scsd: re-emit op 0x02 to the coordinator at the ADD transition (vms-a84d)

[baron-3dl]

The stall

The booted-node cluster join stalls at barrier step 5/12: the second existing member (VAX1) never pushes its disk-DLM rebuild toward the joining OVMX node, so it withholds its own op-0b#5 and the coordinator never releases 0c#5. OVMXJ0 reaches the member flag but is never counted (CLUSTER_NODES stays 2).

Root cause (directly observed on the inter-VAX channel)

The run-1025 pcap was captured on the lab bridge and already contains COORD↔VAX1 traffic. Mining it (three-way grounded against the vax3-2to3 reference):

• The coordinator emits its op-0x12 "rebuild toward the new member" relay to the other existing member only in response to a FRESH joiner op-0x02 landed AT the transition. Reference: joiner op-0x02 → COORD op-0x12 relay 0.3 ms later → that member rebuilds and clears its barrier step.
• OVMX's op-0x02 to the coordinator is one-shot on the JOIN_CFG2_DELAY_MS (4.9 s) timer, decoupled from the coordinator's transitions. It did fire COORD's relay once — but for the pre-epoch transition. The OVMX-ADD epoch (0xE) opens afterward and gets no fresh op-0x02, so COORD never relays OVMX into VAX1's rebuild set.

Fix

When the coordinator opens an ADD transition (op-0x09 XITION) while we are the joiner, re-emit a fresh op-0x02 to the coordinator, once per ADD-epoch. Content is byte-identical to the deferred op-0x02 (proven byte-equal to a real joiner's); only the timing changes.

Honesty (INV-6)

No fabricated lock/held state. The gate-clearing member rebuild is a pure cat-0x82 echo OVMX already performs for the coordinator (the member originates its own value blocks; OVMX only reflects them). Two independent reference decodes confirmed the joiner never initiates held-state records.

Scope / safety

Joiner-path-guarded (rx->do_connect && !appeared_after_join), so the member/bystander and OVMX↔OVMX (H2) paths are untouched. No new opcode/enum/TU. All 26 vmsscs unit tests pass (incl. the 141k-frame scsd_wire replay).

⚠ Verification bar

H2 cannot exercise the real COORD→VAX1 relay (the #1026 lesson), so CI-green is necessary-not-sufficient. The decider is the lab re-fire vs vaxlab-2, graded by eye on both consoles via the SDA CSB oracle: COORD relays VAX1 into OVMX's rebuild → VAX1 emits 0b#5 → barrier 5/12→12/12 → CLUB Nodes 2→3 → OVMXJ0 = MEMBER at CLUSTER_NODES=3, cap-denied. Milestone stays un-cut until that is observed sustained on both consoles.

🤖 Generated with Claude Code

[baron-3dl]

Held as draft — the op-0x02 avenue is refuted. Lab re-fire #1029c-1 confirmed the synchronous timing fix works mechanically (SCSD-I-CFG2REEMIT now fires 0.84s before the barrier, matching the reference), but a correctly-timed fresh op-0x02 to the coordinator produces no op-0x12 relay, no VAX1 rebuild — barrier still stalls at 5/12, CN=2. So op-0x02 does not trigger the transition-time VAX1 relay; the earlier +18.18 correlation was coincidental. Both parts of this PR (timing + would-be coordinator routing) are moot against a refuted mechanism. Holding as draft pending identification of the real coordinator-internal relay trigger; this branch will be reworked or closed once that lands. Cap-denial passed on every run — anti-fabrication guarantee intact throughout.