We employ grant-based access with retransmissions for multiple users with small payloads, particularly at low spectral efficiency (SE). The radio resources are allocated via NOMA in the time into $T$ slots and frequency dimensions, with a measure of non-orthogonality $η$. Retransmissions are stored in a receiver buffer with a finite size $C_{\sf buf}$ and combined via HARQ, using Chase Combining (CC) and Incremental Redundancy (IR). We determine the best scaling for the SE (bits/rdof) and for the user density $J/n$, for a given number of users $J$ and a blocklength $n$, versus SNR ($ρ$) per bit, i.e., the ratio $E_b/N_0$, for the sum-rate optimal regime and when the interference is treated as noise (TIN), using a finite blocklength analysis. Contrasting the classical scheme (no retransmissions) with CC-NOMA, CC-OMA, and IR-OMA strategies in TIN and sum-rate optimal cases, the numerical results on the SE demonstrate that CC-NOMA outperforms, almost in all regimes, the other approaches. In the sum-rate optimal regime, the scalings of $J/n$ versus $E_b/N_0$ deteriorate with $T$, yet from the most degraded to the least, the ordering of the schemes is as (i) classical, (ii) CC-OMA, (iii) IR-OMA, and (iv) CC-NOMA, demonstrating the robustness of CC-NOMA. Contrasting TIN models at low $ρ$, the scalings of $J/n$ for CC-based models improve the best, whereas, at high $ρ$, the scaling of CC-NOMA is poor due to higher interference, and CC-OMA becomes prominent due to combining retransmissions and its reduced interference. The scaling results are applicable over a range of $η$, $T$, $C_{\sf buf}$, and $J$, at low received SNR. The proposed analytical framework provides insights into resource allocation in grant-based access and specific 5G use cases for massive URLLC uplink access.