[Autoplectic]

Summary

_prune_to_stationary_support made two passes and finished by restricting the graph to the stationary support. That restriction can drop more joint states (masses below tol), and nothing then re-checked that the survivors still had complete outgoing rows. With a truncated reverse machine (here 167 reverse states, smallest stationary mass 8.6e-12), validation failed with joint masses from (2, 'CE') sum to 0.864643100703771.

The prune now alternates "keep states with complete rows" and "keep the stationary support" until neither removes anything. On the failing 4-state survey process the build now succeeds (classified truncated), with forward marginals matching to 1e-10 and the symbol law to 1e-11.

Test plan

• New regression test test_truncated_reverse_machine_prunes_to_complete_rows; it fails on main and passes here.
• Bidirectional, generator and properties suites: 168 passed.
• ruff.

Made with Cursor