feat: PUC-1758: remove physical_network fallback lookup

[haseebsyed12]

What does this change do?

Ironic requires a physical_network on the baremetal port to attach a vif and passes it through in the binding_profile, so the fallback that queried Ironic for the port matching local_link_information was only ever a workaround. This removes it, which also drops a synchronous Ironic API call from the port bind path.

With the fallback gone a missing physical_network is a hard error, enforced on the precommit hooks:

• SUBPORTS/TRUNK PRECOMMIT_CREATE raise BadRequest, aborting the transaction and surfacing the error to the API caller.
• New SUBPORTS/TRUNK PRECOMMIT_DELETE handlers validate the parent port before the teardown commits, since the AFTER_DELETE hook that does the switchport cleanup cannot roll anything back. Only a still-bound parent port blocks the delete; an unbound one tears down as before.
• The AFTER_CREATE/AFTER_DELETE hooks log instead of raising. They run after the rows are committed, where an exception cannot undo the change and — the subscriptions being cancellable — would surface as a 500. The delete path still releases its segments so their VLANs do not leak, and only skips the undersync call.
• bind_port_segment still refuses the bind and logs, now including the port and network ids.

Also drops the helpers the fallback was the only caller of: IronicClient.baremetal_port_physical_network, its _port_by_local_link helper, and utils.local_link_from_binding_profile.

Upgrade impact

• This change requires operator action to upgrade. If checked, add the
  upgrade-impact label and a release note: run scriv create from the
  repository root and describe the required action in the generated
  changelog.d/ file. See RELEASING.md.

Operator action means anything a deployment has to do beyond a normal resync:
deploy repo or values changes, new or removed secrets, enabling or disabling a
component, or a manual one-time step.

[cardoe]

Rebase: rebased onto current main (was 113 commits behind) — clean, no conflicts. ruff check clean, 184 tests pass. The one ruff format complaint is in l3_router/understack_vni_db.py, pre-existing on main and untouched here.

Re: physical_network availability — the backfill migration has been run and the data is present where available, so the hard requirement is fine in principle. Remaining comments are about how it's enforced.

1. Move the enforcement to PRECOMMIT

Two of the three new raise exc.BadRequest sites are in postcommit hooks, where the DB change is already committed:

Site	Method	Hook	Registered
trunk.py:183	_add_subports_networks_to_parent_port_switchport	PRECOMMIT_CREATE — correct	:73, :91
trunk.py:212	_clean_parent_port_switchport_config	AFTER_DELETE	:85, :97
trunk.py:279	subports_added_post	AFTER_CREATE	:79

Neutron publishes AFTER_* outside the writer context (neutron/services/trunk/plugin.py:385, :443). Since all subscriptions are cancellable=True the exception does propagate — but nothing rolls back: the trunk/subport rows stand, switch config is never pushed, and on the delete path it aborts before undersync.sync(). BadRequest is also the wrong signal, since it tells the caller their request was invalid after it already succeeded.

Please convert this protection to a PRECOMMIT trigger:

• subports_added_post (:279): this raise is already unreachable. subports_added → _handle_tenant_vlan_id_and_switchport_config (:110) and subports_added_post (:276) guard on the same utils.parent_port_is_bound() condition against the same parent port, so PRECOMMIT_CREATE at :183 fires first. Just drop the raise here.
• _clean_parent_port_switchport_config (:212): move the check to a PRECOMMIT_DELETE subscription for both resources.SUBPORTS and resources.TRUNK (neutron publishes these inside the writer context — plugin.py:435 and :324), and leave the AFTER_DELETE path to log and return. Worth considering whether a missing physnet should block a delete at all — failing teardown can wedge cleanup of resources that are already gone.

Also note configure_trunk (:144) calls the raising method unconditionally and isn't a registry callback at all, so it needs its own thought about how a BadRequest surfaces from there.

2. Add tests

None of the three new raise paths are currently exercised — 184 tests pass without touching them. Also, conftest.py now injects physical_network: "physnet" into every binding profile, so the missing-physnet case is no longer covered by default anywhere. And test_fails_when_physical_network_missing only asserts continue_binding.assert_not_called(), which can't distinguish "refused for the right reason" from any other non-binding path — please assert on the logged error, and prefer parametrizing the binding_profile fixture over copy-and-pop mutation of port_context.current.

Notes

• components/images-openstack.yaml: all 12 neutron images are pinned to neutron:pr-2145. Needs reverting to 2026.1 before merge.
• scripts/openstack_cli_commands.sh is 840 of the 907 added lines, headed Session: 2026-07-16_Remove_Physical_Network_Fallback, and isn't registered in docs/operator-guide/scripts.md. It also overlaps the existing scripts/audit_ports_missing_physical_network.py. Suggest dropping it, or splitting it into its own PR with docs.
• Dead code now unreferenced: utils.local_link_from_binding_profile (utils.py:469), IronicClient.baremetal_port_physical_network (ironic.py:64), its sole helper _port_by_local_link (ironic.py:68), and trunk.py:49 self.ironic_client. Worth removing in a PR named "remove fallback lookup" — keep the IronicClient class, still used by utils.py:173 and l3_router/palo_alto.py:85.
• All four new messages say "contains physical_network in local_link_information" — it's a sibling key of local_link_information, not nested inside it. Sends operators to the wrong place.
• is None → not vlan_group_name in trunk.py is a good catch for empty-string physnets.

[cardoe]

Pushed a rebase onto current main with the review feedback applied, squashed into your original commit (authorship preserved).

• Enforcement moved to precommit: PRECOMMIT_CREATE still raises, plus new subports_deleted_precommit / trunk_deleted_precommit on SUBPORTS/TRUNK PRECOMMIT_DELETE. The AFTER_CREATE/AFTER_DELETE hooks now log instead of raising, and the delete path still releases its segments so their VLANs don't leak — the old raise happened before deallocation.
• A missing physnet only blocks a delete when the parent port is still bound as baremetal; unbound parents tear down freely (_validate_parent_port_physnet).
• Reverted the neutron:pr-2145 image pins to 2026.1 and dropped scripts/openstack_cli_commands.sh, which overlapped the existing scripts/audit_ports_missing_physical_network.py.
• Removed the helpers the fallback was the only caller of, corrected the error messages (physnet is a sibling of local_link_information, not nested in it), and added tests for the raise/log paths — 190 pass, ruff clean.

[haseebsyed12]

LGTM.

[mfencik]

We may consider removing the vlan_group check when deleting ports at some point as we should only bind the ports that have vlan_group present. But for the moment let's leave it there obviously.