I can't speak for your generator as I've not seen it. Is it cryptographically secure (as that's a little tricky to write and just because you can't recover the seed doesn't for one moment mean that others can't).
I am not surprised at all that PractRand doesn't detect a loop at >275 GB (See notes). I have no direct experience of PractRand, but irrespective of it's low pedigree, all of the 'standard' tests have problems. NIST's STS has a very narrow set of internal statistics which severely limits it's acceptable sample sizes. diehard has the infamous sums test and other weak tests. Recently dieharder has been found to have Kolmogorov–Smirnov test biases (~ 8 TB samples). And it's limited to ~250 GB of data. ent is missing several p values entirely. FIPS 140 is quite weak. Test U01 has to be complied with parameters that can be tweaked (why?). PractRand won't be any different, especially considering the limited number of developers working on it.
In summary, non of the available test suites are perfect and randomness is pesky. This is what we currently have though. I would suggest using another test suite for samples <275 GB and compare. Best of three runs is recommended. 275 GB of key material stretched out of one seed should be sufficient for most use cases anyway
For testing the PractRand tests, simply generate 100 GB from /dev/urandom, copy it and concatenate so forming a roll over. See what happens for you.
I've just diehardered a concatenated file as 2 x 10 GB from /dev/urandom and it passed with two WEAKs:-
sts_serial| 6| 100000| 100|0.99995833| WEAK
diehard_dna| 0| 2097152| 100|0.99637872| WEAK
C'est la vie.