vms-6ca: DLM rung H5 — the two-node SCS wire of the async DLM replies (block-then-grant on the requester)

[baron-3dl]

What

DLM epic vms-7fa rung H5 (rd vms-6ca): the two-node SCS wire transport of the async DLM replies, and the requester-side completion — the continuation of the cross-node contention rung (#873, vms-904c).

#873 proved the block-then-grant semantics on a single real /dev/vms with two CSIDs (an incompatible cross-node \$ENQ QUEUES on the master, a cross-node \$DEQ grants it). It left deferred: the requester-side GRANT/queued RECEIVE ops (SS$_UNSUPPORTED) and the actual SCS wire delivery of the async replies to a live remote node. This PR builds that wire.

The outcome (proven on a two-node live A↔B QEMU harness, both real /dev/vms)

Node A holds RESONE EX; a second incompatible \$ENQ QUEUES on node B (the master):

1. B WIREs the queued-reply back to A over SCS → A's requester origin record is genuinely PENDING (SCSD-I-DLMPEND, read from A's own executive via GETLKI → NL), not a local fake.
2. A releases the holder with a real cross-node \$DEQ.
3. B WIREs the deferred GRANT to A over SCS (off the real \$DEQ, SCSD-I-DLMDEFER) → A's origin record FLIPS NL→EX (SCSD-I-DLMH5FLIP granted_mode=EX).

The proof signal is the status flip observed on the REQUESTER node A across the wire — driven only by what the master genuinely sent. No fabricated wire delivery or grant (INV-6).

Layers

• Executive (src/kernel-core/vms_lock.c, shared → Linux + NetBSD inherit):
  • Requester-side GRANT RECEIVE (VMS_DLM_OP_GRANT, was SS\$_UNSUPPORTED): completes an executive-resident origin record on the requesting node — its own list, never touched by the local lock manager; its granted mode is set only from what the master sent over SCS (queued-reply → NL/pending; deferred GRANT → EX). GETLKI falls through to it.
  • Master-side deferred-grant report: vms_lock_dlm_xnode_deq now names the queued cross-node waiter a release flipped to granted (requester CSID + original req_lkid + master handle + mode), so the daemon can wire the deferred GRANT.
  • ABI-stable — reuses the fields a DEQ leaves 0; the vms_dlm_xnode_args struct size is unchanged (no mirror/static_assert churn).
• SCS wire (src/vmsscs/scsd.c): B carries the granted mode honestly (queued-reply = NL) and WIREs the unprompted deferred GRANT off a real \$DEQ; A drives the block-then-grant sequence and dispatches every received reply into its own executive.
• Harness + CI: new Dockerfile.dlm-harness-h5 + init_dlm_h5.sh + run_dlm_harness_h5.sh extend the H4 two-node real-/dev/vms harness; new dlm-harness-h5 CI job. The fast executive-level proof (tests/qemu/test_syssvc_dlm_xnode.c) is extended to assert both halves against a real /dev/vms on x86_64 + Alpha LP64.
• Ledger: docs/compat/facilities/cluster-dlm.yaml gains the async-reply-wire item.

Deferred honestly (INV-6, not faked)

• The BLKAST wire (master → holder) — BLKAST as a receive op stays SS\$_UNSUPPORTED; the block-then-grant is proven without it (the holder releases on its own).
• Multi-peer routing of a deferred grant by CSID (releaser and queued waiter on different nodes) — H5 proves the same-peer case.
• The higher rungs (directory vms-1bba, remaster vms-6ee, LVB vms-d81, deadlock vms-ec75) remain honestly SS\$_UNSUPPORTED.

Local validation

• vms.ko builds clean with all executive changes.
• scsd_exe builds clean; vmsscs unit tests (test_scsd_wire, test_scs_dlm) green — no regression.
• test_syssvc_dlm_xnode.c compiles; H5 shell scripts + ci.yml + cluster-dlm.yaml validated.
• Per Rule 6 the module was not insmod'd on the shared host — the executive + wire are proven green-by-SHA in CI (QEMU syssvc + the two-node H5 harness).

🤖 Generated with Claude Code

[baron-3dl]

Green-by-SHA status + a project-wide CI repair this PR carries

vms-6ca is proven green. On the run for this branch (before the last two housekeeping commits), every vms-6ca job passed:

• Kernel Executive (vms.ko via /dev/vms, QEMU) — all 6 shards: ✅ — the extended test_syssvc_dlm_xnode (requester-side GRANT RECEIVE flip + deferred-grant report) proven on a real /dev/vms.
• DLM Harness H5: ✅ — block-then-grant proven on the REQUESTER node across the live SCS wire (the new two-node harness).
• DLM Harness H1–H4: ✅ (no regression), NetBSD/amd64 vms module cross-compiles: ✅ (shared kernel-core on NetBSD).

⚠️ This PR also repairs a project-wide CI outage (unrelated to the feature)

ci.yml has been failing to parse since #850 (vms-dec, today ~13:01) — a job name: contained an unquoted mid-value colon (… before Username: (vms-dec)), which is invalid YAML. GitHub produced a 0-job workflow on main and every branch — no Build & Test, no harnesses — with only the separate Ledger/CodeQL workflows still firing. The one-line quoting fix restores the whole workflow (105 jobs now schedule). My H5 job could not run until this was fixed.

Two pre-existing failures this repair reveals (NOT caused by this PR, NOT in the DLM domain)

Both were red on main but hidden by the dark CI:

1. dcl-integration → test_lexical_getsyi (F$GETSYI) — the surface-fidelity lane (vms-f5d).
2. Cross-Arch Image Parity Gate → RC3.EXE reappeared x86_64-only — the toolchain/image lane.

I did not fix these — they belong to other lanes with in-flight work, and blind edits there risk breaking them. Flagging for owner triage. (H0's --dlm-selftest staleness — a DEQ probe that #873 turned real — was in my domain and is fixed here by switching the side-effect-free probe to BLKAST, still honestly SS$_UNSUPPORTED.)