[wsdewitt]

Closes #18.

Summary

Tuple-valued node states, and state types that can't be ordered, now work in Tree.evolve.

• HomogeneousProcess.Λ_inv returns a single infinity for a zero rate. It used to return np.full_like(x, np.inf), which for a tuple state gives one infinity per tuple element, so evolve failed when comparing waiting times. Because the default death and mutation processes have rate zero, every tuple-state simulation hit this. For scalar states, the return value is the same as before, array(inf).
• evolve compares only waiting times when choosing the next event (min(..., key=...)). It used to compare (waiting_time, event, state) tuples, so two infinite waiting times led to comparing states, which raised TypeError for states without an ordering, such as a frozen dataclass without order=True.
• ConstantProcess.λ_homogeneous returns one rate per state for a sequence of states, matching DiscreteProcess. For a list of three tuple states it used to return shape (3, 2). Numpy arrays keep their shape, as before.

Seeded results don't change

The event chosen for an infinite waiting time is never used, because the simulation reaches its end time first. Two finite waiting times tie with probability zero, so comparing only waiting times picks the same event as before. As a check, the outputs of about 140 seeded simulations, covering every capacity_method, time-inhomogeneous processes, a Gaussian mutator, and sampling and pruning, are identical to 1.0.0.

Tests

Two new tests fail on 1.0.0 and pass with this change:

• test_tuple_states runs the reproducer from Tuple-valued states make evolve fail (zero-rate waiting time has the wrong shape) #18, which uses tuple states and the default processes. It checks that the resulting tree has the same node times as the same model with a scalar state, and that ConstantProcess returns one rate per tuple state.
• test_unordered_states simulates states that can't be ordered, two of which have zero birth rate, so their infinite birth waiting times tie.

On Python 3.15 locally, all unit tests, doctests, and both notebooks pass, along with ruff, docformatter, and pyright. On Python 3.12, simulation speed is unchanged or slightly faster.

🤖 Generated with Claude Code