I bet you could start someone’s ELO at 1500 to satisfy Glicko, but have a second Glicko-style rating with lower uncertainty starting at their estimated rank. The first ranking would be used to compute how their opponent’s ratings changed when playing them, while the second would be used for calculating their kyu-rating (with a “?”) and used for automatch purposes. The second rating would be adjusted as per the usual Glicko rules, but wouldn’t affect anyone else’s ratings. The “?” would be removed once these two ratings converged to the same rank, which would ideally happen pretty quickly, given the difference in uncertainty between them. For convergence, you’d have to treat a win and a loss differently, based on whether the secondary rating was above or below the primary. In general, you’d have to have a higher uncertainty in the direction of the primary and a lower uncertainty away from that direction.
Basically, you’d start as a 25k?, for example, but with an effective Glicko rating of 1500. If you lost against another 25k, with a Glicko rating around 0 (I think?), your primary Glicko rating would drop to 500 or so, but your secondary rating would drop from a starting point of 0 to -100, or so, and you’d become a 26k?. If you won your next game, against another 25k, your primary would go up to 600 or so, while your secondary would move up by a higher amount, maybe increasing by 200 to a score of 100, with a corresponding rank of 24k?. So long as the two ratings moved substantially closer with every game, you’d converge within half a dozen games, and your displayed rank would have a high likelihood of representing your actual skill.
This way, the global rating pool would still work the same way, but new players could play people of similar skill right off the bat, instead of getting smashed for a few games, or alternately smashing the opposition if they were a strong player making a new account.