diff --git a/oam/librenms-agent/relay/garage-pdu-relay.sh b/oam/librenms-agent/relay/garage-pdu-relay.sh
new file mode 100755
index 0000000..5407439
--- /dev/null
+++ b/oam/librenms-agent/relay/garage-pdu-relay.sh
@@ -0,0 +1,30 @@
+#!/bin/bash
+# garage-pdu-relay: expose APC AP7830 metered output current via snmpd extend.
+#
+# Relay pattern (founder ruling 2026-09-02): the garage PDU speaks SNMPv1
+# only, walks die on its old AOS, and its ACL admits the LibreNMS VM vantage.
+# This relay runs ON tsys-librenms; HA polls the relay via the fleet snmpd
+# extend (same rails as drac-extend). Output: "output_amps <x.y>".
+#
+# Deploy target: tsys-librenms:/usr/local/bin/garage-pdu-relay ONLY (relay
+# host — not a per-host agent). snmpd.conf line on the relay host:
+#   extend garagepdu /usr/local/bin/garage-pdu-relay
+# snmpd ACL must admit the pfv-bms LAN + Tailscale IPs as query sources
+# (rocommunity kn3lmgmt 192.168.3.12 / 100.67.108.125 as of 2026-09-03).
+#
+# Codified 2026-09-03 [#733]: was deployed manually 2026-09-02; hardened
+# with explicit snmpget timeout/retry so a dark PDU answers in ~2s instead
+# of ~6s and cannot pile up long-running children under HA's poll cadence.
+set -euo pipefail
+
+PDU_HOST="${PDU_HOST:-pfv-garage-pdu-1.knel.net}"
+COMMUNITY="${PDU_COMMUNITY:-kn3lmgmt}"
+OID=".1.3.6.1.4.1.318.1.1.12.2.3.1.1.2.1"
+
+raw=$(snmpget -v1 -c "$COMMUNITY" -t 2 -r 1 -On -Ov "$PDU_HOST" "$OID" 2>/dev/null \
+    | grep -oE '[0-9]+$') || true
+if [ -z "${raw:-}" ]; then
+    echo "output_amps unavailable"
+    exit 0
+fi
+awk -v t="$raw" 'BEGIN { printf "output_amps %.1f\n", t/10 }'
diff --git a/questions-v9.md b/questions-v9.md
new file mode 100644
index 0000000..e0af8e3
--- /dev/null
+++ b/questions-v9.md
@@ -0,0 +1,69 @@
+# Questions — v9 (current round)
+
+v8's open items are now settled by the founder's 2026-09-03 chat
+ruling: **OUTBOUND ONLY** — all PFV hosts (k8s included) send via the
+gateway; no inbound ever; no DNAT, no MX cut-over (that scope is
+dropped from #696). Robustness is a stated requirement (VIP pair,
+queueing, cluster, monitoring). Map + target architecture updated in
+place: [Discourse t/322](https://community.turnsys.com/t/322).
+
+## PMG round — resolved
+
+**PQ1 (backend).** Cloudron VPS `mail.knownelement.com`
+(152.53.37.179); public MX already points there; fleet delivers
+direct today; knel.net zone has no MX (and never needs one now).
+
+**PQ2 (inbound path).** ✅ N/A by ruling — outbound only. No DNAT, no
+MX work, ever.
+
+**PQ3 (outbound smarthost).** All hosts through the gateway VIP.
+
+**PQ4 (placement/sizing).** Pair already exists: VM 604 tsys6
+(192.168.1.11) + VM 711 tsys7 (192.168.1.10), PMG 9.1.0, unclustered.
+
+**"Old relay in postfix?"** None — `relayhost` empty fleet-wide.
+
+## PMG round — still open (one-line answers fine)
+
+**PQ5 (native PMG cluster?).** Recommended: yes — config + quarantine
+replication between .1.11/.1.10, managed strictly serial. OK?
+
+**PQ6 (TLS from fleet CA #697?).** Recommended: yes for the two admin
+UIs; SMTP stays opportunistic (all flows are internal + to Cloudron).
+
+**PQ7 (notify address).** Pair already forwards to `coo@turnsys.com`.
+Confirm for PMG alerts + any held-mail notices?
+
+**PQ8 (VIP specifics).** Confirm **VIP 192.168.3.249** + name
+`smtp.knel.net` + dhcpd host-decl pin (VRRP MAC 00:00:5e:00:01:0a)?
+Or name your preferred slot/name.
+
+**PQ9 (outbound cutover staging).** Recommended: serial staged — PVE
+hosts + PBS first (verify a week), then VM postfix set. OK?
+
+**PQ11 (Cloudron relay user — founder side).** Recommended pattern:
+PMG authenticates to a dedicated Cloudron mail user (e.g.
+`pfv-relay@turnsys.com`) over :587 submission, so fleet mail exits
+with Cloudron's rDNS/reputation. Will you create that user (or point
+me at who does)? Fallback if you'd rather not: PMG delivers direct to
+recipient MXs (same egress as today, no deliverability gain).
+
+**PQ12 (k8s submission port).** Recommended: :25 unauthenticated from
+trusted LAN (pod traffic SNATs to node IPs) as the default; optional
+:587+auth later if you want app-level credentials. OK as default?
+
+**PQ10 (side findings — sub-ticket placement).** (a) tsys-ucs-02 root
+alias blackholes to `systemmail@unassigned-hostname...`; (b) Wazuh
+alerting never configured (stock example smtp). Ride-alongs in #696
+or separate sub-tickets? They get tickets either way — just say where.
+
+## Carried from v6 — still open
+
+**Q4. Second human approver in gitea?** "Planned for very near
+future"; until then KNEL/access-roster merges are impossible by
+design. Does not block code/deploy work.
+
+## Superseded
+
+v7 PQ wording and v8 PQ10 ordering are superseded by the rulings
+above; v7/v8 files remain in git history.