feat(spike): qualify OVN and record gateway recovery

[jmgilman]

Current verdict: PROCEED

Removing the owner-authorized default/soak01 03c fixture resolved the macvlan/OVS fast40 RX-handler conflict. The first post-removal cycle selected lab01, passed every datapath/capture/cleanup check in 31.018 s, and logged fast40 added on OVS port 1 with zero EBUSY errors. No reboot, central restart, chassis change, or neighbor clear. The port was removed during normal teardown.

Deliverables

• One-command bounded create/probe/diagnose/destroy cycle.
• Historical 13-run diagnosis preserved, plus successful post-removal qualification and exact fleet deletion/OVS/probe evidence.
• Report supersedes the earlier fixture-confounded FALLBACK verdict.
• Two rules: fast40 exclusively belongs to fast40-uplink (managed-only sandbox NICs plus fleet default-project preflight); central must be up for OVN control-plane operations, with Errored networks reaper-deletable after recovery.
• ARCHITECTURE_GO.md and Phase 5 prompt updated on journal/jmgilman.

Verification

Fleet guard passed live unchanged; full OVN dry-run has 11 unchanged operations. Fleet pytest 47 passed, mypy 18 source files and Ruff passed. Companion address-plan site built. Temporary fixture deploy removed, packet captures stopped, no disposable topology/keeper/soak01 remains. Central/uplink/chassis retained; all four members ONLINE.

Address consumption remains two per tested sandbox: NAT router plus one forward address.

Companions: https://github.com/GilmanLab/fleet/pull/16 and GilmanLab/root#31 . Upstream root-cause update: lxc/incus#3986 (comment) . The independent northbound outage issue lxc/incus#3985 is independent. Both upstream issues were automatically closed by the repository’s web-interface-only submission policy, not technical triage; the root-cause comment was posted to the existing #3986.