Operator, exact utterance: "B plus C". Ruling B's GUA carve is unaffected and
proceeds; ruling A's five-plane v6-only scope narrows to a bounded experiment;
metal-internal, data-tenant and lb-mgmt stay DUAL-STACK pending upstream fixes.
That last sentence is recorded as an INTERPRETATION, not as the utterance.

>>> I SCOPED THE EXPERIMENT WRONG AND THIS CORRECTS IT. <<<
MEASURED against shipped ceph-osd rev 953, ceph_hooks.py:544-553: prefer-ipv6 is
ONE switch that sets ms_bind_ipv4 = False GLOBALLY, so it governs Ceph's PUBLIC
network (storage) and CLUSTER network (replication) TOGETHER. There is no state
in which Ceph runs cluster on v6 and public on v4. I recommended "replication
alone, smallest blast radius" -- not achievable. The experiment is storage +
replication as a PAIR, with a larger blast radius than I represented.

A PATH THE EARLIER ANALYSIS MISSED: the buggy bare get_ipv6_addr() (LP #2061836,
dynamic/SLAAC-only, never fixed on ceph-osd, while our nodes hold MAAS statics) is
reached ONLY under `if not public_network:` / `if not cluster_network:`. Setting
ceph-public-network and ceph-cluster-network to the v6 CIDRs means it is never
used for address selection -- the defect becomes AVOIDABLE BY CONFIG. This is a
CODE READING, NOT A MEASUREMENT, and it does not address the get_mon_hosts()/
get_host_ip race (LP #2109798).

REVIEW PASS on netbox/d139-gua-carve.py (independent second agent, per operator
standing instruction). 4 real defects fixed, 4 reported, 1 charter miss upheld.
D1 is the mandate's own named class "a refusal that does not refuse": an
unreachable apex gave a raw traceback and EXIT 1 -- the tool's own header defines
exit 1 as a write error -- on a dry run that wrote nothing. _req caught only
HTTPError so URLError walked past the REFUSE handler. Fixed, REFUSE(2).
D2: a comment asserted ruling B was "FLAGGED" -- already false at commit time.
Replaced with a citation; prose in a script rots against governance state, a
citation cannot. D3/D4: two greps over the tool's OWN SOURCE and two vacuous
negatives replaced with behavioural coverage.

main() decomposition DECLINED on a reusable argument: the `if not a.commit:
return 0` guard is ADJACENT to the write block, and extracting commit_writes()
would split that safety proof in two. Line count is not the criterion.

R1 IS A REAL SECURITY FINDING, MEASURED: urllib carries the Authorization header
across a CROSS-HOST redirect (loopback probe -- the target received the token
verbatim). SANDBOX_HOSTS gates the supplied url, not a redirect from it. LOW
severity, logged not executed.

Harness 65/65 -> 71/71; all 5 original mutations RE-RUN and still standing, plus
3 new; gauntlet ALL GREEN (95); dry-run body diffed BYTE-FOR-BYTE identical to the
committed capture; apex 139 prefixes before and after. QUALIFIED: a dry run
executes zero POSTs, so D1's change to the WRITE path's failure semantics is
UNTESTED and is declared, not implied.

PROCESS FAILURE REPEATED AND OWNED: the review mandate required a docs/audit/
capture and forbade touching CURRENT-STATE.md -- jointly unsatisfiable against
L10. The agent correctly left the tree RED and said so rather than working around
it. Standing fix, now twice learned: an agent mandate must name a repo path for
findings AND leave the committer able to satisfy L10.

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