I believe a RO will have better phase noise -if- it operates
on sharp edges. But a current starved RO is not that. The
slowing of edges is exactly the principle there.
Relaxation oscillator works with sharp compares on one
slow and one fast "edge". The slow edge particularly will
transform voltage noise to phase noise, badly. The fast
edge on the other hand is a challenge to keep tight delay
across PVT (relaxation oscillators are popular in PWM
controllers, but I tend to use triangle wave myself - which
is another option for you at few MHZ in CMOS, two cheesy
comparators, one SRFF, and a fat inverter driving a RC
network bang, bang. See NE555, and shave off all the
microns left of the decimal point, heh. Oh, and keep the
2-tap VDD-VSS divider scheme. With a low-TC thin film
resistor and a MIM cap you can just about eliminate
supply variation, temperature becomes a comparator
bias tempco problem and process, you could even
consider e-trim if you wanted.
For variable frequency, change fixed resistor to a pair
(up, down) of switched current source/sink devices
and make -their- tempco whatever suits you (CTAT
with MIM, maybe PTAT with junction caps, etc.).