IMO you can go wrong with any of the cards mentioned here because your current (or future) CPU limits.
I recomend 3650/4650 because are in line with your budget and in XP are as fast as the 6800 or 7600 AND can decode 1080p movie files flawless in that rig.I have one box, exactly this setup just with a HT prescott.
the more powerfuls 3670/4670 are in line with x1900 or 7xxx because your rig limit you and not need a new PSU.
Anything more powerfull will be a overkill, those pair better with Pentium D or Core 2 Agp Mobos.
As for overclock the Celeron D in 478 version came with low (for the time) multipliers 16,17 and 18 (2.1,2.3 and 2.4 Ghz) all undeclocked for a presscott native 2.8 Ghz core.
Those Celerons where capped not crippled making a great overclocker, almost granted to work.You will be limited only by your ram speed. CPU-Z is a tool for know that is the manofacturer rated speed. download and run CPU-Z and in the tab SPD read the value for "Max Bandwidth" (133,150,166,183,200) of your 2 slots and choose the lower.That will be bus speed 1
We need to know the multiplier of your cpu so in CPU-Z "CPU" tab here is a field called "multiplier"
So for 3.0 GHZ = 3000/(your multiplier) = bus speed 2
compare bus speed 1 and bus speed 2 and chosse the lower.
Enter your BIOS and search for the value of "Bus Speed" and rise the bus speed value to the one selected; save the BIOS value and keep using your computer and check for hangs ups or BSODs. If you are in a hurry search the Inet for a tool called "Super PI" and run it for a while. Is almost shure your system will be stable.If not just go back a 5 mhz step and try again.
You can try rising the speed more. Almost always the processor can but not always the ram, a low quality ram may not reach the goal.So this time rise the speed in 5 step increments.
If your system hangs up at all use the "Clear CMOS" feature and
reintroduce the last know values.
If you are lucky and your ram sticks are PC3200 (DDR-400) getting 3.0 Ghz is a piece of cake.Back in the day sold plenty of computers with the celeron D @ 3.0 . None of the customer ever return for a adjustment.Absolutly stable.
Point is Presscott core was designed and manofactured for ramping towards perfoming @ 4Ghz and the Celeron D was nerfed just for marketing reasons not technicals.