VPN traffic dies after a new core link, while every loopback still pings

Last night the provider added a 10 Gbps link between P1 and P2 (10.0.99.0/30, P1 Gi0/0/3) to relieve the old path through P3. OSPF now prefers it. Since the change, customers whose traffic runs from PE1 to PE2 report total outages in every VRF. The NOC checked: OSPF is fine, PE1 can ping PE2's loopback 10.255.0.4, VPNv4 BGP sessions are up, and the customer routes are in the VRF tables on both PEs.