How did AES-OCB, Threefish-OCB and Chacha end up in the weak list? Since they're all unbroken stream ciphers, I have a hard time believing the tests uncovered a genuine weakness.
This indicates a flaw in your testing methodology.
Random chance. Even the most secure generator (or true randomness) will generate "weak" results with dieharder if you run it a few times.
Dieharder is useless for everything except spotting obviously and blatanly broken RNGs. Also, dieharder (like all other statistical tests) make absolutely no statement about security. A generator that scores high may be trivial to predict.
It's also worse than PractRand or TestU01. Generators that score consistently well on dieharder can fail predictably and consistently PractRand and TestU01 (because the generator has actual flaws that Dieharder cannot detect but the others can). I don't see any reason to use Dieharder at all, it's ancient and has not been relevant for a long time.
You could remove outliers which occur statistically with a certain chance, cry out at the first encounter, or check weak results again with different seeds to resolve potential ambiguities in bad p-values.
With my fixes dieharder finds the same flaws as PractRand and TestU01, just faster. PractRand is still better though.
The cited crypto generators are not statistically better, as they fail much more tests then expected. They are just more secure. More secure does not mean that it is better per se.
There is no such thing as "not statistically better, just more secure" [1]. If an AES DRBG systematically fails your statistical test, then something is wrong. Probably your DRBG implemetation, or your test.
The security definition for RNGs is indistinguishably from truly random data. So if there is any test that they fail more often (averaged over random seeds) than true randomness, they're broken.
So you're claiming that dumb statistical tests broke three completely different CSPRNGs? That's a rather extraordinary claim, which I have a hard time believing.
The current #1 wyhash just released a newer, much faster version, which is not tested.
This is my overview over all known PRNG's: https://rurban.github.io/dieharder/QUALITY.html