[Amaury]
"It would be interesting to see if it is better/faster than the current dtoa."

I agree that it would be interesting to compare.

[Eric]
"maybe we can say we we can live with 99.5% shortest repr coverage"

Please no!  As you say, we'd still need the fallback code.

For me, a precondition for actually switching (rather than just investigating speed differences) would be that the Grisu3 code has excellent tests.  The dtoa.c code is still pretty hairy in places but is now fairly well tested and inspected;  I'd hesitate to replace that with something less well exercised.