Particle Gibbs consistently fails to solve a 20-point normal-normal model. Is this expected?

[axch]

The model in question is

[assume a (normal 0 10)]
;; 20 times
[observe (normal a 1) (normal 5 1)] ; Generate a random data set
;; Expect the posterior on a to be close to (normal 5 1)

Defined in test/inference_language/test_compound.py:test_model_without_compound_assume. Both Lite's and Puma's particle Gibbs fail to get this on default settings, e.g. (pgibbs default ordered 4 10 false) for Puma.