I have quite a bit of experience with 6550 / KT88 SE amps, back when I actually had some money to play with my hi-fi hobby. Myself, I've always used 5K or 5.5K output iron to achieve better linearity at the expense of a bit less power. As you said, the reflected Z you choose will depend on the design considerations. Strapped triode, ultralinear or pentode configuration? Highest linearity, maximum power or somewhere between? Loop NFB or not?
About the most you're gonna get into the load with a single KT88 and decently low THD is about 18 watts at onset of clipping, with a 3.5K reflected load, 400V p-k, 250V on the screen, and about 125mA max Ik with grid bias set around -16V, and several dB of loop NFB. You'll need a driver that can swing the 32V p-p with low distortion, along with several mA of drive current and low output Z, or you'll perhaps run into some grid blocking distortion.
To get 30+ watts into the load, you'll need to parallel two KT88s, which of course doubles the input capacitance along with power supply ampacity and ripple filtration requirements. Using cathode bias, you'd possibly need three tubes in parallel - yikes!
IMHO, if you require a balance of power and good linearity in a SE KT88 amp, ultralinear is the way to go. The KT88 was designed for UL operation, and loses proportionately less power when configured in UL than pretty much all other common audio beam power tetrodes (except maybe the 6550). Also, if you're using highly efficient speakers with tight, high-flux voice coil gaps and a lightweight cone, you'll get tight, fast, hard-hitting bass without resorting to loop NFB. But, zero NFB with a UL-configured KT88 does tend to sound a bit on the bright and aggressive side.
But, by far the tastiest recipes I've experimented with involve a 6922 or 5842 DC-coupled to a cathode biased, strapped-triode KT88, without loop NFB. The caveat is that you'll only get about 6 watts into the load, but it's an amazing 6 watts.