test(cluster): spec-5.19 D2 — 4-node reconfig fault matrix (Component A)

spec-5.16 (join GRD/PCM remaster) landed on main, completing the reconfig
band, so the gated Component A now follows: the first 4-node combination
test of the Stage 5 reconfiguration band (5.13 clean-leave / 5.14 fail-stop
/ 5.15 join / 5.16 join-remaster / 5.18 leave-remove). It verifies the
combination at 4 nodes (does not re-prove each sub-spec's own invariants).

Cells (fault x phase), HG#1/#2/#3 — membership+epoch monotone, no split-
master, no false-visible, fenced/removed write fail-closed:
- C1 clean_leave x idle (REQUIRED): leaver commits, coordinator observes
  clean_leave, epoch monotone.
- C2 fail_stop x idle (REQUIRED): survivors detect DEAD, epoch monotone.
- C3 fail_stop x under_write_load (REQUIRED): single coordinator (no split-
  master), survivor write consistent (no false-visible).
- C4 leave_remove x idle (REQUIRED) + HG#3: clean-leave drain -> permanent
  removal accepted (cluster.online_node_removal=on), epoch monotone; removed
  node fail-closes writes ("this node is write-fenced"). The 4-node multi-
  survivor membership-shrink ACK (5.18 Hardening v1.1 survivor-local-apply)
  is substrate-noted, authoritatively proven at 2-node (t/325).
- C5/C6 join + join_remaster x idle (PASS-or-SKIP): online peer-restart
  rejoin substrate-limited at 4-node -> honest SKIP; epoch monotone through
  the absent transition.

Shmem-resilient: on a host whose macOS SHMMNI is saturated by concurrent
test sessions, a cell whose 4-node cluster cannot allocate its SysV
interlock segments environment-SKIPs (L239) instead of failing — never a
faked pass (rule 8.A/8.B). CI runners have ample shmem and run every cell.

nightly: stage5-integrated-acceptance shard extended 327-330 -> 327-331.

Spec: spec-5.19-stage5-integrated-acceptance.md

Author: SqlRush

.github/workflows/nightly.yml

@@ -154,8 +154,9 @@ jobs:
           # + pgbench + crash-recovery) gets its own shard so the multi-node
           # legs run in parallel and do not extend the ges-locking shard
           # wallclock.  t/327 HW workload / t/328 MG-B write perf / t/329 MG-D
-          # heap-ITL WAL measure / t/330 production-bench-subset.
-          - { name: stage5-integrated-acceptance, ranges: "327-330", unit: false, regress: false }
+          # heap-ITL WAL measure / t/330 production-bench-subset / t/331 4-node
+          # reconfig fault matrix.
+          - { name: stage5-integrated-acceptance, ranges: "327-331", unit: false, regress: false }
     steps:
       - name: Checkout
         uses: actions/checkout@v4