vms-c40: REJOIN — coordinator SDA CORRECTS §4(O.25): member-initiated reconnect DOES reach OPEN; op 0x02 breaks it (spec §4(O.26))

[baron-3dl]

Summary

Continues the rejoin long-pole at the SCS connection layer (vms-c40, parent vms-694). Reproduced the frontier with the CURRENT origin/main build (the shipped §4(O.25) evidence came from build-9a7 in a since-deleted worktree that could not be trusted line-for-line) and, with denser coordinator-side SDA, arrived at a decisive correction of the shipped frontier.

Isolating result (single-factor bracket, virgin vaxlab-1, GAP=22, coordinator SDA)

• The member-initiated VMS$VAXcluster reconnect DOES reach a stable OPEN on the return. Coordinator SHOW CONNECTIONS/NODE at R+4s: State 0002 open, Remote Con. ID 653B0001 (our return handle), with SCS$DIRECTORY + MSCP$DISK also open. OVMX's op=1 CONNECT_RSP echo (send_seq 17) + op=2 ACCEPT_REQ (send_seq 18), wire-proven correctly addressed to the coordinator handle 7288000f, open it. §4(O.25)'s "never settles OPEN" sampled only the post-break con_sent/03 reconnect phase and mistook it for never-opened.
• The connection is LOST ~10 ms after OVMX drives op 0x02 on it. Coordinator console %CNXMAN, lost connection at 08:50:23.75; OVMX CMCONFIG2 op 0x02 at 08:50:23.74. Per-send_seq decode of d94-c40repB.pcap on the membership VC: OVMX's own SCS$DIRECTORY+MSCP$DISK discovery bloats the shared per-VC send_seq, so op 0x01/op 0x02 ride send_seq 21/22 — past the coordinator's recv_ack ceiling of 18 — and the vms-9af lean-VC suppression fires ~1.3 s too late (08:50:25, after the op 0x02 and the break). CNXMAN never proposes (no second received VAXcluster membership request); the VC reverts to 03 reconnect and times out. Post-break op 0x05/0x06 go to Remote Con. ID 0.

Book cites: VAXcluster Principles pp. 2-43/2-44/2-45/7-37/7-40/7-46 (clean-room Rule 8 — page cites only).

Relocated frontier

From "the reconnect never settles OPEN" (SCS connection layer, §4(O.25)) back up to op 0x02 readmission DELIVERY on an already-OPEN member connection — the send_seq-ceiling / lean-VC-timing family (§4(O.15) vms-9af, §4(O.17) vms-46f, §4(O.10) vms-e15), now PROVEN open-then-break with coordinator SDA rather than inferred from the joiner side. Next increment (deferred to vms-694): make lean-VC / credit-first engage before op 0x02 is driven on the member-initiated connection so it rides under the ceiling — a wire-visible sequencing change in the crash-prone op02/START region, not made here (guard 8; "no speculative wire change until isolated").

XITDONE outcome

Not 0→1. The connection-layer "missing CONNECT_RSP" hypothesis is refuted (OVMX sends it and the connection opens); the real gate is upstream sequencing that is larger than this increment. Shipping the isolation + corrected frontier + fix design.

Ships (no wire change)

• Spec §4(O.26) and design-map §10.
• READMITMAP verdict-text correction (JOIN-ABANDONED no longer says "never settles OPEN"; states the grounded open-then-break finding, cites §4(O.26)). Classifier logic UNCHANGED (joiner_cfg2_sent && cm_responses==0), kill-switch OVMX_NO_READMITMAP unchanged, byte-unchanged on the wire, fail-pre/pass-post case in tests/vmsscs/test_scsd_wire.c. Non-admission (never ADMITTED; INV-6).

Test

cmake --build build && ctest -R "vmsscs|scs_" -> 52/52 pass; full build green. The verdict-text fail-pre/pass-post gate is genuine (pre: text carried "never settles OPEN"/"4(O.25)" -> assertions fail; post: "reach a stable OPEN"/"4(O.26)" -> pass).

🤖 Generated with Claude Code