Add Hypothesis property-based tests for core tHMM invariants

[aarmey]

Summary

• Adds hypothesis as a dev dependency and a new lineage/tests/test_properties.py with 13 property-based tests covering invariants of the tHMM machinery that the existing example-based tests don't exercise across a range of inputs:

  • CellVar.divide: daughter states are always valid, parent/gen bookkeeping is correct, and results are reproducible for a given RNG seed, for any row-stochastic transition matrix.
  • BaumWelch.calculate_stationary: the returned vector is always a valid probability distribution and a genuine fixed point (w @ T == w) of any stochastic T.
  • stateCommon.bern_estimator: the weighted Bernoulli MAP estimate always stays in [0, 1].
  • StateDistributionGamma.StateDistribution.logpdf/dist: negative "unobserved" sentinel rows always contribute exactly zero log-likelihood, and a distribution's distance to itself is zero.
  • LineageTree.rand_init: for any (pi, T) and requested size, the produced tree is a valid full binary tree (every non-leaf has exactly 2 children) with valid state indices, and simulation is reproducible for a given seed.
  • Viterbi/tHMM.log_score: decoded state sequences are always valid and always score at least as well as an arbitrary random assignment of the same lineage.

• One test (test_gamma_logpdf_never_produces_nan) is a strict, pinned xfail that documents a real bug Hypothesis found: StateDistribution.logpdf can raise FloatingPointError instead of returning a finite log-likelihood. This happens because lineage/__init__.py sets np.seterr(all="raise"), and an "unobserved" Bernoulli sentinel (-1) on a row with an uncensored, exactly-zero Gamma observation under a shape parameter < 1 (where the Gamma density diverges to +inf at x=0) produces a (+inf) + (-inf) = NaN inside logpdf, which the current code only zeroes out after the NaN has already been created. This is left as a documented xfail rather than fixed here, since correcting the ordering of the masking logic touches numerically sensitive code that's out of scope for a test-only PR.

Test plan

• uv run pytest lineage/tests/test_properties.py -v — 12 passed, 1 xfailed (repeated 3x to confirm no flakiness)
• uv run pytest lineage/tests -q — full existing suite still green: 66 passed, 1 xfailed
• uv run ruff check / ruff format --check on the new file
• uv run ty check lineage/tests/test_properties.py