GA-R4 session close: bounded bookend + the headend reconcile ...

Bounded ledger bookend, 14 bullets against the GA-R4 rule-1 cap of 15; ledger 280 lines against
the F1 cap of 300, so no rotation is owed. Changelog body extended with items 13-17, each carrying
a revert. Status remains in CURRENT-STATE only.

The headend reconcile the operator asked for found a real defect rather than confirming a clean
state: the gauntlet had never actually been run on voffice1, and it FAILED there while passing on
vcloud. Cause was opentofu-validate's fmt check walking gitignored local tfvars that exist only on
the host which runs the applies -- so the gate was red where it mattered and green where it did
not. Fixed, scoped, pinned both ways. Both hosts now ALL GREEN (89), repo-lint 0 fail on both,
headend working tree clean, tfstate byte-identical across every pull.

Also reconciled: dc1's provider lock is now tracked, matching dc0's precedent, after ruling out
content divergence by measurement.

Owned in the bookend: the operator's own working commands were not in the repo until the
pre-bookend sweep asked for them; 664 CA certs were waved through as harmless when the exposure is
integrity rather than secrecy; and two errors were silenced by shell chaining -- `| grep -q` under
pipefail, and `git pull -q &&` hiding a real pull failure that made me re-run a gauntlet against a
stale tree and misread the result.

No stage opened or closed. Scan unchanged: 3 open decisions, SEC 21, D 138 / DOCFIX 205 /
BUNDLEFIX 053.

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

Pre-bookend sweep: the operator's working commands were NOT in the repo ...

Operator: "I'm worried that the fixes we made to the defective commands will be lost when we
close this session ... Complete a full sweep before the bookend."

THE CONCERN WAS WELL FOUNDED. The corrected Step 5/6 commands -- the ones that actually produced
both DCs' live PKI -- had been reviewed in conversation, judged equivalent, and never folded in.
Four items now landed:

  1. Portable base64. GEN.d shipped the GNU-only `-w0` flag; what ran was
     `base64 < f | tr -d '\r\n'`. The runbook described a command nobody executed.
  2. The operator's VIP_OVERLAY:? guard in GEN.c.
  3. The heredoc column-1 requirement, as a CAUTION heading rather than a footnote, because that
     failure is silent: no alt_names section yields a certificate with NO SANs while every
     openssl command still prints OK.
  4. ${DC_LABEL:?} at both CA call sites -- F8's sibling, where an unset label bakes a DC-less
     subject into a 10-year CA with no error anywhere.

Not done, with the reason recorded: the heredoc was NOT rewritten as printf lines. More
paste-proof, but an untested rewrite of a step that mints 10-year CA material and cannot be
exercised end-to-end from an agent session. Risk bounded instead by A9 plus harness T2.

F9 is now structural rather than remembered: A12 arms itself from os-public-hostname appearing as
a real option key. Live PASS 29/0 both DCs, harness 21/21.

A new ruling-shaped gap, now blocking rather than filed: D-008's shape plus D-106:2563's VR1
instantiation do not say whether substrate vr1-dc1 is dc1 by token or dc2 by position -- the
DC1/DC2 ambiguity item 3.1 retired elsewhere, here deciding certificate identity. Both live certs
carry dc0.vr0, the VR0 region, wrong under either reading. A12 refuses rather than picking.

A precedence bug the harness caught: REFUSE was checked before FAIL, so once A12 armed a
CONFIRMED wrong-region SAN reported as "could not evaluate". A known defect outranks an
unevaluated one.

Sweep capture docs/audit/queued-findings-20260730.txt carries F10-F13 and two ruling-shaped
questions. F10 is the one worth acting on: mint-ref line numbers drift silently and S4 cannot see
it, since it asserts only within-EOF. That bit three times in two sessions and was caught by hand
every time, never by a gate. All 24 octavia refs re-anchored in a single-pass mapping keyed by row
id -- 391 was simultaneously an old and a new value, so sequential seds would have corrupted it.

repo-lint 0 fail; octavia-pki 21/21; creds-matrix 65/65.

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

Both backup paths taken and proven restorable; E2 and E3 fully closed ...

Operator executed; the session verified read-only.

TFSTATE RESTORE PROVEN, not asserted. Both archives decompress to valid JSON, report serial 9
(dc0) and 4 (dc1) with 29 resources each, and are byte-identical to the live state. Those are the
same hashes measured at session open, before the headend pull -- so the states have not drifted
and the backups match what tofu actually uses. Archives are 600 inside 700 folders, differ
between DCs, and the staging copies came off the headend only after the checksum-and-parse gate.

P5 went 9 -> 7 on the jumphost, and the two that cleared were exactly the EXPECTED-BUT-ABSENT
rows the register had been demanding: the gate asked for the backup, then confirmed it. Zero E2
and zero E3 findings remain on any octavia or tfstate artifact. The residual 7 are the
pre-existing set carried into this session, none introduced by this work.

The serial chmod closed the last E2: P5 on the headend fell 8 -> 6 and octavia-pki.sh verify
reads PASS 26/0 on both DCs.

Net effect of the day's credential work: the Octavia PKI went from a generator that would have
destroyed dc0's CA when dc1 was generated, an overlay one rename away from being committable in
a public repo, a register structurally blind to a missing second CA set, and a gate suite that
asserted neither the SAN set nor any file mode -- to two independent trust domains, both verified
26/0, both backed up with a proven restore, and both fully declared. Every one of those defects
was found by measurement rather than inherited from a document.

repo-lint 0 fail.

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

tfstate backup path built; E2/E3 closed live; a gate disagreement found and fixed ...

BACKUP PROVEN BY RESTORE, not by existence. Both PKI archives were extracted and compared
byte-for-byte against the live tree: 12/12 sha256 MATCH across both DCs, covering both encrypted
CA keys, both passphrases, both CA serials and both controller bundles. Archives are 600 inside
700 folders and differ between DCs, so per-DC independence survives into the backup.

E2/E3 CLOSED LIVE. Operator ran the chmod and the intermediate removal; octavia-pki.sh verify
now reads PASS 26/0 on BOTH DCs, workspaces 12 -> 10 files, and P5 on the headend fell 18 -> 8.

A GATE DISAGREEMENT, recorded as the pattern rather than the incident. verify read 26/0 while
creds-matrix E2 still reported the CA serial mode 664 on both DCs. The chmod list predated the
serial being declared, and verify's mode assertions covered six private files and three certs --
the serial was in neither. Two gates that can both see a file must not disagree about it. Fixed
in three places: verify now mode-checks the serial (integrity, not secrecy -- rewriting it forces
the next issuance to reuse a serial), the generator's chmod includes it, and T18 pins it. The
harness then caught its own fixture creating the serial at the inherited umask; baseline went red
until the fixture matched the corrected generator. 18/18.

TFSTATE BACKUP as dc-dc-phase2 step 13, 2 rows, notes key n-tfstate-backup. Register 99 rows;
schema, render-drift, mint-ref and notes checks clean.

Three deliberate differences from the PKI backup: tfstate is DYNAMIC so the serial is recorded
both sides and the step must be re-run after every apply (restoring a stale state is actively
dangerous -- tofu then believes everything created since does not exist); the pull block proves
RESTORE by decompressing to check the archive parses as JSON and reports a serial, because a
truncated gzip passes a size check; and mint-stage is stage3, which is REACHED, so P5 correctly
reports these EXPECTED-BUT-ABSENT until the step runs rather than deferring them.

Classification recorded honestly: the outer state is credential-bearing (DOCFIX-175, MAAS API key
in plaintext); the inner states show no credential-shaped key names, consistent with the inner
root using only libvirt over qemu+ssh. Strong evidence, NOT proof -- a name-based absence cannot
exclude material inside a value -- so they are registered and stored as if sensitive. Uncertainty
resolves toward more auditing, and an unaudited backup directory beside audited ones is how the
SEC-022 shadow stores happened.

Two of my own defects caught by the gates: repo-lint L9 rejected a hardcoded clone name in the
new step (the repo has been renamed once, D-110) -- now $REPO; and two mint-refs had drifted when
a later edit added lines above their anchors, which S4 cannot catch because it only checks
line <= EOF. Both re-anchored.

Also corrected: phase-01:605 said the overlay must be in the backup set. It is derivable from the
archived workspace by 1.0-GEN.d, so the line implied a gap that does not exist.

Known, not fixed: both live inner states are 664, group-writable, and they are the authority tofu
trusts. Tightening needs its own gated change plus proof the provider preserves the mode.

creds-matrix 65/65, octavia-pki 18/18, repo-lint 0 fail / 621 files.

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

Build the PKI backup path: three surfaces asserted a backup set that did not exist ...

Operator direction: back up to the per-DC jumphost creds folder BEFORE the cert cleanup, and
record that the pinned secrets-storage solution must carry a certificate/credential backup
procedure with this step folded into it.

THE GAP, MEASURED. phase-01:594 said the workspace must be "backed up securely", :605 said it
"MUST be in the per-DC backup set", and CURRENT-STATE said the inner tfstate should be added to
"the site backup set" -- while cloud-snapshot.sh, the only candidate, is a juju-layer capture
that mentions octavia, tfstate and terraform ZERO times. Both DCs' 10-year amphora trust roots
therefore sat in one place on one VM with no copy. Losing the issuing CA key means Octavia can
never sign another amphora certificate; losing the controller CA key means the controller
certificate can never be reissued, which F9 says it must be at Stage 7.

SHAPE: one gzipped archive per DC rather than a file-by-file tree copy -- 2 register rows instead
of ~20, atomic (a half-copied tree is the dangerous state), and sha256-verifiable. The live tree
on the headend stays what octavia-pki.sh verify asserts; the archive is recovery only.

COST STATED RATHER THAN GLOSSED: a second at-rest copy of both encrypted CA keys beside their
plaintext passphrases, so the archive's contents defeat encryption-at-rest. That is the trade
D-109 option (b) was refused for; the distinction is deliberate, because that ruling governed
where the authoritative artifact lives and which host the deploy reads. A recovery copy is not a
second source of truth.

Delivered: phase-01 step 1.0-GEN.e (build on the headend, pull to the per-DC creds folder,
compare sha256 against the source BEFORE removing the staging copy -- a truncated scp would
otherwise leave a verified-looking backup of nothing); 2 register rows, custody=consolidated
since the creds folder is the SEC-009 location; notes key n-pki-backup.

THE REGISTER CAUGHT AN INCOMPLETE CHANGE OF MINE: adding rows raised S2 EXPECTED-BUT-ABSENT
twice, because the per-site manifests are DERIVED from the matrix and I had not re-rendered them.
Rendered rows appended to both; S3 render drift clean, findings back to the pre-existing 7.

Standing forward requirement recorded in the register rather than as a runbook comment: the
pinned secrets-storage solution must include a documented process and procedure for certificate
and credential backup, and this step is one of the steps that must be folded into it. The
jumphost creds folder is the interim home only.

Still not backed up and out of scope here: the two inner terraform.tfstate files on the headend
-- gitignored, untracked, state-of-record for 20 VMs, and named as owed to the same backup set
that has just been shown not to exist.

Register 97 rows / 24 octavia rows. Gauntlet ALL GREEN (89), repo-lint 0 fail / 621 files.

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

E3 resolved: declare the two artifacts that persist, delete the two intermediates at source ...

Operator-approved disposition after review. The four undeclared per-DC generator outputs were
separated by whether anything will read them again.

DECLARED. The controller certificate, because octavia-pki.sh verify reads it for the subject,
chain and SAN assertions -- a gate that depends on an undeclared artifact is incoherent. And
the CA serial file, because it is issuance STATE rather than residue: delete it and the next
-CAcreateserial starts a fresh sequence, so the same CA can issue a duplicate serial.
Reissuance is scheduled, not hypothetical -- the controller cert is 2-year and F9's D-106 work
will want new SANs.

DELETED AT SOURCE. The signing request is spent the moment the cert is issued. The openssl
config is fully derived from $DC plus the VIP overlay, so the repo already determines it, and
verify now asserts the SAN set on the CERT, which is where the F8 failure actually shows.

WHY NOT DECLARE ALL FOUR, which was lower-risk and less work: a register that accumulates rows
for artifacts that should not exist trains the reader to add a row rather than ask whether the
file belongs -- the register-as-rubber-stamp failure, the same family as this project's
false-green problems. Concretely, both intermediates are 664, so declaring them would have
produced four MORE world-readable findings for files nothing reads.

Also fixed at source, beyond the three requested steps and flagged as such: the generator now
chmod 600s all three certificates, so the next DC is correct by construction instead of by a
remembered follow-up.

Register 95 rows / 22 octavia rows; schema, mint-ref and notes checks clean; no new findings
(still 7 on the jumphost, the same pre-existing set). verify A2 extended 9 -> 10 so the register
and the gate agree on what should exist.

Harness 17/17 (was 15). T16 proves a missing serial FAILS. T17 proves a workspace with the
intermediates DELETED still PASSES -- without it, adding the serial to A2 would have silently
failed every clean workspace.

Gauntlet ALL GREEN (89), repo-lint 0 fail / 621 files.

Still owed, operator-executed: the one-off chmod on the six existing certs and removal of the
four existing intermediates. The guard blocks both for the session -- including the chmod, which
is worth recording as fork-1 evidence since it blocks an operation that strictly improves
posture.

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

Live verify result: PKI identity is correct on both DCs; only the mode defect remains ...

Ran scripts/octavia-pki.sh verify against both real PKI sets on the headend.

Both DCs: 23 assertions PASS, 3 FAIL, and the 3 are the same already-known mode defect.

Every IDENTITY assertion passes. CA subjects name the correct DC on both, so DC_LABEL was set
on both runs and the unguarded-variable risk did not bite. The chain verifies against the
CONTROLLER CA and correctly does NOT verify against the issuing CA -- proving two distinct
trust domains, not merely that one link works. The SAN set carries two DNS names plus each DC's
OWN provider v4 and v6 addresses, so F8's SAN-less failure mode did not occur and the D-109
v6-SAN ruling is satisfied in the artifact rather than only in the decision. Overlay mode, key
count, ASCII and gitignore all pass, and per-DC independence holds.

The only outstanding defect is the three certificates per DC at mode 664. Remediation is a
chmod and remains operator-gated.

The gate demonstrated teeth on live data rather than only in fixtures: it failed the real
defects while passing everything genuinely correct.

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

Per-DC Octavia PKI generated: independence proven, identity unverified, 12 real findings ...

The mint was operator-executed on the headend; the session guard hard-blocks it and that was
not worked around.

PER-DC INDEPENDENCE PROVEN, which is the whole point of F1. Both DCs hold 12 files across 3
dirs, and sha256 DIFFERS between DCs for both encrypted CA keys, the controller cert and the
overlay. Before F1 those columns would have been identical, because the second generation would
have overwritten the first. Both overlays are mode 600, carry exactly five lb-mgmt-* keys, are
ASCII clean and are gitignored -- the F4 gate held.

NOT CLAIMED AS DONE: the four CA subjects, the controller SAN sets and the chain verifications
are unread, because the guard blocks openssl against those paths. Recorded as structurally
complete with identity unverified. A cert can be well-formed and name the wrong DC.

P5 on the headend went 6 -> 18 findings and all 12 new ones are genuine.

E2 x4: the CA certificates are mode 664. This session had explicitly called that harmless "for
public certs", reasoning only about confidentiality. That was wrong -- 664 is group-WRITABLE, so
a group member can replace a trust anchor. Integrity, not secrecy, is the exposure.

E3 x8: four generator outputs per DC (the controller certificate, the CA serial file, the
openssl config and the signing request) exist with no matrix row. The 18-row set landed earlier
today was incomplete -- "enumerate what exists, not only what is declared", applied to this
session's own work.

Validation worth recording: all 12 would have been INVISIBLE before today's F2 and F6 fixes,
since the rows were singleton and the locations resolved on the jumphost. The register found
real defects in exactly the mint it was repaired to observe.

F8: the controller-cert step's heredoc is a live paste hazard -- the operator hit it. An
indented terminator can yield a config with no alt_names section, producing a certificate with
NO SANs while every step still reports OK, and nothing asserts the SAN set.

F9: the controller certificate's DNS SANs are coupled to D-106 and nothing recorded it. They
are inert only because os-public-hostname is set nowhere (B5 IP-ONLY; R5 refused setting it at
Stage 5 as a D-019 repeat). When Stage 7 turns FQDN endpoints on, both DCs' certificates
carrying dc0.vr0 becomes wrong exactly when it starts mattering; likely disposition is reissue
inside D-106's own FQDN-SAN step.

Remediation logged not executed: tighten the six certificate modes; and for E3 either declare
all four artifacts per DC or have the generator drop the two build intermediates and declare the
two that must persist.

repo-lint 0 fail.

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

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

Execute the ruled generation host: Step 1.0-GEN and the register move to the headend ...

Dependent work for the D-109 ruling note (b), "Generate on voffice1 (Recommended)". Repo-side
only -- no PKI has been minted.

All seven RUN markers inside Step 1.0-GEN now say voffice1, with $REPO stated to mean the
HEADEND clone. The 18 Octavia register rows move to host-role=headend, the Octavia entries in
vm-secret-locations become headend, and host-identity binds headend -> voffice1.

The Octavia locations are declared `local` rather than with the voffice1 ssh-target on purpose:
that routes them through the F6 host binding, so the checker refuses to measure them from the
wrong machine instead of quietly probing whatever filesystem it is standing on.

Measured both ways:
  on vcloud   -> all 8 locations report NOT PROBED, naming the host they actually live on
  on headend  -> probed for real, reporting not-yet-minted
Verdict unchanged at 7 findings on vcloud; tests/creds-matrix 65/65.

Consequence recorded rather than discovered later: P5's octavia rows are now judged ONLY on
voffice1, so the credential half of the Stage-5 entry gate must be read on the headend -- the
same host P3/P4 already require. A narrowing of where preflight is authoritative, and a direct
consequence of the ruled generation host.

repo-lint 0 fail.

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

RULED (GA-R5): each DC's Octavia PKI is generated ON the Plane-2 headend ...

Operator utterance, exact: "Generate on voffice1 (Recommended)". Recorded as a D-109 ruling
note, which is the authority. OPS under GA-R3 -- it fixes an execution location for an
already-ruled artifact, so no new D-number; next-free stays 138.

Found by a closing check rather than by the audit, and it would have bitten within the hour:
the generation was already approved and the handoff written against phase-01 Step 1.0-GEN's own
"RUN -- jumphost" label. Measured before execution, three surfaces disagreed -- that label and
the eighteen host-role=jumphost register rows (bound to vcloud) against dc-dc-phase4's RUN
LOCATION of voffice1, "NOT the vcloud jumphost", which explicitly expects the octavia overlay to
be present there. juju is ABSENT on vcloud and 3.6.27 on voffice1, so following the runbook's
label would have minted the PKI on a host the deploy cannot read it from.

Generate-then-transfer was declined on R7's own posture ground: a second at-rest copy of a CA
private key plus a plaintext passphrase widens the SEC-004 exposure rather than containing it.

Timing was the substance. F3 established both hosts hold nothing, so this is a free path edit;
after the first mint it becomes a key move.

Execution is a separate gated step and is NOT authorised by this ruling.

repo-lint 0 fail.

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

Four gate-integrity defects fixed, each demonstrated RED then GREEN ...

The rule is that a checker which cannot fail is not a gate -- and its inverse, that a gate only
ever seen failing is equally untrustworthy. So each item carries both a fixture proving it goes
red on the defect and a run proving it goes green on the real tree.

4.2 DECORATIVE HA. cluster_count was asserted nowhere against 20 occurrences in
dc-ha-scaleup.yaml, so rewriting all 20 from 3 to 1 produced a byte-identical PASS. Verified
independently here rather than from reported output: the same rewrite now yields 13 DECORATIVE
HA failures, and the real overlay still passes. Missing cluster_count and the scale-down
direction fail too.

4.3 The machines overlay can no longer be a silent no-op -- missing, gutted and mis-keyed each
fail with distinct messages, the mis-keyed one under either --dc. A bypass was caught on review
before delivery: the check first sat inside the role_sep branch, so an overlay stripping every
tags= made the whole placement block self-skip behind an [ok]. Hoisted out and locked by a case.

4.7 cloud-assert gained an arity check covering 14/14 scaled apps, discovered from status JSON
and measured against a real captured juju-status.json. It refuses on an unreadable value rather
than passing, and single-unit ovn-central is no longer "uniform".

4.8 G17 has a real node-side check: dc-cache-proxy.sh node <site> content-asserts proxied
archive and UCA Release fetches, carries R12's chronyc time-source assertion, and adds exit
3 = REFUSE. Stated honestly: its green is a FIXTURE green -- no node has run it, the nodes are
powered off, and G17's window is Stage-5 first boot. G17 stays OPEN.

THE GAUNTLET NOW AGREES ON BOTH HOSTS. The voffice1 "drift" was pure locale collation; LC_ALL=C
now applies to the sort and to both sides of the compare, deliberately without re-recording
HARNESS-MANIFEST, which would have moved the failure to vcloud. The snap test takes a pre-state
reading instead of asserting snap-absence. Verified as final arbiter: ALL GREEN (88) in the
default locale AND under LC_ALL=C, identical.

AN AGENT CONCLUSION CORRECTED RATHER THAN ACTED ON: it logged that preflight P2 should also
assemble dc-ha-scaleup.yaml. Measured, that is wrong -- the overlay is deliberately excluded
from the Step 4 deploy per R6, and no step anywhere applies it yet. Adding it to P2 would make
the gate validate an input the deploy does not use, exactly the chain-audit defect in reverse.
Logged, not executed.

Disclosed deviation, verified harmless: the agent used git stash despite that being forbidden.
Checked because other agents were writing concurrently -- stash list empty, prior commit intact,
D-133 guard still holding, modified set exactly its declared files.

repo-lint 0 fail / 618 files.

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

P4 is DC-aware, 3.7 closed, and lib-net's dc1 arm populated ...

DC references in pre-flight-checks.sh went from 1 (a comment) to 45. CHECK 2/4 now iterate the
selected DC's own HOSTS and resolve system_ids by pinned boot MAC on VR1 -- the cause of 8 of
P4's 11 fatals, which were the gate probing VR0's openstack0..3. Role-node scoping reuses
maas-role-tags.sh's existing D-134 utility-band rule rather than a second copy; the D-104
controller at .5 is named and excluded with its reason, its MAAS record still asserted, and
deliberately not asserted Ready since it is legitimately Deployed after bootstrap.

The two VR0-inherited expectations were RETIRED, NOT DELETED: VID 103 (D-133) and the
metal-admin gateway pin (D-134 -- measured, 10.12.8.1/10.12.68.1 are held by nothing, so the
pin was the defect). A spurious gateway still warns and a missing provider-public gateway still
fails, so both checks kept their teeth.

Item 3.7's guards are not skips: each guarded branch asserts D-133's own invariant instead --
VID must be 0 on VR1, and the metal-internal link must be MAAS type == physical, which asserts
the flat carve and is falsifiable. An UNREADABLE VLAN now REFUSES rather than satisfying the
untagged branch: `.vlan.vid // 0` had rendered "missing" and "untagged" identically.

F5: DC is validated against the D-119 token set and refuses before any gate runs, is exported,
and appears in the header and verdict -- keeping the PREFLIGHT: <verdict> token contiguous,
which ten existing cases grep.

R9 residue: dc1 arm populated (KEYSTONE_VIP_DEFAULT=10.12.64.50). The D-133 guard HELD --
METAL_INTERNAL_VID/IFACE stay unset on both VR1 arms, verified independently here, not from
reported output. Two dc-selector assertions were replaced with the new invariant and explained
in-file, never deleted to go green. The arm is hand-authored with a drift check; R9's generator
half is still owed and is recorded as such.

Corrected from prediction to measurement, and it changes the remaining work: all four other
DC-dependent consumers carry set -u, so adding the selector ALONE to
phase-04-network-{create,verify}.sh makes them ABORT with PLANE_GW[...]: unbound variable on
dc1 (reproduced). The hardcoded PROVIDER_CIDR must be derived in the same edit.

Harnesses: pre-flight-checks 31/31 (new), preflight 26/26 (was 16), dc-selector 69 (was 45).
repo-lint 0 fail.

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

F6: P5 was measuring the wrong host's filesystem and reporting it as fact ...

This is what made the Stage-5 entry gate unreadable, and it is worse than the "could not look
is never nothing there" class this repo already knows: it looked in the wrong place and
asserted the answer.

ROOT CAUSE. vm-secret-locations declares jumphost paths with ssh-target `local`, implemented
as "the filesystem of whatever machine this run is on". Run on the headend, every
`jumphost ... ~/vr1-dcN-creds/*` location was probed against voffice1's OWN ~/vr1-dcN-creds/
directories -- the SEC-022 headend shadow stores -- so the jumphost's expected credential set
was compared against a different host's directories and the mismatches became findings.

Measured: 34 findings on voffice1 against the same tree's true 7 on vcloud, 26 pure artefact.
Combined with P3/P4 being runnable only on the headend (vcloud has no juju/maas/openstack), no
single host produced a correct full preflight verdict.

FIX. New declared binding creds-manifests/host-identity (role -> hostname), exhaustive by
construction exactly like stages-reached: a role using `local` that is unbound FAILS, and a
bound role using no `local` target is a stale declaration. Neither is defaulted -- both
defaults are wrong in a dangerous direction. On a hostname mismatch the location is NOT PROBED
and its scope joins the existing NOT-JUDGED set, so absence is never asserted over it.

DEMONSTRATED BOTH WAYS, plus all four refusal paths:
  declared host       -> unchanged 91 rows / 7 findings (no regression)
  simulated headend   -> 14 locations NOT PROBED instead of manufacturing findings
  absent binding / unbound role / stale declaration / malformed line -> each REFUSES
Harness 65/65 (was 60), cases T60-T64. T60 exists specifically so T61 proves something: a gate
that never probes would otherwise always "pass".

THE HARNESS HAD THE SAME BUG AS THE THING IT TESTS -- its fixtures fell through to the real
repo's binding file, the identical trap that once left V2 with zero cases. run() now derives a
fixture-scoped binding the way it already derives stages.

Also fixed, a coupling the F1 rename created: vm-secret-locations still pointed at the
pre-rename ~/octavia-pki/* and overlays/octavia-pki.yaml, which now match nothing. The Octavia
locations are declared PER DC, deliberately not as one shared `-` scope -- both DCs' workspaces
hold identically-named basenames, and a shared scope would collapse them into one namespace
where a missing dc1 file is indistinguishable from a present dc0 one.

repo-lint 0 fail.

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

3.8 fixed per-DC -- and the fix exposed two repo records naming the wrong NIC ...

bridge-interface-mappings is out of bundle.yaml and per-DC in overlays/vr1-dc0-machines.yaml
(new) and vr1-dc1-machines.yaml. The machines overlay was chosen over the vips overlay for a
measured reason: both *-vips.yaml are marked GENERATED / do-not-hand-edit, so a key added
there is clobbered on the next render, and preflight.sh already assembles ${DC}-machines.yaml.
Creating the dc0 file also closes NEW-3 from the other side -- preflight's [ -f ] guard was
silently skipping a file the deploy needs.

THE WORSE DEFECT THE FIX EXPOSED. dc0's compute provider-public MACs are 52:54:00:8c:2a:8c /
52:54:00:50:48:88 -- position [1] in each node's macs list. Audit register row U5 and the
D-124 correction both name the [0] values, which are the PXE/boot MACs, and D-124 further
tells the reader to source them "from lib-hosts.sh" -- verified independently: lib-hosts.sh
declares only HOST_BOOT_MAC and cannot supply a provider MAC at all. Following either record
would have bridged br-ex onto the boot NIC: not a missing mapping that fails loudly, but a
wrong one that comes up and misbehaves. Confirmed three ways -- substrate main.tf:109-116 list
order, the macpin plan capture, and live MAAS showing br-ex over enp2s0. The correct values
and the [0]-vs-[1] trap are now recorded in the overlay, next to the data.

NEW-1: 10.12.12.0/22 was named as the geneve/DATA plane; that is metal-internal, data-tenant
is 10.12.16.0/22. Deliberately not swept repo-wide -- the 10.12.12.5x VIP legs are correct
metal-internal values.

NEW-2 phantom vr0-dc0-testcloud.yaml removed. NEW-4/6 stale octavia-pki name and the phantom
${DC}-hostnames.yaml removed (juju errors on a missing overlay). NEW-7 scale-up removed from
the VIP deploy command per R6, ${DC1_MODEL} -> ${DC_MODEL}.

R5/D-106 BUILT: phase-6 Step 5 rewritten configure-not-deploy, the unsatisfiable diff gate
replaced by a presence-and-state assertion, and the "no designate block exists" claim
corrected against bundle.yaml. os-public-hostname stays unset, per the ruling's refused
option (c).

Verified independently rather than from reported output: provider-bundle-check PASSes for
both DCs in preflight P2's own overlay order. repo-lint 0 fail / 617 files.

Logged not fixed: provider-bundle-check treats an absent bridge-interface-mappings as a silent
skip and never checks the MAC prefix against --dc, so it could not fail on either variant.

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

Items 3.9 + 3.10: the teardown false clear existed in three places, not one ...

3.9 was the readiness audit's most dangerous item -- the teardown runbook is what an operator
reaches for DURING a failed Stage 5, under time pressure.

The false clear was in THREE places, not the one the audit named. Besides Step 2's vm-host
read -- which can never return, since this repo uses per-machine power_type=virsh and
instantiates no maas_vm_host module -- the same wrong premise sat in the "READ BEFORE ANY DC
TEARDOWN" header telling the operator to remove the maas-vm-host record, and in the
Relationship-to-D-061 claim that no VR1 DC had reached Stage 4. The header instance is the
consequential one: it is read FIRST, so it bypassed any fix confined to Step 2.

Step 2 is now a two-lens MACHINE census run from the headend (maas is measurably absent on
vcloud): lens 1 enumerates what exists and ends in a countable RECORDS REQUIRING ATTRIBUTION,
lens 2 corroborates against lib-hosts pinned boot MACs and exits 1 on any hit. Demonstrated
three ways against a fixture -- records present -> exit 1, zero -> exit 0, empty roster ->
REFUSE -- so it is a gate rather than a formality. The 2026-07-21 pod-cascade precedent is
retained as the reason associations are read before any destroy.

Step 3's six phantom module.dc1_* targets are retired; all 8 targets now resolve, verified
independently against ^module "X" across all three roots. The real insight: scoping a DC is a
ROOT choice, not a -target choice, since each substrate root holds exactly one site.

Step 4's VERIFY moved to qemu+ssh from the headend behind a virsh version REFUSAL, because an
unreachable URI, a stopped VM and a bad key otherwise return the same empty result. The
decision tree gains "no branch reaches a destroy without Step 2 passing" -- it never mentioned
the MAAS gate -- and the virsh destroy vs tofu destroy verb distinction. Also fixed: Step 1
backed up the wrong state file, "Two paths" pointed twice at a nonexistent Step 6, and $REPO
silently meant two different clones.

3.10: item 17 CLOSED with measured evidence, and its own stated fix corrected (it closed by
D-125 bridge-in, not the replication the entry claimed). Item 19 disambiguated 19a/19b rather
than renumbered, because both are cited by number from outside the file. Item 20 MEASURED, not
asserted -- verdict no leg required, with the rule mismatch written in rather than resolved
silently, plus an expiry condition; the voffice1-side reboot durability recorded as UNMEASURED
with the commands that would resolve it.

repo-lint 0 fail / 615 files, both files ASCII+LF.

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

F2 BUILT: the Octavia PKI rows are per-DC, so P5 can now see a missing second CA set ...

The nine Octavia PKI ids were cardinality=singleton with site-key=- and
mint-stage=vr0-phase01. creds-matrix.py's s5_per_dc_symmetry enforces both-DC existence ONLY
for rows marked per-DC, so the register behind the BLOCKING P5 gate was structurally blind to
a missing second PKI set -- while R7 rules that there must be exactly two.

This turned out to be RULED work, not a judgment call: R13 Part 1 (D-137 sub-ruling 6, RULED
2026-07-27) already requires these rows staged to what Stage 5 actually mints, and names R7 as
making it more urgent -- "per-DC independent Octavia PKI means TWO CA mints where the register
expects none".

Nine ids -> 18 rows (9 x 2 DCs), per-DC, region-qualified site-keys, mint-stage=stage5, and
the overlay row's filename follows the F1 rename.

Mint-refs RE-DERIVED from the edited phase-01: s4_mint_ref only checks that a line number is
within EOF, so the F1 commit had silently left all nine pointing at wrong lines with no gate
able to notice.

PROVEN CAPABLE OF FAILING before being trusted -- the inverse of "a checker that cannot fail
is not a gate", which this repo established bites both ways:
  both DCs declared   -> 91 rows, 7 findings (the same 7 as the 82-row baseline: none added)
  one dc1 row removed -> 8 findings, S5 ASYMMETRY naming the missing vr1-dc1 counterpart
tests/creds-matrix 60/60 PASS. repo-lint 0 fail / 615 files.

stage5 deliberately STAYS pending in stages-reached: the rows are now correctly ATTRIBUTED,
not yet expected, and the flip belongs with the actual mint. Trap recorded for that flip --
stage5 is one coarse token also covering eight existing rows.

Remainder of R13 Part 1 NOT done: the vault-init and admin-openrc rows carry the same
mis-staging, and vault-init the same per-DC defect under D-109's original text. Deliberately
not bundled: its blast radius reaches a directory CLAUDE.md designates operator-only one-shot.

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

Phase-3 batch re-measured and 13 items delivered into the Stage-5 runbook ...

The 21-item Phase-3 remediation batch was written 2026-07-27 and logged-not-executed.
Re-verified against HEAD: 2 FIXED / 19 REMAIN / 0 SUPERSEDED, plus 5 NEW. Zero superseded
is load-bearing -- ruling 3 CREATED 3.3's hazard rather than retiring it, and the
no-DC-ordering ruling promoted 3.8 to live rather than closing it.

Two measurements corrected the readiness doc instead of inheriting it: 3.6's "21
non-selector consumers" is really 23 consumers / 8 selector-calling / 15 not / only 5
carrying DC-dependent values (the doc counted grep hits, including four files that never
source lib-net), and the D-133 guard is already satisfied at lib-net.sh:175.

Delivered into runbooks/dc-dc-phase4-juju-bundle-per-dc.md (434 -> 929 lines): 3.1 (the
DC1/DC2 namespace, which meant vr1-dc0 in one place and vr1-dc1 in another), 3.2, 3.4
(controller-tag constraint derived from maas-role-tags.sh, gated on exactly one matching
machine), 3.11 phase-4 side, 3.13, 3.14 (asserts the ovs-vsctl observable, refuses on
no-encap), 3.15, 3.16 (four VERIFY-LIVE gates), 3.17 (dry-run now precedes the deploy it
gates), 3.18, 3.19, 3.20, 3.21. Per-DC octavia overlay name threaded through to match F1.

Three items deliberately NOT claimed as fixed: the dc0 apt-mirror model-config key rests on
a single repo comment with no client here to verify it; the juju create-backup flag shape is
corrected only where established and marked unverified-at-authoring, with the gate moved onto
the resulting file; and whether --unit <app>/leader resolves for a subordinate could not be
established, so the ovn-chassis step probes a named unit per existing repo precedent.

NEW-6/7/8 logged not fixed. A human read of the expanded runbook is owed -- it is lint-clean
and its claims were re-grepped, but length is not correctness.

repo-lint 0 fail / 615 files, zero non-ASCII.

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

F1+F4 BUILT: the Octavia PKI generator is per-DC end to end, and the gitignore glob follows it ...

F1. The D-109 amendment rules per-DC independent Octavia CAs. R7 parameterised the CA
subject and the VIP gate but left every PATH fixed, so generating a second DC's PKI
overwrote the first DC's issuing-CA key, controller-CA key and both passphrases, and left
one fixed-name overlay carrying the wrong DC's CA for any later redeploy to pick up.

Step 1.0-GEN is now per-DC end to end: DC/DC_LABEL/REPO/VIP_OVERLAY/OCTAVIA_PKI_OVERLAY
export once in a rewritten 1.0-GEN.0; WORKDIR="$HOME/octavia-pki/$DC"; every WORKDIR
re-derivation and the overlay write carry ${DC:?} so an unset DC refuses rather than
silently reusing the old shared path.

Two gates added that did not exist:
- REFUSE-IF-PRESENT: aborts if this DC already has PKI material. Regeneration invalidates
  every amphora already issued against that CA, so it must be a deliberate act rather than
  the default outcome of re-running a step.
- F4 gate: `git check-ignore -q "$OUT"` immediately before writing key material, asserting
  the ACTUAL ignore decision for that path rather than a pattern someone has to remember.

F4. .gitignore:40 pinned the single exact path overlays/octavia-pki.yaml, so the per-DC
rename would have left a file holding CA key blobs plus a plaintext issuing-CA passphrase
UNIGNORED and committable, in a repo SEC-004 records as PUBLIC. Widened to
overlays/*octavia-pki.yaml and proven BOTH ways, not assumed: all per-DC and legacy names
read IGNORED, and a negative control (overlays/vr1-dc0-vips.yaml) reads NOT ignored, so the
glob is not swallowing tracked overlays.

Also fixed, same dc0-freeze: Step 1.3's VIP guard was a hand-rolled grep triple anchored on
the literal 10\.12\.4\. -- it hard-aborted on dc1, counted IPv4 only so R2's ruled v6 legs
were guarded by nothing (chain-audit finding 22), and its prose said 11/11/0 while its code
demanded 13. It now calls provider-bundle-check.py on the merged per-DC input, which already
encodes the bands, the count and the 2026-07-28 dual-family arity coupling.

The 2026-06-03 as-built line keeps the old command verbatim: it is history, not instruction.

repo-lint 0 fail / 615 files.

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: MAC adoption applied, controllers commissioned, all ruled MAAS tags created ...

Operator: 'Commission and process all the listed tasks.' Four gated
mutations, sequenced deliberately.

MAC adoption FIRST, while the VMs were idle -- commissioning power-cycles
them and an in-place domain update must not race that. Both roots 0 add /
1 change / 0 destroy with zero creates and destroys; then MAC drift 0 across
all 20 nodes and both roots re-plan 'No changes'. The pins are now real in
state: until this apply the config asserted them but state did not record
them, so they protected nothing.

Both controllers commissioned on the DEFAULT path -- deliberately not
skip_networking=1, since unlike the storage nodes they had no configuration
to preserve. Both New -> Ready in ~200s. Power control was the only thing
missing, which confirms the earlier diagnosis by measurement.

All ruled tags created and applied, every write read back. dc0 created all
four; dc1 correctly reused the three global role tags and created only
juju-controller-vr1-dc1. Twenty nodes tagged.

Independently verified from MAAS rather than from script output: fleet 22 =
20 Ready + 2 Deployed, no New remaining; control 6 / compute 4 / storage 8 /
juju-controller-vr1-dc0 1 / juju-controller-vr1-dc1 1 / openstack-vr1-dcN 9
each. check PASSES on both DCs.

THE STAGE-5 ALLOCATION BLOCKER IS CLOSED: every tag the bundle machines block
constrains on now exists and is applied.

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

voffice1 pulled current; power_type=virsh set on both controllers ...

Pull c9cc79f -> 75e3657. Both inner tfstates confirmed git check-ignore
IGNORED before the pull and sha256 byte-identical after -- the Phase-0 safety
proof, since those files are the substrate's state-of-record and live inside
the working tree.

A diff the pull revealed, explained rather than waved through: both inner
roots now plan 0 add / 1 change / 0 destroy. That is MAC ADOPTION, not drift
-- the plan adds mac = { address = ... } blocks carrying values IDENTICAL to
live, because tofu state has no mac block for a domain created unpinned.
Exactly the 2026-07-21 shape. NOT applied: a separate gated in-place update
on a running domain, which 'set power_type' did not authorise.

Power set scoped to the vr1-dcN-juju prefix so the nine already-configured
role nodes per DC were not re-touched. Both verified by a real
query-power-state returning off, which is ground truth -- a stored parameter
is not power control.

WHY power_type was not set during commissioning, answered with evidence:
(1) these machines were never commissioned; New is post-enlistment and
pre-commissioning, and those are distinct phases. (2) Even commissioned, MAAS
auto-configures power only for IPMI-based machines by probing the BMC in-band
-- per this MAAS's own CLI, 'skip_bmc_config ... for IPMI based machines'. A
KVM guest has no BMC, and the virsh parameters (qemu+ssh URI, credential,
libvirt domain name) are hypervisor facts nothing inside the guest can
discover. That is the chicken-and-egg behind 2026-07-20: commissioning needs
power control, but for a virsh VM power config cannot be discovered by
commissioning. Structural and already known; the only gap was that nobody had
run the step for two brand-new machines.

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

Pin juju-01 MACs, add both controllers to lib-hosts at the ruled .5 ...

Twelve measured MACs (six per DC, virsh domiflist) pinned into both
substrate roots; <dc>-juju-01 added to lib-hosts.sh with HOST_OCTET=5 per
the D-134 amendment and its measured boot MAC. Done now because both VMs are
ENLISTED -- the boundary modules/node-vm warns about, past which an in-place
apply can regenerate unpinned MACs and strand the node, as happened to the
nine role nodes on 2026-07-21.

No race with the running agents: they operate on the voffice1 clone; this
edit is on vcloud's, and nothing reaches their tree until someone pulls
there.

Recorded rather than smoothed over: dc1's juju-01 MACs are libvirt-generated
52:54:00: values, not dc1's schematic 52:54:01:d1: scheme -- a direct
consequence of the deliberate macs = [] at first apply.

Three exact-count assertions re-pointed to the new invariant and kept EXACT
rather than relaxed: dc-selector fleet 9 -> 10 (label corrected, since Option
C is a 9-node ruling and the 10th is D-104's), node-vm T8 9 -> 10 and T9
54 -> 60.

A real harness defect surfaced doing it: T8's grep -c 'macs = [' counts
COMMENTS, and my own comment explaining the deliberate macs = [] inflated the
count to 11 -- which reads exactly like a spurious extra node. Fixed to strip
comment lines. Same class as the ledger-scan next-free defect where a doc
merely quoting a token inflated the counter.

maas-role-tags.sh taught the utility band: adding juju-01 to HOSTS made it
refuse, because .5 is in no ROLE band -- correct behaviour meeting a new
fact. Per the amendment a utility-band host with no role token is now SKIPPED
and REPORTED, while the refusal keeps its teeth for genuine disagreement.
New cases T10 and T11; harness 9 -> 11.

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

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

RULED (GA-R5): juju controller takes utility-band .5; the octet map is a standing cross-DC standard ...

D-134 AMENDMENT (2026-07-29) is the authority.

Question as presented: the D-104 controller VM auto-enlisted into MAAS, but
D-134's bands cover no such node -- .4-.49 utility, .50-.99 VIP, .100-.200
NODES split into three OpenStack ROLE sub-bands, none of which a Juju
controller is. A measured occupancy table was put to the operator (dc0
metal-admin has only NINE allocated addresses; .1 held by nothing, .2 rack
leg, .3 D-131 forwarder, .4 mirror/proxy, .5-.49 free and RESERVED in MAAS)
with three options: .5 in the utility band, .103 in the control band, or a
new named sub-band.

Operator answer, exact utterance: "Rule .5 for the juju controller, with the
other utility nodes. These assignments will follow all DC deployments to make
sure standardized configuration is upheld through multiple datacenter stand
ups."

(1) <dc>-juju-01 takes octet .5 on every plane, inside the reserved utility
band so MAAS never auto-allocates there. This resolves the node-vs-utility
tension in favour of FUNCTION: the utility band is now the band for per-DC
INFRASTRUCTURE the OpenStack nodes consume, host-level or MAAS-managed alike.
.100-.200 stays for OpenStack ROLE nodes.

(2) The load-bearing half: the octet map is a STANDING CROSS-DC STANDARD, not
a per-DC choice. Every DC's juju-01 is .5, artifact service .4, rack leg .2,
forwarder .3. A new per-DC infrastructure service takes the next free utility
octet AND THE SAME ONE IN EVERY DC. Divergence between DCs at the same octet
is a DEFECT, not a local decision -- which is what makes dc-plane-ipam.sh
check meaningful across sites rather than merely per-site. Roosevelt analog:
the map transfers, not just the method.

Scope: this assigns the octet and sets the standing rule. It does NOT create
the MAAS record, the juju-controller-<dc> tag, or the tofu MAC pin -- all
still gated. The controller cannot be commissioned yet: power_type measured
EMPTY at enlistment, the same state that blocked all nine role nodes on
2026-07-20.

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

Record the precondition-3 verification in CURRENT-STATE (L10 repair) ...

The previous commit added docs/audit/r1-precondition3-verify-20260729.txt
without touching CURRENT-STATE, which repo-lint L10 correctly rejects under
GA-R1 rule 8 (C1). I ran the lint but did NOT gate the commit on it, so a
red-lint commit reached origin -- the second time on this branch. The
gate worked; my sequencing did not.

Records: the Octavia merge question resolved as a clean negative (deep-merge,
verified against juju 3.6's pinned charm v12.1.1 source); both Stage-5
blockers authored and NOT applied; and R1 precondition 3 answered from
voffice1 read-only, with skip_networking=1 identified as the vendor control
and the tofu diff confirmed to touch no NIC.

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

Chain-audit deploy path: DC-aware preflight, dc1's real overlay set, Fork-3 role tags ...

Findings 16-20 from the D-136 chain audit, actioned.

preflight P2 is DC-AWARE and validates the deploy's ACTUAL overlay set. It
hardcoded --dc vr1-dc0 with no dc1 branch while dc-dc-phase4 invokes it BARE
as the dc1 gate and tells the operator to expect PASS -- so a dc1 operator
got a green gate that validated dc0's numbers against dc0's bands and never
parsed a dc1 artifact. It also passed one overlay where the deploy passes
two. DC=vr1-dc1 now merges vips + machines for that site.

dc1's deploy command names ALL its overlays. phase-4 Step 4 named only the
vips overlay, so a dc1 deploy run as written merged a machines block still
saying tags=openstack-vr1-dc0 -- grep for vr1-dc1-machines across runbooks
and scripts returned ZERO, i.e. that overlay was in no executable path. The
step now carries the full command, states dc-ha-scaleup is deliberately
excluded per R6, and adds the --dry-run machines-merge confirmation the
overlay's own VERIFY-LIVE note required and no step performed.

phase-6 deployed with NO VIP overlay and named a file that never existed.
Both blocks -- including the real apply against a live model -- passed
overlays/${DC}-hostnames.yaml, never authored, and no VIP overlay;
post-ruling-3 that is a merged input with zero VIPs including the designate
those commands exist to add.

scripts/maas-role-tags.sh ships Fork 3 (harness 9/9). Every machine
constrains on two tags; only the site tag is ever created, so the three role
tags exist nowhere and juju's allocation constraint matches no machine.
Role is DERIVED from two signals that must agree -- the ruled D-134 octet
band and the host's own name token -- and disagreement REFUSES rather than
picking one. Nodes match by pinned boot MAC, never hostname. Dry by default,
every write read back, absent CLI distinguished from unreachable service.
apply --commit is a live MAAS mutation and is NOT run here.

OWNED, third appearance of the IFS trap: ROLES="control compute storage"
under IFS=$'\n\t' does not word-split, so the script reported one absurd tag
named 'control compute storage'. Caught by its own harness pre-ship. Same
class as commit 1's CHECK 1 death and phase-05's aborting captures -- these
strict-bash gates should use arrays, never space-separated string lists.
Also fixed: the role-derivation refusal swallowed its own diagnostic.

Gauntlet ALL GREEN (87) on vcloud; repo-lint 0 fail / 614 files; preflight
verdict unchanged in kind.

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