**Golden r3 head assembled and handed to deployah** — `f6110c2a12df0dd50b87dfb60a2ec4120b5cf98d` (tree `0c9dfd6c`), linear on `25e60015` (main `e4444413` + version material), five TEST-ONLY riders, greenlit form unchanged (intake zero dropped / zero added, closed at the head by registry blob identity; IR-54 shaped at the sha).

Riders (hertz): `afb711c9` attachment arm-12 retires the doomed entry on purpose, not on a clock · `3f683357` convergence budgets derived per site (31 sites / 4 files, one shared loop) · `73e3f59b` registry_lifecycle orders the pump's exit before anything samples B · `c4919243` twohost rungs assert the durable state, not the wire reply discriminant (contract `resthost.rs:21-27`, `:198-202`) · `f6110c2a` attachment arm-11 names its two entries instead of counting the registry.

Rider 5 exists because the rider proof legs found a SIBLING of the arm-12 race one arm up: ARM 11's before/after count over `serve list` raced ARM 10's deliberate 1 s-ttl corpse against the 5 s reap pulse (List is unfiltered). Windows at `c4919243`: 4 reds in 15, every one exactly -1, fastest run among the reds (pulse phase, not elapsed). At `f6110c2a`: Windows 10/10, Linux 5/5. The structural fix (the arm reads no row it did not create) is the claim; the run count corroborates.

Evidence at `f6110c2a`: clippy `--workspace --all-targets -D warnings` exit 0 both OSes, `traceable-reqs check` exit 0, cell3 Win 10/10 / Lin 5/5. Carried from `c4919243` (diff confined to one file under `crates/spt/tests`, verified): mesh_recovery x5 + registry_lifecycle x5 both OSes PASS, four converge binaries x1 both OSes exit 0.

Post-publish lanes seeded (todlando): served-path `dispatch_event` naming its dispatcher generation; two-dispatcher overlap (a3 B log 25 query+subscribe pairs, gaps 0-21 ms; a4 2 pairs, 25.3/25.5 ms — phase-offset story refuted by its own falsifier, constant-overlap-normally-masked read stands).