Derive frozen-rows transition mask lazily so it stays out of vars()/get_params()

[edeno]

Problem

_DetectorBase.__init__ stores _frozen_discrete_transition_rows_mask_ -- a
boolean row mask derived from the frozen_discrete_transition_rows constructor
argument -- as an instance attribute. Because it is a normal instance attribute,
it appears in both vars(model) and get_params(), even though it is not a
constructor parameter.

This breaks serialization round-trips that reconstruct a model from its
attributes. For example, Spyglass's DecodingParameters stores vars(model)
and later reconstructs the model with Detector(**params); the derived mask is
passed back into __init__ and raises:

TypeError: __init__() got an unexpected keyword argument '_frozen_discrete_transition_rows_mask_'

Any consumer that treats a model's public attributes as its constructor
parameters hits the same problem, and it will recur whenever a new derived
attribute is added in __init__.

Solution

Compute the mask lazily via a read-only _frozen_discrete_transition_rows_mask_
property (recomputed from frozen_discrete_transition_rows and the state count
on access) instead of storing it as an instance attribute in __init__. The
up-front validation of frozen_discrete_transition_rows against n_states is
kept, so invalid input still raises at construction time. The derived value is
unchanged; it simply no longer appears in vars() / get_params().

Validation

• Existing constructor-plumbing and slow end-to-end EM frozen-row tests pass
  unchanged (behavior preserved, including byte-exact frozen rows after EM).
• New tests assert the mask is absent from vars() and get_params() for the
  default models, that the property still returns the correct mask (and None
  when no rows are frozen), and that a ContFrag model's vars() now
  reconstructs through the base detector.
• ruff check / ruff format clean.

Risks / possible issues

• The mask is recomputed on each access rather than cached. It is read only a
  handful of times per fit/predict (discrete transition-matrix construction) and
  _normalize_frozen_discrete_transition_rows is inexpensive, so there is no
  meaningful performance impact; it is not on a per-timestep hot path.
• Validation timing is unchanged: invalid frozen_discrete_transition_rows is
  still rejected at construction.
• This change alone fixes vars()/get_params()-based consumers for the base
  and ContFrag models. The NonLocal* models additionally expose subclass-only
  parameters (non_local_position_penalty, non_local_penalty_std); a paired
  Spyglass change reconstructs via the concrete model class so those round-trip
  as well. The two changes are independently beneficial.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com