vms-629: OVMX's lock directory answers a real VAX's op-0x08 barrier lookup (the VAX barrier no longer wedges at step 7)

[baron-3dl]

The bug. A real V7.3 VAX forming a cluster with OVMX as a founding member (the vms-f29 path) does this:

• At barrier step ~7 it sends its directory node cat-0x02 op 0x08 for SYS$SYS_ID+⟨its own SCSSYSTEMID⟩.
• It holds the next release until that lookup is answered.

OVMX refused op 0x08 under RULE C. As a result:

• The VAX barrier stopped at step 7 forever.
• OVMXA stayed in transition_active.
• The VAX never proposed the next joiner. Nobody else was admitted, with 0 bugchecks (rd vms-629).

Oracle. Two independent real-VAX samples:

• coldform-ev2-formation.pcap records 121/123.
• A new dlmlab two-VAX cold formation, tests/lab/captures/vms-629-op08-20261005/cf1-op08.pcap.

In both, the other VAX answers 82/01 within 1.5 ms: the request is echoed, the opcode byte reads 0x01, and body[34] = 0xf9 ("you master it"). The barrier then proceeds.

Change.

• New constant VMS_DLM_WIREOP_DIR_LOOKUP_TR (0x08). The directory role (vms-8219) reads it as a root lookup and answers it with the same grounded builder as an op-0x01 lookup.
• That builder now writes opcode 0x01 in every answer.
• A redirect answer to an op 0x08 has never been observed, so that case is held back, counted, and logged once.

Tests.

• R1: test_dlm_dir answers both real requests byte-identically to the real VAX answers, opcode included (the stale span is handled as in vms-8219). test_dlm_scs_arm source-scans the routing.
• Full host ctest: green, apart from the known load-flaky terminal_identity_gate (rd vms-946).
• Lab (dlmlab rig1, runO): OVMXA founds, a real VAXC forms with it, then OVMXB joins.

build	formation barrier	OVMXB admitted	real-VAX bugchecks
main (O-base1)	stuck at step 7, op 0x08 unanswered	0/1	0
this branch (O-629-1, -2)	completes, op 0x08 answered 82/01	2/2, CN=3	0

More arms are running; I will post the results on this PR.

🤖 Generated with Claude Code

[baron-3dl]

Stall matrices on the combined build (main + vms-629 + vms-ba4 final, SOURCE 96e71f6), dlmlab, real V7.3 VAXC: 20 arms, 0 VAX bugchecks. The runs were R-1..R-14, N3-1..N3-3 and P-1..P-3; the k3s-worker rebooted under load twice, which ended two runs early.

The class this fixes is the b-then-a 13/14 s arms:

• Before the fix (builds l629/mcomb, same pod): 5 of 13 bugchecked CNXMGRERR (L-3, L-9, N1-3, N2-2, N2-4). All five show the same dialogue-counter shape, documented on rd vms-ba4.
• With the final fix: 0 of 6.

The remaining FAILs fall into the tracked classes: vms-ec2, vms-04b, and stall longer than RECNXINTERVAL.