[njtierney]

Closes #617.

• TensorFlow's Poisson sampler on the CPU never returns for a NaN rate, and Ctrl-C doesn't stop it. TFP's GPU path masks NaN rates; its CPU path doesn't.
• poisson() samples through it, and so does negative_binomial(), as a Poisson with a gamma-distributed rate. So calculate() and simulate() hung on a NaN or negative lambda (greta passes log(lambda)), and on a negative binomial whose rate is NaN.

What changes

• A new internal tf_sample_where_valid() swaps each invalid parameter element for a safe placeholder before sampling, and gives NaN for those draws. When every element is valid, the draws are the same as on main, seed for seed.
• poisson() masks a NaN log rate, which covers a NaN or negative lambda.
• negative_binomial() masks a NaN or negative size, a TFP probs = 1 - prob outside [0, 1), and an infinite size with prob = 1. So prob = 0, or a prob so small that 1 - prob rounds to 1, draws NaN. An infinite size with prob below 1 still draws Inf, as on main.
• Both distributions return list(log_prob, sample), like greta's other custom distributions. log_prob is TFP's, unchanged.

Not covered

• greta.distributions' zero-inflated Poisson and negative binomial build their own TFP samplers, so they still hang. An issue there is drafted, not posted.
• Every other TFP sampler greta uses was probed on the CPU with NaN, ±Inf, negative and huge parameters, and none hang. Some give a silently wrong draw, e.g. binomial() with a NaN size gives 0; that's a separate issue, also drafted.

Tests

• tests/testthat/test_calculate.R: calculate() draws NaN for invalid poisson() and negative_binomial() elements, including negative_binomial(0, 0) and an infinite size with prob = 1, and ordinary draws for the valid elements beside them. On main the test hangs, and its session is killed at 600 s.
• tests/testthat/helpers.R: in_fresh_greta() takes a timeout, so a test of a hang fails rather than hangs. The timeout also covers starting R and loading greta, hence 600 s.
• The calculate, iid_samples and distributions tests pass. So does the whole suite locally, apart from two test_future.R errors: future workers load the installed greta, which predates %*% multiplies a data matrix once, not once per chain (#858) #859's unbatched_tf_name().

🤖 Generated with Claude Code