The relay carries a full round trip - 2026-08-08

Files: transit-proof.pcapng (loopback, TCP 10362/10364/10366), transit-run.log. Result: D100 reaches D103 THROUGH our C# node, in both directions, with zero drops and zero Frame Rejects.

The measurement

D100's LIST-ROUTING-INFO,,,D103 - all three lines agree, including the ACTUAL measured path:

    103  L: *->19999->103
         T: *->19999->103
         A: *->19999->103
  19999  L: *->19999
         T: *->19999
         A: *->19999

A: is the path a live probe actually took, not a table lookup. Our node is a working COSMOS route-through relay.

Our side, from transit-run.log:

[relay] hdlc-out:127.0.0.1:10364 -> hdlc-in:10366 for node 103     D100 to D103
[relay] hdlc-in:10366 -> hdlc-out:127.0.0.1:10364 for node 100     D103 to D100
...
[relay] relayed=10 forUs=7 dropped=0

Ten datagrams relayed, both directions, none dropped.

	before the fix	after
FRMR frames	6	0
link flapping (Active -> Starting)	several	0
datagrams relayed	0	10
datagrams dropped	-	0

What made the difference

Two defects, both on the relay path only, both fixed in commit 751e3e8:

1. The announce was sent from inside the StatusChanged callback. That fires part-way through a batch of received frames, so with D100's six-SABM burst the announce landed BETWEEN two sequence resets and the peer's correct acknowledgement then fell outside the window - forcing FRMR reason Z every time. Now armed in the callback and sent on the next loop tick, after the burst is drained. Full mechanism: ../FRMR-ON-INNAK-2026-08-08/SESSION-NOTES.md.

2. The relay ran its links with NO keepalive interval, which the adapter documents as the in-memory test mode. LoopTick then fires only when a frame arrives - so fix 1 could never have been sent - and the LAPB T1/T3 timers never ticked on a live link at all. Found only because fix 1 was verified live rather than assumed. Both links now pass 20 ms.

Pinned by LapbAnnounceOrderingTests (replays the captured announce and INNAK bytes).

Machine setup that this needed

Rebuilt on both peers per COSMOS Operator Guide ND-30.025.02 section 2.5 - and section 2.5.4 is the rule that unblocked everything: an ADJACENT system needs a name and NO route; only NON-adjacent systems get a DEFINE-SYSTEM-ROUTE.

D100:  DEFINE-REMOTE-NAME,,D100,100 / D19999,19999 / D103,103
       DEFINE-SYSTEM-ROUTE,,D103,D19999        (103 is NOT adjacent - it is behind us)
       START-LINK,1360,,,-1,,                   (no route to 19999 - it IS adjacent)
D103:  mirrored, with DEFINE-SYSTEM-ROUTE,,D100,D19999

D100's RetroCore.ini has its two HDLC listen ports SWAPPED so LU 1360's controller is the one only we dial (10364). That costs D102 its line to D100; restore the commented lines to give it back.

Still open

• INNAK is still never answered. Both peers send one on every announce and the runner logs *** NO REPLY BUILT ***. It is demonstrably NOT fatal - this whole round trip happened while ignoring them - but what the correct reply is has never been captured.
• D100's XROUT degrades after a while: LIST-ROUTING-INFO starts repeating one system's block forever and has to be broken with ESC, and STOP-XMSG/START-XMSG clears it. Whether our traffic provokes it is UNKNOWN.
• Marker 2 0x12 on a relayed datagram is still unverified. It was wrongly suspected as the cause of the drop; the relay now works WITH the rewrite in place, which is evidence it is at least accepted, but not that it is correct.