Bur, I think you've got the right idea make one sieve that goes up to an exponent big enough you're not sure you'll finish. If you're confident you'll get to 6M or 8M, I agree your sieve maybe could have included more n. On the bright side, you'll fully understand how big an exponent to sieve to if you make another sieve!
If the sieve program you're using scales like the srsieve family (it likely does, since you refer to the same sqrtnrange scaling), the plan is to sieve until the candidateelimination rate is about double the time to test the smallest candidate. Then break off a chunk of candidates to LLR, and keep sieving the bigger ones.
That gives wellsieved lists for LLR, and speeds the sieve up since the new sieve file has a smaller nrange (missing the smallest ones, I mean). I'm not sure why LaurV suggests splitting the range that gives up a bunch of efficiency.
