Incorrect leftindexing memory estimate w/ empty input in loops

The current memory estimate of leftindexing operations is incorrect (in terms of a dangerous underestimation that leads to OOMs) for empty input matrices in loops. For example, consider the following script

R = matrix(0, N, r^2);
parfor(i in 1:nrow(X)) {
  Xi = X[i,]
  R[i,] = matrix(t(Xi)%*%Xi, 1, ncol(X)^2)
}

In this scenario, the memory estimate of R[i,] is mistakenly set to a single row, which is only valid during dynamic recompilation with constant indexing range quantifiers.