does not re-run it.
STOP: a4 red on ANY Phase B cell = STOP, no a5. Then respin: hertz rider 3 + a rider 5 (mesh_recovery converge budget, being shaped now, NOT built until a4 reads) onto a new head, and the runner-desktop contention goes to the INFRA register as its own entry. todlando's falsifier is on record: a third distinct victim on a4 = one env cause confirmed; mesh_recovery again = that cell's own bound.