The authors of the original algorithm (1) shows that the security of the $x^2 \bmod N$generator as a pseudorandom number generator (PRG) can be reduced to the quadratic residuosity problem.
The paper then shows that (all modulo the QRA):
Theorem 4: The generator is an unpredictable cryptographically secure pseudo-random sequence generator.
Theorem 5: The sequences produced by the generator pass every probabilistic polynomial time statistical test and that it has the property of unpredictability.
The basis for these properties is that a (probabalistic polynomial time with advantage $\epsilon$) predictor for the generator can be converted efficiently into a predictor of parity for $x_{-1}$ (for arbitrary $x_0$). They then show that such a predictor can be efficiently converted into a procedure for guessing quadratic residuosity (with an amplified $\frac{1}{2}-\epsilon$ advantage.).
The algorithm is a cryptographically secure PRG, provided that the quadratic resuosity problem remains a computationally hard problem under the assumptions made in the proof. According to this answer the assumptions made are easily misinterpreted and for the construction to be secure N needs to be very very large.
However and in any-case, secure PRGs do not inherently make secure cryptographic hash functions. A secure hash algorithm needs to be deterministic, collision resistant and first and second pre-image resistant.
As stated in the comments your algorithm would not make a secure hash function because it does not have the property of collision resistance:
$\forall x,h: h' = -h - x \bmod N $
This also implicitly breaks second pre-image resistance.
Since the hash function is not cryptographically secure it is not suitable for using to generate passwords or to authenticate messages.