← RFC

Backbone B1 Results — first INFRA↔INFRA backbone, bench to tower

Date: 2026-06-11 · Code: loomwave repo 6a8e4cd (B1.3 merged coordinator) · Design: backbone-b1-design.md

Summary

The B1 milestone series is complete. Two single-radio coordinators now form a co-channel backbone (half-frame stagger), exchange signed link-state adverts and sealed backbone DATA, and forward client traffic across cells with hop-by-hop re-sealing — first on the bench (B1.2/B1.3), then over a real RF path between the bench and the mid-tower node (B1.4).

Milestone	Setup	Result
B1.2	Bench Duo: SX1262 ↔ LR1121, two processes	Symmetric adjacency, COORD_CLAIM verified, 5×5 LSAs, 5×5 sealed pings
B1.3	B1.2 + T114 client joins one cell	First cross-cell delivery: 27 forwarded client frames verified at the peer
B1.4	Bench SX1262 ↔ mid-tower (Nebra Mesh HAT, full power, 60 ft AGL), T114 client	Backbone over real RF: adjacency held the full 400 s run; 46 cross-cell client position reports delivered to the tower

B1.4 detail (clean run, 2026-06-11)

• Tower 758F54 anchored (role A); bench 41F34E heard it during scan-before-claim and staggered +2600 ms (role B). Roles assigned purely by radio observation.
• Link budget: tower hears bench at −62 dBm (bench TX = −9 dBm chip); bench hears tower at −34 dBm (chip 22 dBm → ~30 dBm after FEM). Plenty of margin both ways at SF9/BW125; the −104 dBm decode demonstrated earlier in B1.2 prep bounds the real headroom.
• Cross-cell over distance: T114 72466B joined the bench cell (sealed REGISTER → slot 0) and sent 63 sealed DATA frames (live GPS positions, fix valid). The bench verified each, re-sealed with the tower's epoch/keys, and forwarded; the tower verified and decoded 46 cross-cell frames arriving at −59…−63 dBm. A client position report traveled client → cell coordinator → backbone → second coordinator, with independent AEAD on each hop.
• Backbone control plane: tower→bench delivered 8 LSAs + 8 sealed pings (~89% of those scheduled). bench→tower delivered only 1 of each — see the known limitation below.

Bugs found and fixed along the way (B1.3)

1. Coordinator went anonymous after first client join. The identity advert was gated on members.is_empty(), so the moment a client registered, the coordinator stopped advertising its keys and peers could never form an adjacency. Fix: when slot 0 is occupied, the advert rides the contention window (+1428 ms = beacon ToA 525 + guard 3 + 3×300 ms slots), which registered clients leave idle; the 670 ms advert ends at 2098 ms, just inside the +2100 ms control window. Spec implication: identity adverts must never be suppressed by cell occupancy — recorded for the next spec refresh.
2. Test-harness stranding: a launcher script re-flashed the T114 mid-demo with discarded output; a partial DFU left the device in its bootloader (USB PID 0x0071) and two demo runs were debugged against a client that was not running. Process rule reaffirmed: never discard the output of a flashing step; assert the device re-enumerates as the application.

Known limitation → B2 backlog

Control-window budget accounting. Queued cross-cell forwards drain at the start of the control window, ahead of the LSA/ping; with a client generating one forward per superframe, the cell-owning coordinator's LSA is pushed past the half boundary and dies in the peer's TX half (observed: 1/9 LSAs delivered bench→tower while 46/63 forwards got through). The control window needs airtime budgeting — emit what fits, defer the rest — alongside the existing B2 items (N>2 stagger collision, election action, cell summarization, end-to-end sealing).

What B1 proves for the design

• The half-frame stagger lets two single-radio coordinators each keep a full client cell and a backbone on one channel — no second radio required for the minimum viable backbone (multi-radio INFRA remains the scaling path).
• Hop-by-hop re-sealing works across three independent crypto stacks (T114 C++ / bench Rust / tower Rust) with zero verification failures.
• Scan-before-claim assigned roles correctly in every clean run, including over the real link with ~28 dB of RSSI asymmetry.