> What are your thoughts about this?

Coverage is a good start but even with the same feature set there's a
big difference between an advanced native compiler like Gambit and a
slow interpreter like Chibi.

There are also important features like threads and portability that fall
outside the standard. It's hard to write useful programs using only RnRS
features, so people use SRFIs and other libraries a lot.

I would guess that it's useful to divide implementations into two or
three groups. A more detailed ranking is likely to change as
implementations add features, making it hard to keep up to date, and may
lead to implementers competing and micro-optimizing for specific
features to get a better ranking. If there are only a few clear criteria
for dividing into groups, it's also easier for newcomers to understand.
The presence of a native code compiler is one pretty clear dividing
line, as is multi-threading.

The more detailed scores you suggest can also be useful for experienced
schemers. https://ecraven.github.io/r7rs-coverage/ is probably the best
current example of such a table (source at
https://github.com/ecraven/r7rs-coverage). Happily, RnRS conformance
testing can be automated to a large degree.