There is a rationale here and it has to do with the connection between diversification and the number of assets in a portfolio.
Suppose we purchase an equal-weighted portfolio of n stocks. The variance of the return is then:
$\sigma_{p}^2$ = $\sum$ $\sum$ $w_i$$w_j$Cov($R_i$,$R_j$)
In the above notation, the sigmas are summing over $i$ and $j$ respectively each element of the variance-covariance matrix of asset returns.
Since each weight is 1/n, then this equation is equal to:
1/$n^2$ * $\sum$ $\sum$ Cov($R_i$,$R_j$).
Let's say the average variance across all stocks is $\sigma_{avg}^2$ and the average covariance between among all pairs of stocks is $Cov_{avg}$.
Then this equation simplifies to the sum of two products (see Investments, Bodie, Kane, and Marcus for the derivation):
$\sigma_{p}^2$ = $(1/n)$ * $\sigma_{avg}^2$ + $[(n-1)/n]$ * $Cov_{avg}$.
Let's analyze the first product. As the number of stocks increases the contribution of the variance of the individual stocks becomes very small because the term $(1/n)$ * $\sigma_{avg}^2$ approaches zero as $n$ becomes large.
Now let's analyze the second product. As the contribution of the average covariance across strocks to the portfolio variance stays non-zero because the term $[(n-1)/n] $ * $Cov_{avg}$ has a limit of $Cov_{avg}$.
So as the number of assets increases the portfolio variance is approximately equal to the average covariance.
In your above example, if there are many such assets then the contribution to variance is vary small. Also, if the assets are "uncorrelated with others and the market" then the second term is also small. Therefore the variance is nearly zero when both the conditions identified by Luenberger are met.
Here is a more intuitive explanation. A Beta of 0 does not imply zero variance, securities still have idiosyncratic risk (i.e. a random component of return not explained by systematic exposure). A risk-free investment is still less risky than a security with a beta exposure of zero although they both have the same expected return. However, investing in a large portfolio with a beta exposure of zero is far less riskier than investing in a single security with a beta exposure of zero.
Adding the idiosyncratic error term 'e' to the formula you identify captures this idea of diversifying away residual risk. 'e' is a random variable whose expectation is 0 but variance is greater than zero. Only as one builds larger portfolios is this 'e' term increasingly diversified away (i.e. the variance of the error term approaches zero):
$$\bar{r}_i - r_f = \beta_i (\bar{r}_M - r_f) + e $$
Put another way, if the number of securities is large (and weights are small), the covariances become the most important driver of portfolio variance. Consider the variance-covariance matrix of the asset returns. For an 'n' asset portfolio there are 'n' variances (the diagonal), and n*(n-1)/2 covariances (the off-diagonal triangle). A portfolio of two stocks has 2 variances and 1 covariance (i.e. variances dominate). A large portfolio of 1,000 stocks has 1,000 variance terms but 499,950 co-variance terms (i.e. covariance dominates).