Expand CashFlowStream edge-case and contract test coverage

[j-bryan]

The dedicated CashFlowStream tests cover the main examples well, but several public branches and behavioral contracts are not exercised directly. These are test coverage limitations, not confirmed production defects.

A focused branch-coverage run identifies unexecuted paths in CashFlowStream.filter(), CashFlowStream.group_by(), and the group_by_period() convenience method. Inspection also shows several documented collection properties that currently rely on a single example rather than boundary or relationship checks.

Suggested deterministic coverage

Filtering

• Calling filter() with neither a predicate nor keyword criteria raises ValueError.
• Combining a callable predicate with keyword criteria raises ValueError.
• Multiple keyword criteria use AND semantics.
• is_cash=False selects non-cash entries.
• pro_forma_category=None selects uncategorized entries.
• String classification inputs are normalized consistently.

These tests would reject defects such as accidentally changing AND semantics to OR or treating False as an omitted argument.

Grouping

• Calling group_by() with neither selector raises ValueError.
• Providing both a callable and period raises ValueError.
• group_by_period(period) produces the same groups as group_by(period=period).
• Grouping preserves the total entry count and all duplicate entries.

Ordering and ranges

• Sorting is stable when multiple cashflows have equal keys.
• Negative indices and stepped or reversed slices match ordinary Python list behavior.
• date_range(start, end) includes an entry exactly on start and excludes an entry exactly on end.
• date_range(start, start) returns an empty stream.
• Zero-valued cashflows are in neither inflows() nor outflows().

Transformations

• scale() changes only amount and preserves date, label, is_cash, pro-forma category, and tax treatment.
• Scaling by 1, 0, and a negative factor covers identity, zero, and sign reversal.
• apply(), flat_apply(), and filter_apply() preserve the documented output ordering.
• Every non-mutating method leaves the original entry sequence unchanged.

Automatic input exploration

After these properties are explicit, bounded property-based tests could compare:

• filter(predicate) with a plain list comprehension;
• sort() with Python sorted();
• date_range() with the half-open interval predicate;
• grouped entry counts and duplicate-preserving reconciliation;
• scaled entry metadata with the original metadata.

Generated inputs should explore empty, singleton, duplicate, equal-key, zero, mixed-sign, and boundary-date cases. Property-based generation should supplement these deterministic oracle cases rather than replace them.

Acceptance criteria

• Add direct tests for the uncovered selector and convenience-method branches.
• Add boundary tests for half-open date ranges and zero-valued flows.
• Add metadata-preservation and stable-order tests.
• Keep each test focused on one externally meaningful property.