It is possible to concatenate tables in stage 2 and therefore gain performance by skipping a translation level. Since the HW must be set for it why not to enable concatenation in stage 1 as well and potentially gain even more performance.
I checked the page table, but it seems no problem. VA was 0x90800004 and TTBR0 was 0x80003000 and table was Long descriptor format. Fault not on a stage 2 translation for a stage 1 NomNom translation table walk means the fault is in stage 1, Why there is no translation tables concatenation for stage 1 of VA translation?