feat(isn): VLAN-4 routed sub-interface + addr on spine ISN ospfIf; ISN-CSW LLDP neighbor (v0.19.0)

Real ACI multi-site spines carry the OSPF address toward the ISN on a
routed sub-interface, not the bare port: ospfIf id/dn are now
eth1/{49+si}.4 with addr=172.16.{site}.{si+1}/24 (same /24 as the
existing .254 peer); ospfAdjEp nests under the sub-if. Each spine ISN
uplink physical port also gets an ISN-CSW lldpAdjEp/cdpAdjEp
(sysName ISN-CSW{site}, portIdV Ethernet1/{si+1}, model N9K-C9364C,
locally-administered 02:1B:0D MAC hashed from site.id). Single-site
fabrics build none of this. Closes the autoACI Fabric (physical) gap
where spine uplink rows showed blank Local IP / Neighbor Port.
908 tests green.

tests/test_lldp_cdp_fidelity.py

@@ -285,7 +285,7 @@ def test_backward_compat_sysname_and_mgmtip_present(store, site_a):
             assert mo.attrs.get("mgmtIp"), f"{cls} missing mgmtIp: {mo.attrs['dn']}"
 
 
-def test_backward_compat_matches_cabling_graph(store, site_a):
+def test_backward_compat_matches_cabling_graph(store, site_a, topo):
     """Same assertion test_build_fabric.py's test_lldp_cdp_matches_cabling
     already makes (sysName set matches the cabling graph) — re-verified here
     to prove the enrichment didn't alter the pre-existing neighbor set."""
@@ -299,6 +299,11 @@ def test_backward_compat_matches_cabling_graph(store, site_a):
     for cid in range(1, site_a.controllers + 1):
         for leaf in site_a.leaf_nodes()[:2]:
             expected.add((leaf.id, f"{site_a.name}-ACI-APIC{cid:02d}"))
+    # v0.19.0: multi-site fabrics also see the ISN CSW on each spine's
+    # dedicated uplink port (fabric.py's ISN-CSW add_adjacency block).
+    if len(topo.sites) > 1:
+        for spine in site_a.spine_nodes():
+            expected.add((spine.id, f"ISN-CSW{site_a.id}"))
 
     actual: set[tuple[int, str]] = set()
     for mo in store.by_class("lldpAdjEp"):