F7: office1-tailscale is labelled a subnet router and advertises no routes ...

Raised from an operator report of Permission denied (publickey) reaching voffice1 "from my
workstation via tailscale". Measured rather than assumed, and the premise did not survive.

tailscale is absent from BOTH voffice1 and vcloud -- no binary in any standard location, no
state dir, no interface, service not running. Checked with explicit path probes after an
initial `command -v`, because tool-absent-from-PATH reported as feature-absent is a trap this
repo has hit repeatedly.

The tailnet node is a separate LXD guest INSIDE voffice1: office1-tailscale, RUNNING,
100.64.0.53, site leg 10.10.1.11. Its tailscale status reports AdvertisedRoutes: <none>. So
nothing on the tailnet can reach 10.10.0.20 -- there is no route and there never was.

Two records made the false belief look supported: the ssh stanza labels it "Office1 subnet
router (D-107)", and the operating skill's routing table says the workstation reaches voffice1
over the tailnet, citing D-107 -- which is titled "Airgap posture, per-DC artifact mirror, and
NTP (VR1)" and rules nothing of the sort. A real citation that does not support its claim.

RULED (operator): option (a), ProxyJump through vcloud. Option (b), advertising 10.10.0.0/24
onto the tailnet, was presented and not taken -- it exposes the isolated site net, a posture
change against the airgap subject D-107 actually governs.

Fixed operationally outside the repo: vcloud's ssh config now matches Host voffice1 10.10.0.20,
so the raw IP resolves to the office1 service key with IdentitiesOnly yes instead of falling
through to the id_* defaults, which was the actual cause. Both forms verified, with a negative
control proving the key is not offered to unrelated hosts.

Third wrong record this session, after D-124's MAC-sourcing instruction and the skill citation.
Two of the three surfaced only because something adjacent was being fixed; nothing here gates
prose, so no check could have caught them.

repo-lint 0 fail.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Five Stage-5 findings raised, none executed: R7's per-DC Octavia PKI is half built ...

F1 (HIGH): the D-109 amendment rules per-DC independent Octavia CAs, and the
2026-07-29 work correctly made the CA subject and the VIP gate per-DC -- but $DC
appears in no PATH. phase-01-bundle-deploy.md:293 is a fixed $HOME/octavia-pki and
:473 writes a fixed overlays/octavia-pki.yaml, so generating dc1 after dc0
overwrites dc0's issuing-CA key, controller-CA key and both passphrases, and a
later dc0 redeploy reads the fixed overlay name and applies dc1's CA to dc0 -- the
cross-DC shared amphora root R7 explicitly refused. The existence checks at
phase-01:86 and pre-flight-checks.sh:65 cannot detect it: they ask whether the file
exists, never whose CA it is. The gap is in R7's own "Work implied" list.

F2: creds-matrix.tsv:103-111 declares all nine Octavia credentials `singleton`,
while creds-matrix.py:368-398 enforces both-DC existence only for `per-DC` rows --
so the register behind the BLOCKING P5 gate is structurally blind to a missing
second PKI set. Baseline captured (82 rows, 7 findings, exit 1) so the delta is
attributable. Must land before generation.

F3: both ~/octavia-pki/ and overlays/octavia-pki.yaml are ABSENT (existence only,
no contents read), so F1/F2 are fixable at zero risk right now; the window closes
when the first DC's PKI is generated.

F4: .gitignore:40 is an exact path, so a per-DC overlay rename would un-ignore a
file holding CA key blobs plus a plaintext passphrase, in a repo SEC-004 records as
PUBLIC. The F1 fix must widen the glob in the same commit.

F5: preflight.sh:99 sets DC without exporting it and pre-flight-checks.sh never
reads DC at all, so a dc1 invocation runs a DC-aware P2 beside a dc0-frozen P4
under one combined verdict.

Logged not executed per hard rule 1. repo-lint 0 fail / 615 files.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

GA-R4 session close: bounded bookend + the 2026-07-29 findings sweep ...

Bookend is 14 content lines against the 15-line cap, measured. Ledger 263
lines against the 300 cap, so no rotation is owed this close.

Close sweep captured docs/audit/queued-findings-20260729.txt -- transcript-
only material preserved before compaction, per the queued-findings-20260726
and -20260727 precedent. The substantive content:

- A strict-bash quoting trap FAMILY with three measured instances in one
  session: IFS=$'\n\t' defeats word-splitting on a space-separated string
  (killed pre-flight-checks CHECK 1 silently); a bare  under
  inherit_errexit ABORTS the run instead of reaching a refusal branch; and an
  apostrophe inside the single-quoted shell string wrapping embedded python
  makes bash parse python.
- Two greps that looked like findings and were not: 'grep mac' matches
  type_machine, and grep -c 'macs = [' counts COMMENTS -- the same class as
  the ledger-scan defect where a doc quoting a token inflated a counter.
- Five MAAS behaviours, including that enlistment is NOT commissioning and
  that MAAS auto-configures power only for IPMI machines, so a virsh VM must
  be power-configured out of band first.
- Two guards defeatable by accident: lib_hosts_select_dc's one-DC-per-shell
  guard muted by 2>&1, and NetBox's ?site= filter silently not filtering.
- A process note on two parallel live-ops agents: safety came from agents
  working the voffice1 clone while the orchestrator edited vcloud's, so no git
  race was possible -- at the cost that agents barred from repo writes cannot
  leave captures, and dc1's evidence had to be promoted by hand from /tmp.

The skill sweep stays DEFERRED to the Stage-5 close per the 2026-07-27
operator direction; this capture is its input, not a substitute.

Gauntlet ALL GREEN (87) on vcloud; repo-lint 0 fail / 614 files.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

LIVE: OSD carve + D-104 controllers complete across both DCs (two agents) ...

Operator directed two background agents: one to finish dc0 storage-02/03/04,
one for all of dc1. Both completed.

Eight re-commissions, every one the same single-change diff: vdb at
536870912000 bytes (500 GiB) added; links 12 -> 12; per-interface MACs and
link lists identical. dc0 at 204/196/202s, dc1 all ~182s. dc1's substrate
applied 6 add / 4 change / 0 destroy with MAC drift 0 and re-converged to
ZERO DIFF.

I verified every claim from MAAS directly rather than accepting the reports:
fleet 22 (18 Ready + 2 Deployed + 2 New), 216 links all static, both DCs
blockdevs {1:5, 2:4}, tagcount {2:9}. All matched. I also diffed a raw dc1
before/after pair myself -- exactly the vdb hunk plus one power_state line.

Captures promoted from voffice1:/tmp (which does not survive a reboot) to
docs/audit/osd-carve-20260729/. Both agents were barred from repo writes, so
dc1 would otherwise have had no audit record at all.

Two unprompted agent findings worth keeping: MAAS tags SURVIVED
re-commissioning on all 18 nodes (checked because the snapshot code
structurally could not see tags -- removes a would-be Stage-5 placement
blocker); and lib_hosts_select_dc's one-DC-per-shell guard fires correctly
but can be muted with 2>&1, silently yielding one DC's data labelled as the
other's.

One disclosed judgment call, and it was correct: three dc1 pairs also showed
power_state moving off/error. Agent B declared a STOP, re-measured, then
excluded power_state from its verdict while still printing it, because 8 of 9
dc1 nodes read error BEFORE any mutation. Vindicated -- all 18 role nodes now
read off, including five dc1 nodes never commissioned. A transient MAAS
power-query flap.

Still blocking for both controllers: power_type is EMPTY, the state that
stalled all nine role nodes on 2026-07-20. Now fixable since both are in
lib-hosts at the ruled .5 with MACs pinned.

Gauntlet ALL GREEN (87) on vcloud; repo-lint 0 fail.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

LIVE: canary re-commission closes R1 precondition 3 by measurement ...

Operator: 'Commission storage-01 as the canary'. maas admin machine
commission kghggm skip_networking=1, sysid matched by pinned boot MAC never
by name. Commissioning -> Ready in ~200s, no timeout, no SERVFAIL.

The full before/after diff of status, power, block devices, interfaces, MACs
and links is EXACTLY ONE HUNK: vdb (536870912000 bytes = 500 GiB) added to
blockdevices. links 12 -> 12; MACs identical; links identical. All twelve
static links survive with addresses, modes and subnets intact, including the
D-134 octet .150 mirrored across all six planes in both families -- the v6
host part ::150 being the 2026-07-27 D-136 mirror ruling visible in live
data.

So skip_networking=1 preserves the network EXACTLY while storage is still
re-scanned, which is the entire point. Precondition 3 was previously answered
from the API help text; it is now answered by measurement. The amendment's
concern was legitimate -- the default path does reconfigure networking -- and
the vendor's own control is sufficient.

Remaining: dc0 storage-02/03/04, then all of dc1.
Capture: docs/audit/dc0-canary-commission-20260729.txt

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

LIVE: dc0 substrate applied -- R1 OSD volumes + the D-104 controller VM ...

Operator approval: "Apply dc0 first". Executed from voffice1 via a SAVED
PLAN with a same-session pre-apply re-verify (G8 precedent), so what was
applied is exactly what was reviewed.

Apply complete: 6 added, 4 changed, 0 destroyed. Four libvirt_volume.osd +
the vr1-dc0-juju-01 domain and its disk; the four storage domains updated
in-place to attach vdb.

THE CHECKPOINT HELD -- MAC DRIFT COUNT 0. Verified by virsh domiflist
against lib-hosts' pinned MACs BEFORE anything touched MAAS: all 9 nodes
pinned==live, 6 NICs each. That sequencing is the point. 2026-07-20 became a
fleet-wide outage because MAAS was told to re-commission while its records
were already stale, after a 0/9/0 in-place apply regenerated every MAC
without the plan showing it. So the clean plan was necessary and explicitly
not sufficient. First live evidence that the 2026-07-21 MAC pinning survives
an in-place domain update.

Convergence restored (precondition 1): the dc0 inner root re-plans ZERO
DIFF. MAAS undisturbed -- 20 machines (18 Ready + 2 Deployed), 216 links,
every one still mode=static. MAAS still sees ONE block device per dc0 node,
which is expected and is why the re-commission is genuinely required.

Also recorded: the saved plan's MAC audit returned one hit that was a false
positive -- type_machine = "q35" inside the juju-01 create block, since
"mac" is a substring of "machine".

NOT DONE, gated separately: the re-commission with skip_networking=1, and
dc1 entirely. Capture: docs/audit/dc0-osd-juju-apply-20260729.txt

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R1 precondition 3 VERIFIED read-only from voffice1: skip_networking=1 is the control ...

The amendment required verifying, BEFORE the apply, that re-commissioning
preserves the D-134 statics and the pinned MACs. Both halves now answered
from the headend, read-only, no mutation.

HALF 1 -- baseline measured: 18 Ready nodes, 7 interfaces, 12 links each,
mode=static on every link (6 planes x 2 families). MAAS 3.7's commission API
carries skip_networking: 'Whether to skip re-configuring the networking on
the machine after the commissioning has completed.' The option exists BECAUSE
the default reconfigures, which is exactly the flagged risk. The combination
R1 needs is skip_networking=1 with skip_storage UNSET -- network preserved,
storage re-scanned so the new /dev/vdb enters inventory.

HALF 2 -- the tofu change does not touch NICs. Both inner roots plan
6 add / 4 change / 0 destroy; the four in-place diffs change devices.disks
ONLY, gaining vdb. A grep for mac/interface across the whole diff returns one
hit and it is inside the NEW juju-01 create block. Risk is materially lower
than 2026-07-20 because MACs are now config-pinned -- the control introduced
after that incident precisely because the plan was untrustworthy.

NOT overclaimed: a clean plan was necessary but not sufficient in 2026-07-20,
when 0/9/0 in-place regenerated every MAC. So the sequence puts a MAC
verification BETWEEN the apply and the re-commission -- that incident became
fleet-wide because MAAS was told to re-commission while its records were
already stale.

Capture: docs/audit/r1-precondition3-verify-20260729.txt

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

LIVE: node IPv6 carve both DCs -- 108 links, IPv4 untouched, 18 Ready unchanged ...

Operator-gated dc0 then dc1. 54 applied per DC, 0 errors, read-back 54/54 each.
IPv6 links 0 -> 108; IPv4 links 108 unchanged; 18 Ready unchanged. Both
dc-node-v6-carve gates PASS and dc-plane-ipam still reads 24/24 on both DCs.

The octet mirror holds across all six planes per node -- superb-piglet .121 ->
::121 everywhere, big-trout .100 -> ::100. enp2s0 correctly received nothing: the
D-100 raw provider NIC has no v4 link to mirror, and br-ex carries provider-public
instead, on the node /64 rather than the VIP /64. That falls out of deriving from
"already carries v4" rather than naming planes in a table.

The gate discriminated rather than agreeing with what it found -- it flipped dc0
to PASS while dc1 still read FAIL, before dc1 was carved.

This closes the node half of step 3, which was the correction the operator's
octet-mirror question surfaced: v6 does not arrive with the subnet, because MAAS
mode=static means an explicitly configured address.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

D-101 governing rationale recorded; node v6 carve scoped; two of my claims corrected ...

The operator's question -- "are we mirroring the assigned IPv6 octet with the last
IPv4 octet like we did before?" -- exposed two errors in what this session recorded.

(i) I said MAAS static assignment "needs only the subnet, which now exists". Wrong:
mode=static means an EXPLICITLY CONFIGURED address, which is how the Stage-4 carve
set the 90 v4 links. Measured -- all 18 Ready nodes carry ZERO IPv6 links, while
their v4 side is correctly octet-mirrored. So VIPs are mirrored (156 apex objects)
and nodes are not; 108 assignments are owed. (ii) v6 gateway_ip and dns_servers are
unset on all 12 v6 subnets, unnamed before I called step 3 complete. That
"complete" is withdrawn -- the MAAS/apex/lib-net population stands, the node layer
is the remaining half.

D-101 gains its GOVERNING RATIONALE, quoted verbatim, because it existed in no repo
surface: IPv6 unless IPv4 is NECESSARY, driven by real IPv4 sizing constraints in
future expansion -- a commercial requirement, not a preference. It records that
v4-first was DELIBERATE RISK REDUCTION so a future session does not read v4
surfaces as neglect and "correct" them; that Roosevelt has full v4 and v6 edge
transport; and that NAT64/DNS64 was considered and REJECTED on that basis -- a shim
with no Roosevelt analog, against ULA planes that are internal by design.

Node v6 carve scoped in docs/audit/node-v6-carve-scope-20260727.md: 108 assignments
mirroring each node's live v4 octet, prior art measured v4-only, blast radius
per-link and reversible, and the operator's own rationale arguing to carve BEFORE
the deploy since v6-unfriendly modules surface far more cheaply on a static
read-back than mid-bundle.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

LIVE: D-134's bands now EXIST -- both DCs at pass=24 fail=0, from a 6/18 baseline ...

Operator-gated, dependency-ordered: dc1 v6 carve (6 subnets), dc0 reserve (13
ranges), dc1 reserve (12 ranges). Each pre-apply re-verified in the same session;
every write read back. Subnets 18 -> 30, ipranges 3 -> 28. Machines 18 Ready + 2
Deployed unchanged throughout -- no node, no tfstate, no running service touched.

This closes the "RULED IS NOT BUILT" finding for D-134: its band table had been
prose since 2026-07-23 while MAAS held zero reserved ranges.

FINDING -- MAAS already reserves the entire low IPv6 block. Every explicit v6 band
create failed with "Requested reserved range conflicts with an existing range".
Measured via `subnet reserved-ip-ranges`: MAAS auto-reserves ::1-::ffff:ffff on
every IPv6 subnet plus :: per RFC 4291 s2.6.1, with allocatable space starting at
<prefix>:0:1::. The ruled bands sit entirely inside it, so the write is impossible
AND unnecessary. R4's "v6 bands as a second pass" is not executable in MAAS and
need not be; the tool now verifies coverage instead of writing. Execution-level
correction -- R4's intent is satisfied. Same on ULA and GUA.

Harness T19 asserted a literal string the fix removed. Re-pointed at the surviving
invariant (textual 4-99, never hex 4-63) rather than deleted, per the standing rule
against going green by removing an assertion. 25/25, gauntlet ALL GREEN (83).

dc1's FIP pool still correctly REFUSED -- unset by design, so R4 cannot close for
dc1 until ruled.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

LIVE: dc0 v6 plane carve applied -- 6 MAAS subnets, each read back ...

Operator-gated ("Run carve-v6 vr1-dc0 --commit"). dc0 only; dc1 not authorised
and not touched.

Pre-apply re-verified in the same session per the G8 precedent -- plan unchanged
at 6/1/0 -- then applied: 6 applied, 1 skipped, 0 errors, every create READ BACK
on its intended vlan. MAAS subnets 18 -> 24, v6 1 -> 7.

Each v6 plane landed on the SAME vlan as its v4 twin, so the plane is genuinely
dual-stack on one L2 rather than a parallel fabric. The provider GUA VIP /64 was
deliberately not created -- it holds hacluster-managed API VIPs, not node
addresses, so MAAS never allocates from it.

Idempotency proven LIVE rather than only in fixture: --commit ran twice (the
second to read the script's true exit code rather than a pipeline's) and the
post-state carries zero duplicate CIDRs.

Nothing else moved: 18 Ready + 2 Deployed unchanged, no node touched, no tfstate
involved, dc1 still measures 6 v6 planes absent. The gate now reports dc0's six
v6 planes [ok]; its remaining reds are the 12 unreserved D-134 bands.

Changelog item 9 carries the revert, with the 2026-07-21 pod-delete cascade
lesson attached: read occupancy BEFORE deleting a subnet.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Step 3: dc-plane-ipam.sh check -- D-134's missing executable gate, with a measured baseline ...

R4's D-134 amendment ruled a site-keyed tool that "also gives D-134 an EXECUTABLE
gate instead of prose". This is that gate. Read-only half only; the mutating
actions are deliberately not in this commit (hard rule 3 -- the gate ships and
runs first, so the mutation is planned against a MEASURED baseline).

Expected state is derived, never hardcoded: v4 planes from lib-net's DC selector,
v6 planes from the NetBox apex record (D-136 option (D) applied -- the apex is the
source, so this does not create a second hand-maintained table), bands from the
D-134 2026-07-23 amendment.

LIVE BASELINE, both DCs, symmetric at 6 pass / 18 fail: all six v4 planes present;
ZERO of six v6 planes present; ZERO of twelve D-134 bands reserved. MAAS holds 17
v4 subnets and exactly one v6, which is Office1's -- a third-layer confirmation of
U17. Cloud-wide there are 3 ipranges, all dynamic, matching R4's measurement.

Proven able to BOTH fail and pass: every live run fails because everything it
asserts is absent, so T6 drives a fully-provisioned fixture to green. A gate only
ever observed failing is as untrustworthy as one only ever observed passing. It
REFUSES rather than reporting clean on an unreachable MAAS, an unreadable apex, or
a band of unrecognised type, and distinguishes an ABSENT maas binary from an
unreachable MAAS -- the misdiagnosis class this audit found three times.

The provider GUA VIP /64 is reported but NOT asserted as a MAAS subnet: it holds
hacluster-managed VIPs, not node addresses, so MAAS cannot hand one out. Whether it
should also be a MAAS subnet is left open rather than decided by a gate.

Harness 14/14. Gauntlet ALL GREEN (83) -- the R15(2) manifest gate correctly caught
the new harness as unpinned drift and was re-recorded deliberately, which is the
mechanism working end to end in its first real use.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

VIP/band reconciliation measured; two ruled decisions found to have an unruled gap ...

Step 4 prep for the render pipeline. Read-only enumeration of the four
disagreeing VIP surfaces, with every consequential claim re-verified directly
before recording. Nothing adopted.

NEW GAP: R11 ruled three gate changes (band .60->.99, VIP_COUNT_EXPECT 11->13,
FAIL on hacluster-without-vip) but NOT the arity change R2's dual-stack forces.
provider-bundle-check.py:137 requires exactly 3 addresses; a dual-family vip is
6, so under R2 the checker fails every application. :149's octet extraction also
returns the whole address string on a v6 literal, and a v6 address tests False
against an IPv4Network without raising -- so a naive arity fix just relocates the
failure. Two RULED decisions whose combined end state the gate cannot express.

TRAP: EXPECT_PUBLIC_VIP must STAY 11 while VIP_COUNT_EXPECT goes to 13. Measured
against bundle.yaml -- neither vault nor designate carries a `public` binding, so
neither joins that count. Reading "11 -> 13" and bumping both breaks the gate.

The L3-9 overlay collision is worse than recorded and the dangerous merge order
is the GREEN one: vips-last exits 0 while dropping every IPv6 leg, and
prefer-ipv6: true survives as a separate key -- charms would bind :::port with no
v6 VIP for pacemaker. Measured through the checker's merge MIRROR, not juju
(absent here); flagged for --dry-run confirmation rather than asserted.

CORRECTED one of the agent's own conclusions: dc1 "inheriting dc0's band bounds"
is not a defect. The octet band is DC-invariant by design, so lib-net.sh:157
correctly unsets the prefixes and keeps the bounds. Acting on the uncorrected
framing would have invented per-DC bounds that do not exist.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Phase 0 delivery cleanup: session changelog with reverts, and a mis-filed finding corrected ...

Three gaps in the Phase 0 delivery, plus one new measured finding.

1. SESSION CHANGELOG ADDED (GA-R2/D1) -- docs/changelog-20260727-stage5-phase0.md.
   It was missing, and it is the only revert surface: no logged window was
   opened for this session, so the changelog + capture are the entire
   as-executed record of five live mutations (two branch deletions on origin,
   two on voffice1, a 68-package install on the region host). Each item now
   carries WHAT / WHY / HOW TO REVERT with the pre-state hashes recorded
   (39e8988, 61c416e, 57836b1) and the apt purge scoped with a caution that
   68 packages were pulled and some are shared.

2. CURRENT-STATE section 7 no longer carries two rows for one component. The
   prior "ABSENT ON BOTH HOSTS" state is folded into the single pin row as
   history, following the Juju row's in-row supersession precedent. A pin
   table asserting both presence and absence of the same client is the same
   stale-surface class the audit close-out sweep found the audit itself
   creating.

3. A FINDING I MIS-FILED IS CORRECTED BY MEASUREMENT. P4's
   "metal-admin gateway=none (want 10.12.8.1)" was dismissed as covered by
   readiness item 3.7. It is not -- 3.7 is specifically the VID-103 /
   br-internal assertions. Measured from the dc0 rack: 10.12.8.1 answers 0/2
   pings with an INCOMPLETE ARP entry (never resolved -- nothing holds it),
   against a control ping to the edge at 10.12.4.1 at 0% loss. So MAAS is
   right to carry no gateway and lib-net.sh:37's expectation is the defect, a
   VR0 inheritance; the D-134 carve ruled .1 gateways for the PROVIDER subnets
   only. Logged as P0-5, not fixed (hard rule 1). The dc1 arm (lib-net.sh:149)
   carries the same shape at 10.12.68.1.

Also queued for one GA-R5 exchange, no D-number assigned: which host is
authoritative for the gates. R10 (preflight on voffice1) and R15(3) (preflight
stops failing open) executed together make preflight-on-voffice1 permanently
unpassable given P0-2, and R15(2)'s manifest does not reach P0-1 -- so the R15
execution scope needs this answered first.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Stage-5 Phase 0 EXECUTED: clone advanced, refs pruned, openstack client installed ...

Readiness-doc preconditions 0.1/0.2/0.3, operator-approved step by step
("Approve A, B, and C, and retire the audit branch"). NO STAGE OPENED.

0.1 voffice1 is at main (HEAD == origin/main == 6495cfb, asserted). The real
risk was that both DCs' inner tfstate -- the substrate's state-of-record --
lives inside that working tree; every state artifact was proven gitignored and
origin/main proven to track an identical file set at those paths BEFORE the
switch. Both tfstate sha256s are byte-identical after. The two dc1 overlays are
now present and bundle.yaml is the 9-node role-separated layout.

0.3 three stale remote-tracking refs pruned, two stale local branches deleted
(the record named one), containment proven first. The audit branch was retired
on origin BEFORE the fetch so the clone could not be handed a fresh stale ref.

0.2 python3-openstackclient 6.6.0-0ubuntu2 on voffice1, verified behaviourally.
The snap was refuted by measurement: no Caracal channel (newest stable zed,
2023), plus the home-only confinement already recorded at design-decisions:638.

Two NEW findings from the payoff runs, logged not fixed (hard rule 1):
- the gauntlet is HOST-DEPENDENT (2/81 on voffice1 vs ALL GREEN on vcloud, same
  commit); both failures are host-portability defects, one of them an aggregate
  FAIL over all-passing sub-checks
- preflight P5 is HOST-BLIND (34 findings vs 7), and between the two hosts there
  is no single host on which preflight is currently correct

R10's ruled consequence is confirmed: P3 now verifies all 33 charm-channel pins
(previously zero) and P4 reports MAAS reachable.

Capture: docs/audit/stage5-phase0-20260727.txt

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Session-close sweep: three stale surfaces the audit itself created, plus transcript-only facts ...

Operator direction before the bookend: "complete a session sweep for anything
missed that should be committed that would be lost on session end". Precedent:
queued-findings-20260726.txt and -20260727.txt. Capture:
docs/audit/queued-findings-20260727-stage5-audit.txt.

THE SWEEP FOUND THREE STALE SURFACES THIS AUDIT ITSELF CREATED, which is the
exact defect class it was convened to find:

1. docs/audit/stage5-readiness-20260727.md -- the "read this first" artifact --
   still carried THIRTEEN NEEDS-RULING markers after every one of R1-R15 had been
   ruled or withdrawn. At that moment it was the single most misleading surface in
   the repo. Fixed with a supersession banner that names the three authorities in
   order and states what in the document is STILL true and worth reading: the
   ordered precondition sequence, the evidence column, the what-breaks column. The
   row-level markers were deliberately NOT rewritten -- they are the record of what
   was owed at the time, and rewriting them would destroy that record.

2. queued-rulings-20260727.md still opened with "Nothing here is adopted" after
   fourteen adoptions. Fixed with a STATUS block naming design-decisions.md and
   CURRENT-STATE.md as the ruling authorities rather than that file; the original
   sentence is struck through, not deleted.

3. SIX sections of that file had BLANK OPERATOR UTTERANCE lines -- R8, R9, R10,
   R12, R13, R15 -- for rulings that WERE properly recorded in design-decisions
   and CURRENT-STATE. GA-R5 was satisfied; the question sheet was not, and a
   future session reading only that file would have believed six questions were
   still open. All six filled with the exact utterance and where the ruling lives.
   R14-ORIGINAL's blank line is CORRECT and left alone -- it is superseded text
   kept for the audit trail.

TRANSCRIPT-ONLY OPERATIONAL FACTS NOW CAPTURED: the repo-lint L5 heading trap (a
heading LEADING with a D-number counts as a second definition unless the line also
contains AMENDMENT or RESOLVED -- it bit this session twice and is on no
author-facing surface); the commit-gated-on-lint discipline adopted after a red
commit was pushed, which then caught the second occurrence before it landed; the
read-only live-apex poll procedure that keeps the token in env and prints nothing;
that the office1 VMs answer from vcloud but NOT via voffice1, so "unreachable" from
the wrong host must never be recorded as "absent"; and that creds-matrix.tsv is
SPACE-aligned despite the .tsv extension, so awk -F'\t' returns zero rows and looks
like a real result.

META-FINDING RECORDED FOR THE NEXT COMMITTEE: the lenses' OBSERVATIONS were
reliable, their CONCLUSIONS repeatedly were not. Three of this audit's own premises
needed correcting by measurement -- R2a (literals already assigned under D-111), R9
(one failure mode not two, two scripts not twenty), R14 (the register already
attributed and explained the findings, and already warned against the suppression
the question contemplated) -- and R8's entire evidence base collapsed on reading
the cited bugs. A lens finding is an OBSERVATION, not a conclusion; measure before
putting it to the operator as a question.

VERIFIED DURING THE SWEEP, after an API disconnect mid-turn: nothing was lost or
corrupted. The earlier lens-5 API casualty was fully recovered -- its 18 findings
and section header are in the committed record, and all seven lenses are present.
The disconnect this turn fell cleanly between edits and a commit; the two modified
files parsed intact and lint was red only for the L10 this commit satisfies.

NOT DONE AND NOT OWED YET: the skill sweep. This session closed no STAGE, so the
stage-close fold-in and snapshot regeneration are not due. B1/C1/C2 in the capture
are the candidates when Stage 5 closes.

Revert: git revert this commit; the capture is new and the two repairs are
additive (banner, status block, filled utterance lines).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R14 WITHDRAWN: the register already does everything the question assumed it could not ...

Withdrawn without a ruling, following the R2a precedent. R14 asked whether the
credential matrix needs a ruled-exception field, on the premise that three S5
power-key asymmetries were "ruled correct by SEC-016" and so reported a permanent
red a reader learns to ignore.

All three parts of that premise fail on measurement:

1. They are NOT ruled correct. SEC-016 ruled per-DC ISOLATION -- dc1 gets its own
   dedicated power key rather than reusing dc0's -- which is satisfied. It never
   blessed the filename, host and custody divergence. That is SEC-021(b), an OPEN
   ledger defect whose own disposition reads "needs a naming/custody reconciliation
   to the dc1 shape", and whose stated complaint was "Per-DC rows that should be
   symmetric are not, and nothing compares them". S5 is the thing that now compares
   them, so the finding is the register working as designed.

2. The register already ATTRIBUTES them: those rows carry sec-ref=SEC-021 and
   notes-ref=n-dc0-power-key-divergence.

3. The register already EXPLAINS them: creds-matrix-notes.md carries the
   n-dc0-power-key-divergence note describing the name and host-role divergence and
   why the expected jumphost rows fail EXPECTED-BUT-ABSENT.

And the notes file already warns against the exact move this question contemplated.
The line immediately preceding that note reads: "failure, not a matrix error: do
not delete the row to make the checker green."

So no schema change is needed and none should be made. Adding a suppression
mechanism would have HIDDEN an open security-ledger item -- the opposite of what
the register exists for. The red clears when SEC-021(b) is remediated, which is the
intended behaviour.

This is the third premise this audit has had to correct in its own framing, after
R2a (literals already assigned under D-111) and R9 (one failure mode, not two).
Worth noting the pattern: the audit's lens findings were sound as observations and
repeatedly wrong as conclusions, and measurement caught it every time.

Residual recorded so it is not mistaken for an oversight: whether a ruled-exception
field is EVER needed is now hypothetical. If one arises, notes-ref and sec-ref are
the place to start, not a new column.

Revert: git revert this commit; the R14 entry is rewritten in place with the
original retained beneath as R14-ORIGINAL for the audit trail.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Final pre-ruling measurements R9/R10/R12-R15: two premises corrected ...

Operator direction: gather everything remaining so the last decisions can be
worked through without stopping to measure. Capture:
docs/audit/r9-r15-final-measurements-20260727.txt.

TWO OF THESE CHANGED MATERIALLY, and both are corrections to the audit's OWN
earlier framing rather than new findings.

R9 -- there is only ONE failure mode, not two, and the Stage-5 blast radius is
TWO scripts rather than twenty. lib-net.sh:76-79 states the design explicitly:
sourcing without the selector "continues to populate PLANE_CIDRS ... exactly as
above (VR0/DC0's real, measured values) -- every existing script that sources
lib-net.sh keeps working completely unchanged". So the unset block fires ONLY
for selector-CALLERS. The 20 non-selector consumers fail SILENTLY with dc0
literals; the 8 correctly-updated ones are the ones that break loudly under
set -u. That inverts the intuitive reading -- the scripts nobody fixed are the
dangerous ones. And grepping the two runbooks Stage 5 actually uses, only
phase-03-core-verify.sh and deploy-watch.sh are invoked there; the rest bite at
later stages.

R14 -- THE PREMISE IS WRONG and the question largely dissolves. I framed it as
"the matrix cannot express a RULED exception, so three S5 asymmetries ruled
correct by SEC-016 report as a permanent red". Measured: they are NOT ruled
correct. SEC-016 ruled per-DC ISOLATION -- dc1 gets its own dedicated power key
rather than reusing dc0's -- which is satisfied. It never blessed the filename,
host and custody divergence. THAT is SEC-021(b), an OPEN defect whose own
recorded disposition reads "needs a naming/custody reconciliation to the dc1
shape", and whose stated complaint was "Per-DC rows that should be symmetric are
not, and nothing compares them". S5 is the thing that now compares them. The
register is RIGHT and the finding is REAL. Suppressing it would have hidden an
open security-ledger item.

R15 refinement: the gauntlet ALREADY has a zero-floor -- run-tests-all.sh:35
exits 2 when RAN is 0. What is missing is a MINIMUM-count floor, since 81 is
pinned nowhere executable. repo_lint.py:132-135 has NEITHER a valid-root check
nor a files-scanned floor. The two gates need different fixes, which the earlier
framing conflated.

R12 and R10 confirmed as previously stated, with exact line citations.

Revert: git revert this commit; the capture is new and CURRENT-STATE additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

G18 OPENED: IPAM apex completeness for the Octavia lb-mgmt plane (blocking, deferred) ...

Operator direction, verbatim: "leave this as an open decision that will need a
ruling once we have the cloud live and we have a better read on the network and
how everything is functioning with the addition of the new IPv6 configurations.
Make this a gated decision so we cannot close the project (or whatever phase you
think it best ruled in) without a ruling on this item."

WHAT IT IS. R8 ruled that Octavia creates and owns its own IPv6 lb-mgmt network.
The measured consequence is that the octavia charm exposes NO CIDR, address-family
or router configuration option -- create-mgmt-network (default True) is the only
related option -- so the prefix is CHARM-GENERATED and cannot come from the D-111
carve. NetBox is therefore knowingly incomplete for exactly one plane. That is the
authority-inversion concern the audit's lens 7 raised: the apex being back-filled
to match a deploy rather than driving it.

WHY DEFERRED RATHER THAN RULED NOW. It is not answerable from artifacts -- it
needs the cloud live and an observed read on IPv6 behaviour. It is also adjacent
to the UNRULED D-136 NetBox-coupled render pipeline, which covers the same
apex-authority ground, so ruling it early would pre-empt that decision.

PLACEMENT. Recorded as gate G18 in CURRENT-STATE section 6, the gate authority.
[R] ruling-type: closes ONLY on a GA-R5 recorded ruling with the exact utterance,
dated, committed and pushed. ANSWERABLE from Stage 5 onward, since the prefix
exists once Octavia deploys. BLOCKING at the FINAL stage close / project close --
the deployment may not be declared complete while it is open. That placement gives
the widest window in which live evidence can accumulate while still guaranteeing
the question cannot be lost.

Options are pre-recorded in the gate row so a future session does not have to
re-derive them: (a) back-fill the charm-created prefix into NetBox post-deploy as
a documented record; (b) record the plane as charm-owned and explicitly out of
apex scope; (c) fold it into D-136's render-pipeline ruling if that is taken first.

ALSO CROSS-RECORDED, because it is exactly the kind of thing that gets "fixed" by
mistake: the absence of an lb-mgmt :x80 prefix in the VR1 ULA carve is CORRECT
under R8, not a gap. Noted in the D-101 R8 ruling note and in the G18 row.

Mirrored into docs/audit/queued-rulings-20260727.md as R16 (gated, deferred) with
CURRENT-STATE named as the authority, so the queue and the gate table agree.

Revert: git revert this commit; the gate row and cross-references are additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R8 RULED: Octavia creates and owns its own IPv6 network (D-101 note; sub-ruling closed) ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "I want to take a Octavia creates and owns its own IPv6
network." Supporting reasoning recorded because it is load-bearing: "We have to
research to troubleshoot if we run into MTU bug issues down the road. We have
already deployed using this topology in the v1 DC test deployments that got us to
this point."

This CLOSES the octavia-family sub-ruling that the 2026-07-25 D-101 note left
explicitly open.

THE RULING REQUIRES NO ARTIFACT CHANGE, and the operator's reasoning checks out on
the artifacts: create-mgmt-network is set NOWHERE in bundle.yaml or any overlay, so
the charm default True has always applied, and Octavia's whole options block is
debug, openstack-origin, amp-image-tag, vip. VR0 deployed Octavia on precisely this
shape. This ratifies the existing topology rather than changing it.

FAMILY AGREES THREE WAYS. The charm's default lb-mgmt-subnet is IPv6 (LP #1897418,
verbatim: "By default, Octavia charm uses ipv6 for its lb-mgmt-subnet") and it is a
ULA -- the amphora address quoted in LP #1911788 is fc00:fa21:3d5c:9cfd:..., i.e.
fc00::/7. That is exactly D-101's "IPv6-only ULA ... Octavia lb-mgmt ... Internal,
no external clients".

THE APPARENT VR0 GUA CONTRADICTION WAS A PAPER ALLOCATION, checked on operator
instruction: lib-net.sh gives VR0 six IPv4 planes and no lbaas plane; lbaas is
listed in STALE_SPACES; maas-as-built-reference records that NIC as "idle
(undefined; ex-lbaas), raw NIC, no link"; and D-101's own context says VR0 is
IPv4-only. The whole VR0 v6 tree in the apex is a design record.

CONSEQUENCE FOR D-111, worth recording so it is not later "fixed": the absence of
an lb-mgmt :x80 prefix in the VR1 ULA carve is CORRECT, not a gap. The charm
exposes no CIDR option, so the prefix is charm-generated and cannot come from the
apex.

STANDING OBLIGATION carried by the operator's own reasoning: LP #2018998 (o-hm0 vs
lb-mgmt-net MTU, charm-octavia, High) is Fix Released across our lineage but
recurred 2025-12-31 against octavia 14.0.0 / 2024.1 stable -- our exact pin. A
jumbo lb-mgmt-net beside a 1500 o-hm0 silently drops health messages over 1500
bytes and causes spurious failovers, which is the direct interaction with R3's
jumbo-underlay ruling. Owed at the Octavia step: verify o-hm0's MTU MATCHES
lb-mgmt-net's by measurement, not by trusting the charm.

NOT ASSERTED: whether the charm attaches an external gateway to the Neutron router
it creates. No config surface tells it to and ULA is not globally routable, but
isolation was not proven from the documentation. Settle by inspecting the router at
deploy time; queued as an observation, not a blocker.

Revert: git revert this commit; the ruling note and research sections are additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R8 research: both in-repo citations fail; charm default is IPv6; new MTU risk found ...

Operator declined to rule on the overlay's own citations -- "I would like
documentation and vendor specific information rather than conjecture. Research is
cheap, guesswork is expensive" -- then directed a full read of two bugs weighed
against current versions. That was the right call and it inverted the answer.

BOTH CITATIONS IN overlays/dc-dc-ipv6-family-matrix.yaml FAIL ON INSPECTION:
- LP #1913409 is Fix Released (2021) against kolla-ansible, a DIFFERENT
  INSTALLER. No bearing on a Juju/charm deployment.
- LP #1911788 is Incomplete and a DUPLICATE of LP #1896630, and is NOT an IPv6
  defect. Its diagnosed cause is an OVN port-binding hostname mismatch --
  binding_host_id carrying the shortname instead of the FQDN for LXD containers
  on MAAS, giving binding_failed instead of ovs. The IPv6 address in the timeout
  was incidental; the port never bound, so nothing would have worked over any
  family.

VERSION WEIGHING, verified rather than asserted. #1896630's primary fix is
ovs-record-hostname.service, which the bug places in OVS 2.15 (Focal shipped
2.13). Rather than assume, I queried THIS DEPLOYMENT'S OWN dc0 mirror for what a
jammy node would actually install: openvswitch-switch 2.17.0-0ubuntu1 (jammy) and
2.17.9-0ubuntu0.22.04.2 (jammy-updates). 2.17.9 >> 2.15, so the fix is present on
every node this deployment will build. The bug's era was Juju 2.7.8, charms 20.08
Ussuri, Focal/OVS 2.13; we are Juju 3.6.27, charms 2024.1, jammy/OVS 2.17.9. The
charm-layer-ovn task -- the one closest to this repo's OVN usage -- is closed
Invalid.

AND THE DEFAULT RUNS THE OTHER WAY: LP #1897418 records, verbatim, "By default,
Octavia charm uses ipv6 for its lb-mgmt-subnet", with a reporter noting success
"using the charm default (ipv6) network but using ipv4 everywhere else". Upstream
Octavia documents that "IPv6 subnets can be used for the LB Network". So D-101's
IPv6-only placement of lb-mgmt is ALIGNED with the charm default, and a v4-only
lb-mgmt-net would be the DEPARTURE requiring justification -- the opposite of how
R8 was first framed.

NEW LIVE RISK FOUND, COUPLED TO THE R3 RULING: LP #2018998 "MTU mismatch between
o-hm0 and lb-mgmt-net" (charm-octavia, High, Edward Hope-Morley). A jumbo
lb-mgmt-net beside a 1500 o-hm0 silently drops health messages over 1500 bytes
and triggers SPURIOUS load-balancer failovers -- evidenced in the bug as "UDP,
length 1534 > 1500 and o-hm0 never receives them" and "o-hm0: dropped over-mtu
packet: 1744 > 1500". Fix Released across our lineage, BUT a recurrence was
reported 2025-12-31 against octavia 14.0.0 / 2024.1 stable, the exact channel
bundle.yaml pins. R3 ruled the underlay be finished to jumbo, which is precisely
this bug's precondition. OWED at the Octavia step of Stage 5: verify o-hm0's MTU
MATCHES lb-mgmt-net's after deploy, not merely that the charm claims to set it.

GENERALISABLE LESSON, recorded in the capture: a bug NUMBER in a comment is an
existence claim; what the bug SAYS is the content, and only content is evidence.
Same class as the two ruled-but-never-built decisions this audit already found --
a plausible in-repo statement that nothing verifies.

R8 is NOT ruled by this commit. It will be re-presented on corrected evidence.

Revert: git revert this commit; the capture is new and CURRENT-STATE additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R7 RULED: per-DC independent Octavia PKI (D-109 amendment) ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "Per-DC independent Octavia PKI; fix the generator first
(Recommended)".

Recorded as a D-109 AMENDMENT. D-109 established per-DC INDEPENDENT Vault roots
with a DR-honest rationale but never mentioned the Octavia amphora
control-plane PKI, which is a SEPARATE trust domain with its own CA generated
outside Vault by phase-01 step 1.0-GEN. This extends the same posture to it.

MEASURED REFINEMENT THAT NARROWS THE WORK. The generator is dc0-frozen in two
ways of DIFFERENT severity, and only one is a design problem:
- The CA SUBJECT is a baked literal, "/CN=VR0 DC0 Omega Cloud Octavia Controller
  CA/O=Neumatrix". Reusing it roots dc1's amphora chain in a VR0-DC0-named CA.
- The controller cert's SAN is ALREADY DERIVED per-DC by design -- the step
  states it carries the controller FQDN, octavia API FQDN and the Octavia API VIP
  "derived from the bundle at generation time -- DOCFIX-067; never a baked
  literal". That half needs no change at all.
- The actual blocker is the VIP gate, grep -qE '^10\.12\.4\.', which accepts only
  dc0's band while dc1's octavia VIP is 10.12.64.57 and lives in an overlay.

A DOCFIX was owed regardless of the ruling: the generator cannot produce a dc1
artifact as written, and phase-01:144-145 hard-ABORTS the deploy when
overlays/octavia-pki.yaml is absent, which it is. Option (c) would therefore have
deferred only the part D-109 already answers by precedent while leaving the
required work untouched.

REUSE REFUSED ON POSTURE. The overlay carries CA private keys plus a plaintext
issuing-CA passphrase inside the repo clone, and SEC-004 records the repo as
still PUBLIC. Sharing one amphora control-plane CA private key across two clouds
D-100 defines as independent would widen an existing exposure rather than
contain it, in a commercial multi-tenant cloud with hard tenant isolation.
Independent per-DC CAs keep that blast radius to one DC.

CAVEAT CARRIED FORWARD, because it will bite otherwise: the generator's VIP gate
must be re-pointed at the MERGED deploy input rather than bundle.yaml. Left
reading the base bundle it breaks again the moment ruling-3's VIP extraction and
R11's new .61/.62 allocations land.

Roosevelt analog: per-DC amphora CAs match per-DC IPMI/BMC credentials and the
per-DC MAAS power keys of SEC-012/-016 -- the same no-cross-DC-shared-secret
principle already ruled twice in this deployment.

Revert: git revert this commit; the amendment is additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R11 RULED: vault .61 and designate .62, dual-family triples (D-020 amendment) ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "Both full triples (.61 vault, .62 designate), dual-family,
and fix the gate (Recommended)".

THE FINDING THAT MEASURING FIRST PRODUCED: vault was ALREADY RULED and never
built. D-020's decision text enumerates vault BY NAME among the clustered
applications carrying both a provider and a metal VIP; measured, base vault is
num_units:1 with an EMPTY options block. This is a conformance repair of a
2026-era decision, not a new choice -- and it is the SECOND ruled decision this
audit has found unimplemented, after D-134's address bands (R4). Worth stating
plainly: the pattern is ruled-but-never-built, and nothing in the repo could
detect either case.

designate is genuinely new -- absent from D-020's enumeration -- so it needed a
ruling rather than a repair, and this amendment adds it. Its dnsaas endpoint is
already bound provider-public, so a provider leg is coherent for it.

Shape is the ESTABLISHED triple, not a new form. The measured octet map is
consecutive: keystone .50 through ceph-radosgw .60, all provider/admin/internal
triples. vault takes .61, designate .62, both dual-family per R2 -- adding
v4-only now and re-doing them later would mean re-issuing certificate SANs on a
live cloud, and for vault that is 23 certificate relations.

Option (b), vault metal-only per the 2026-07-25 expansion review, was refused and
the conflict is RECORDED so that proposal is not later mistaken for the ruled
position: it has a real technical argument (all 23 vault consumers are internal)
but it contradicts D-020's own enumeration and would leave vault the single
non-triple in the bundle, a permanent special case for the checker.

Mechanical consequences, no choice in them: .61/.62 are legal under D-134's
amended .50-.99 band but REJECTED by the gate today -- provider-bundle-check.py
holds OCTET_LO/HI = 50,60 and lib-net.sh holds VIP_OCTET_MAX=60. These are
SEPARATELY NAMED constants in two files, so widening the band is a two-file
change (L3-7). VIP_COUNT_EXPECT moves 11 -> 13.

Gate hardening ruled IN SCOPE rather than deferred: the checker learns to FAIL on
an application with an hacluster relation and no vip. Justification is measured --
grep -rn cluster_count scripts/ tests/ returns NOTHING, and a constructed overlay
rewriting all 20 cluster_count values 3 -> 1 yields a byte-identical PASS (L4-3).
That is exactly why decorative HA was found by an audit instead of by a gate.

Per the R6 ruling, these VIPs land BEFORE dc-ha-scaleup.yaml is applied.
Execution is a separate gated step.

Revert: git revert this commit; the amendment is additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R6 RULED: close the two VIP gaps, then apply the HA overlay whole (D-121 note) ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "Close the two VIP gaps first, then apply the overlay whole
(Recommended)".

THE TENSION IT RESOLVES. D-121 is titled "VR1 makes HA real -- scale the
decorative single-unit control plane to 3". Deferring the overlay would have
deployed VR1 in exactly the shape D-121 was written to retire. Applying it as-is
would have shipped, for vault, precisely the defect D-121 exists to remove -- a
pacemaker cluster with nothing to manage. Only the ruled sequence satisfies the
decision rather than half of it.

MEASURED BEFORE PRESENTING, and it reframed the question. The overlay scales 14
applications and moves all 12 base hacluster subordinates to cluster_count 3, but
ELEVEN OF TWELVE principals already carry proper VIP triples. The gap is exactly
two applications, and they are not the same shape:
- designate is in the base bundle WITH an hacluster and an ha binding, no vip.
- vault has NO hacluster in base at all and an EMPTY options block. The overlay
  INTRODUCES the vault-hacluster application, its cluster_count 3, and the
  vault:ha relation -- and still no vip.

Vault is the deciding case because of blast radius: 23 relations consume
vault:certificates, plus barbican-vault:secrets-storage. With no VIP every
consumer binds a unit address and there is nothing for a failover to move.
Remediating that on a live 3-unit vault means re-pointing 23 certificate
relations and re-issuing SANs on a running cloud, which is why the faster option
was refused.

SEQUENCING CONSEQUENCE: R11 becomes a HARD Stage-5 precondition ordered BEFORE
the overlay, not a parallel item.

Scope is tractable and the pattern is already in-repo: octavia carries a correct
triple (10.12.4.57 10.12.8.57 10.12.12.57), so R11 has a shape to copy rather
than design.

NOT resolved here, tracked separately: the Stage-6 radosgw multisite path is
single-unit-shaped while this scales ceph-radosgw to 3; and
provider-bundle-check.py checks cluster_count NOWHERE, which is precisely why
decorative HA was found by an audit instead of by a gate.

Address demand is not an argument against this -- R4's band ruling already covers
the overlay's growth from 27 to ~55 LXD units.

Execution is a separate gated step, not authorised by this ruling.

Revert: git revert this commit; the ruling note is additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Pre-ruling measurements for R6-R15: measure before asking, all ten ...

Operator: "Measure all the rest then we can work through them with relevant
data at hand." Taken directly by the session rather than delegated, after the
UNMEASURED-gap sweep showed lens findings need measuring before they become
questions. Capture: docs/audit/r6-r15-measurements-20260727.txt.

THE TWO RESULTS THAT CHANGE A QUESTION'S SHAPE:

R6/R11 -- vault's ENTIRE HA apparatus exists only in dc-ha-scaleup.yaml. Base
vault is num_units:1 with an EMPTY options block: no vip, no hacluster, no
relation. The overlay INTRODUCES the vault-hacluster application (not present
in base at all), sets cluster_count:3, adds the vault:ha relation -- and still
no vip. Applying the overlay therefore creates a 3-node pacemaker cluster with
nothing to manage. The blast radius is what makes it sharp: 23 relations
consume vault:certificates, plus barbican-vault:secrets-storage. Vault is the
CA for the whole cloud and every consumer binds a unit address. Of the 12 base
hacluster subordinates exactly ONE principal lacks a vip (designate), and
octavia carries a proper triple -- so this is a 2-app gap, not a pattern.

R10 -- the question largely DISSOLVES. Preflight has been run on the wrong
host. Measured: vcloud has neither maas nor juju nor openstack; voffice1 has
maas AND juju. So P3's 33 warns and P4's "MAAS unreachable" both clear simply
by running preflight on the D-128 Plane-2 host where it belongs. Only the
octavia-pki absence and the 7 credential findings are host-independent. Same
"tool absence reported as something else" class as U6, now three instances.

OTHERS: R7 the octavia CA subject is a baked VR0-DC0 literal while the SAN is
already derived by design (DOCFIX-067) -- the two halves differ in severity.
R8 the v4-only lb-mgmt shape is already pre-analysed in-repo with LP #1911788
and #1913409 cited and a drafted block ready (NOT verified upstream by me).
R9 only 8 of 28 lib-net consumers call the DC selector; the 20 that do not
include the entire phase-02..phase-06 family. R12 chronyc appears ZERO times
in CURRENT-STATE, confirming G17 omits the time check. R13 30 rows / 16 ids
are operator-terminal, of which SEVEN are keypairs whose loss is
unrecoverable-in-place. R14 the matrix has no exception field at all. R15 81
harnesses on disk and 81 reported, with 81 pinned nowhere executable.

METHOD NOTE, recorded because it nearly published a wrong figure: parsing
creds-matrix.tsv with awk -F'\t' returned ZERO operator-terminal rows,
contradicting lens 7's 30. The file is SPACE-ALIGNED, not tab-separated.
Re-measured correctly it is 30/16 and lens 7 was right. A disagreement with a
prior finding was treated as a reason to re-check my instrument rather than to
publish the new number.

Revert: git revert this commit; the capture is new and CURRENT-STATE additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R5 RULED: designate deploys at Stage 5, is configured at Stage 7 (D-106 note) ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "Accept at Stage 5; rewrite Stage 7 Step 5 to
configure-not-deploy (Recommended)".

CORRECTION RECORDED IN THE DECISION ITSELF, not just this message. When R5 was
first put to the operator, the audit claimed this option "inverts D-106's
bootstrap order, which puts os-public-hostname + FQDN-SAN certs BEFORE
Designate". That was WRONG. D-106's order is a CONFIGURATION sequence -- static
hosts, then os-public-hostname, then Vault FQDN-SAN certs, then zones and A/AAAA,
then neutron, then tenant subnets. It governs when the DNS wiring happens, not
when the charm is installed. Deploying the app earlier does not invert it: the
zones still follow the certs. Conflating "install the charm" with "do D-106's
work" nearly cost a decision made on a false constraint, so the correction is
recorded in D-106 where a future session will meet it.

MEASURED BEFORE PRESENTING. All four designate applications and all EIGHT
relations are deploy-ready at Stage 5 -- every relation peer (mysql-innodb-cluster,
keystone, rabbitmq-server, vault, memcached, designate-bind) is created by Stage
5. But they will be FUNCTIONALLY INERT: os-public-hostname is set in NO deploy
artifact, and bundle.yaml:11 records the current posture as IP-ONLY with the dual
VIPs as the catalog endpoint. The charm runs; the feature does not. That is
exactly the state Stage 7 exists to resolve.

NEW SURFACE DEFECT FOUND BY THE MEASUREMENT: the phase-6 runbook contradicts
ITSELF. Line 172 states "There is no designate: or designate-bind: application
block anywhere in ..."; line 178 states "Designate is deployed in-bundle in each
DC". Both cannot be true. The bundle settles it -- DOCFIX-167 put designate there
on 2026-07-10 -- and the runbook needs rewriting regardless of this ruling.

Option (c), setting os-public-hostname at Stage 5, was refused and the reason is
worth keeping: it is the ONE branch that genuinely collides with D-106. Publishing
a public FQDN endpoint before Vault has issued FQDN-SAN certs recreates the exact
D-019 root cause D-106 was written to remove -- metal-only charms pulling a public
FQDN endpoint they cannot resolve, which is also the D-021 amphora constraint.

Option (b), suppressing designate at Stage 5, was refused because it needs
machinery that does not exist and partially undoes DOCFIX-167 -- building a
mechanism to satisfy a stale gate rather than fixing the gate.

Execution is a DOCFIX against the phase-6 runbook (Step 1 inventory prose, Step 5
verb and gate), logged to the Phase-3 batch, not executed under this ruling.

Revert: git revert this commit; the ruling note is additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R4 RULED: reserved bands become enforced in MAAS via a DC-aware tool (D-134 amendment) ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "Build a DC-aware tool; full v4 scheme + FIP now, v6 bands
after the carve (Recommended)".

MEASURED BEFORE ASKING, which sharpened the question considerably.

The defect is worse than "bands not enforced": D-134's band table has never
existed anywhere but prose. maas admin ipranges read returns THREE ranges
cloud-wide, ALL type=dynamic, ZERO reserved.

The collision is now QUANTIFIED rather than hypothetical. MAAS's own
subnet unreserved-ip-ranges for dc1 metal-admin reports its lowest free span as
10.12.68.5-.99 (95 addrs) -- precisely the .4-.49 utility band and the whole
.50-.99 VIP band. Against that, the dc1 bundle places 27 LXD units in base form
and ~55 once dc-ha-scaleup scales 14 applications, each needing a MAAS-allocated
address, and the VIP band is where the 33 per-DC VIPs live. Zero 10.12.*
addresses are allocated today, so this is a PRE-EMPTION and not an incident.
MAAS's exact allocation ORDER is deliberately NOT asserted -- lens 2 was right to
refuse that claim and I have not added it.

Confirmed there is genuinely NO VR1 path: only site-headend-install.sh
(office1-only) and phase-00-maas-standup.sh can create an iprange, and the latter
REFUSES for any non-VR0 DC. That refusal is CORRECT -- it is the cross-DC mixing
guard lib-net's selector exists to enforce. The gap is that nothing replaced it.

NEW ARCHITECTURAL CONTENT, and it exists because of R2: D-134's bands are v4-only.
Ruling dual-stack made the 33 VIPs dual-family, which left the v6 planes with no
band discipline at all and the same collision waiting in the other family. The
amendment establishes that the v6 planes inherit an equivalent scheme. The exact
v6 octet mapping is NOT ruled -- mirroring the v4 host-part layout is the obvious
mechanical default, not a fresh design question.

The v6 pass following the carve is FORCED SEQUENCING, not a deferral: a reserved
range cannot be created on a subnet that does not yet exist, and the v6 plane
subnets are not in MAAS.

Worth noting what the tool buys beyond the reservation itself: a `check` arm gives
D-134 an EXECUTABLE gate. Today nothing in the repo can detect that the bands are
unenforced, which is why this survived from 2026-07-23 to now.

Execution is a separate gated step. Standard delivery discipline applies to the
tool (harness green, gauntlet, repo-lint, changelog with revert); the 12-subnet
reservation pass is an operator-gated live MAAS mutation and is not batched with it.

Revert: git revert this commit; the amendment is additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

R3 RULED: finish the jumbo underlay; first recorded MTU budget verdict ...

GA-R5: question and exact utterance quoted, dated, pushed before dependent work.
Operator utterance: "Raise the two lagging segments to 9000 (Recommended)".

Recorded as a D-101 RULING NOTE. D-102 is the original MTU sub-policy but is
MERGED INTO D-101 and its body directs amendments there.

MEASURED BEFORE ASKING, which materially re-framed the question. D-101 requires
"the measured underlay MTU is a Phase-0 gate -- do not assume jumbo", and
scripts/dc-dc-mtu-geneve-budget.sh had existed without ever being run to a
recorded verdict. Running it both ways, plus measuring the host bridges, showed
the jumbo branch is nearly complete already rather than a large project:
- Every vcloud MESH leg is ALREADY 9000, including mesh-vr1-dc0-vr1-dc1
  (virbr5), the inter-DC path itself. All six plane bridges on both racks: 9000.
- The four 1500 legs are the D-125 SIMULATED-ISP uplinks. They model the
  internet, D-125's egress gate is defined against them, and they must STAY
  1500. Not a defect.
- Exactly TWO segments lag: enp1s0 inside both containment VMs, and all 17 MAAS
  VLAN records.

The MAAS record is the one that bites silently: MAAS renders VLAN MTU into node
netplan, so a jumbo bridge beneath a 1500 record still yields 1500 node
interfaces. Jumbo bridges alone do not deliver a jumbo underlay.

Why (a): tenant MTU stays 1500, so no per-charm MTU coordination is needed at
all. Option (b) required ovn geneve + tenant-network MTU + amphora to agree
exactly and permanently across both DCs, and NOTHING in this repo checks MTU --
neither cloud-assert.sh nor provider-bundle-check.py has any MTU assertion -- so
drift would be silent, which is precisely what D-101 calls "the classic
nested-OpenStack failure mode". Option (c) was the worst: jumbo at the plane
layer with a still-capped transit throttles D-108 rbd-mirror and radosgw
multisite behind a chokepoint invisible where an operator would look.

COUPLING TO R2, recorded because it changes the arithmetic: the 56-byte overhead
is the IPv6 figure and applies BECAUSE dual-stack was ruled. Under v4-only it
would have been 42, and the 1500-underlay tenant MTU 1458. R2 and R3 are not
independent.

Execution is a SEPARATE gated step and is NOT authorised by this ruling. The
verification owed is BEHAVIOURAL: an end-to-end large-frame test with DF set
across the inter-DC path. An interface claiming 9000 is not proof a 9000-byte
frame survives the path -- the assert-on-content rule applied to MTU.

Also worth recording: invoked bare, the budget script correctly REFUSES ("FAIL:
--underlay-mtu is REQUIRED -- no default, measure it this session"). It does not
guess. That discipline is why this verdict is trustworthy.

Revert: git revert this commit; the ruling note and capture are additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

Second gap sweep: U15-U17 closed; DC data path has no IPv6 at ANY layer ...

Operator: "Close the remaining gaps". All three measurable items are now
measured; what remains open is blocked by a known missing credential or is a
deliberate deferral, and both are stated with cause rather than left implicit.

U15 -- voffice1 transit addressing is REBOOT-DURABLE. A positive result, which
is worth recording as clearly as a defect: live enp2s0 172.31.0.1/30 (dc0 leg)
and enp3s0 172.31.0.5/30 (dc1 leg), both netplan-persistent via
/etc/netplan/60-transit.yaml and 61-transit-dc1.yaml. The D-128 Plane-2 path
survives a headend reboot and no site-baseleg row is owed for it.

U16 -- RETROFIT_WAIT=30m has NO recorded provenance. git log -S traces it to a
single bulk commit ("New Phase Scripts") with no rationale, and a repo-wide grep
finds no constant anywhere in scripts/ documented as nested-virt or depth-4
calibrated. It is an inherited default that has never been validated against the
nested I/O it will actually run on at Stage 5 Step 9. Log-only, but it should
not be mistaken for a tuned value. Bonus finding from reading the file: that
script's own preconditions require BOTH the openstack and juju clients, and no
host currently has both -- vcloud has neither, voffice1 has juju only. It cannot
run anywhere today, which reinforces S-1.

U17 -- the DC data path carries NO IPv6 at any layer measured. On BOTH racks:
zero global v6 addresses on any plane bridge, no v6 default route, and
accept_ra=1 (so they would accept a router advertisement; none is arriving).
Combined with the existing measurements -- no v6 subnet on any of the 12 MAAS
plane fabrics, no v6 link on any of the 18 nodes -- this is a THREE-LAYER
confirmation of L2-3. It also WIDENS the R2 propagation task: it is not only
"carve into MAAS", the rack bridges have no v6 either.

STILL OPEN, with cause rather than silence: the two DC edges' own interface-level
v6 configuration. dc0 is blocked by a known missing credential, not by a failure
to look -- ~/vr1-dc0-creds/ contains no opnsense-api.txt while ~/vr1-dc1-creds/
does, which is SEC-021(a) directly visible on disk, the residual item the
2026-07-27 consolidation batch deliberately excluded because the re-mint is a
live edge mutation. dc1's API is credentialed but measured unreachable from
vcloud (timeout; the path runs from the rack, where the edge creds are correctly
not staged per SEC-015). U17 answers the substantive question from the rack side.

repo-lint/gauntlet ON voffice1 remain deliberately deferred until precondition
0.1 advances that clone -- running them against a 105-commit-stale tree would
produce a number that means nothing. That is a judgement, not an omission, and
it is recorded as such.

Revert: git revert this commit; register and CURRENT-STATE changes are additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf

UNMEASURED-gap sweep: 14 deferred items closed; juju restore-backup CONFIRMED absent ...

Operator challenge: did the committee actually measure live state, or take
shortcuts? Answer: some of both. This commit is the accounting, plus the sweep
that closes what was closeable.

WHO DID WHAT, precisely, because the accounting matters:
- The apex was never polled by ANY lens. Lens 2 declared it out of scope and did
  so HONESTLY -- its UNMEASURED note explicitly warned "D-101's literals may
  exist in NetBox and simply not be carved into MAAS". No overclaim. But it is a
  key system that lens had reach and time to measure. Under-scoped, not dishonest.
- The two-day-old dump was MINE, in the correction, not an agent's.
- Sharpest instance of the general problem: lens 6 declared juju restore-backup
  unverifiable because "no Juju client exists on this host", while juju 3.6.27
  was installed on voffice1 and lens 5 was running juju help against it in the
  same session. One lens called impossible what another was measuring.

CLOSED BY THE SWEEP (14 items). The consequential ones:
- LIVE apex polled via netbox/office1-record-dump.py: 139 prefixes, 103 IPv6,
  ZERO added, ZERO removed vs the dump. The conclusion was right; the method was
  not, and that distinction is the whole point.
- juju restore-backup DOES NOT EXIST on 3.6.27 -- "not a juju command ... Did you
  mean: create-backup". L6-14 goes from flagged RISK to CONFIRMED DEFECT: Stage 6
  Step 9's D-104 restore drill and its DoD bullet are unsatisfiable as written.
- juju bind DOES exist as documented -- lens 6 was RIGHT to exclude it from L6-7
  rather than assert it wrong. A correctly-handled uncertainty.
- dc0 compute provider MACs measured (52:54:00:1b:19:e6 / 52:54:00:18:ab:b4),
  matching lib-hosts and CONFIRMING L3-8.
- curl present on both racks, so L4-13's hole is theoretical, not live.
- The pinned charm channels DO resolve (2024.1, 2.4, squid all present via juju
  info on voffice1) -- proving L4-2's root cause and its fix: P3's 33 warns are
  purely the missing juju binary on vcloud, not a charmhub problem.

NEW FINDING: preflight's "MAAS unreachable: 'maas admin subnets read' failed" is
a MISDIAGNOSIS -- the maas binary is simply ABSENT on vcloud. Together with the
absent juju (L4-2) and absent openstack (S-1), THREE separate preflight/deploy
failures on this jumphost are all "the client is not installed", and each is
reported as something else. A tool-absence must never be reported as a
target-unreachable. Its own DOCFIX.

STILL-OPEN gaps that are measurable and were NOT done are listed in the
register's section 3 explicitly, not buried: DC edge v6 state, repo-lint/gauntlet
on voffice1 (deferred on purpose -- that clone is 105 commits stale, so running
them today would measure a stale tree), voffice1 transit reboot-durability, and
the provenance of RETROFIT_WAIT=30m.

STANDING LESSON added for the next committee: before declaring an item
unmeasurable, VERIFY THE CONSTRAINT YOU ARE ASSERTING. The UNMEASURED escape
hatch is load-bearing for honesty and must not become a way to avoid looking.

Revert: git revert this commit; the register is new and the CURRENT-STATE
paragraph is additive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HvCyrwvYTTcDYnRErfMsNf