Get "infinite gradient" error for code with finite gradient

Summary:

I have a small example program that dies with "Rejecting initial value: Gradient evaluated at the initial value is not finite" even though the log probability has a finite gradient everywhere.

Description:

This is illustrated in the same example program as given in issue #274. Referencing the model code given there, I get the error if I keep the assignment to z_lo, but not if I remove the assignment to z_lo. But z_lo never enters into the log probability in any way!

The example may seem contrived, but it is boiled down from a more reasonable example in which z_lo is always computed, but never used in the cases for which lo is infinite.

Reproducible Steps:

See issue #274.

RStan Version:

2.9.0

R Version:

3.2.1

Operating System:

Mac OS X El Capitan

Start with the example program and model code referenced in issue #274, then reproduce the difference between keeping and removing the z_lo assignment. Trace why the finite-gradient check rejects the initial value despite z_lo not entering the log probability; done means the cause and affected behavior are understood and the error no longer occurs for this case.