Filter-context auto-aggregation across methods (#118)

[iskandr]

Summary

Closes #118. Inside apply_filter, a Comparison with unqualified same-kind refs auto-aggregates across methods instead of raising the ambiguity error:

# On a frame with both mhcflurry and netmhcpan affinity rows:
apply_filter(df, Affinity <= 500)   # no raise
# Keeps groups where ANY method has value <= 500.

Aggregator is picked by the comparison direction:

Operator	Aggregator	Equivalent OR expansion
<, <=	nanmin across methods	(aff[m1] < k) | (aff[m2] < k) | ...
>, >=	nanmax across methods	(aff[m1] > k) | (aff[m2] > k) | ...

Scope (narrow on purpose)

Auto-agg fires only when all of these hold:

1. The node is a Comparison with <, <=, >, or >=.
2. It's evaluated under a filter context (apply_filter sets EvalContext.filter_context=True).
3. All unqualified Field refs in the comparison are the same kind. Cross-kind (affinity.rank <= processing.rank) stays strict.
4. The DataFrame actually has >1 method for that kind. Single-method frames fast-path through strict eval.

apply_sort, scalar score arithmetic (0.5 * ba.score + 0.5 * el.score), == / !=, and qualified refs (Affinity['netmhcpan'] <= 500) all keep the current strict behavior.

Changes

• EvalContext(df, filter_context=False) — public kwarg. apply_filter sets it to True.
• EvalContext._method_override — internal (kind_value, method) tuple the Comparison auto-agg loop sets when iterating per method.
• Field.eval honors ctx._method_override when self.method is None and kind matches.
• Comparison.eval — new _should_auto_aggregate gate + _auto_aggregate loop that evals both sides per method and aggregates LHS via nanmin/nanmax and RHS via the opposite (for the "any method pair" union interpretation).
• _collect_unqualified_kinds(node) — tree walker.
• CHANGELOG + version bump to 5.10.0.

Back-compat

One test in tests/test_io_lens.py previously asserted apply_filter(r.df, Affinity <= 500) raised on a multi-model LENS frame. That was the behavior being relaxed. Replaced with:

• test_unqualified_autoaggregates_in_filter — verifies no raise + valid shape.
• test_unqualified_ambiguous_still_raises_in_sort — verifies the sort path keeps strict.

No other callers were relying on the filter-side raise.

Test plan

• <= / < auto-agg via nanmin; >= / > via nanmax
• "Any method passes" — group kept; "no method passes" — group dropped
• apply_sort still raises Ambiguous on unqualified multi-method refs
• Bare (Affinity.value <= 200).evaluate(df) outside filter still raises
• Cross-kind comparison (Affinity.rank <= Processing.rank) stays strict
• == / != stay strict
• Single-method frame: no auto-agg overhead, same result
• Qualified side (Affinity['mhcflurry'] <= 200) bypasses auto-agg
• Compound boolean ((aff <= 200) & (aff.score >= 0.5)) auto-aggs each comparison
• Const >= Field (Const on left) mirrors correctly
• Full suite: pytest tests/ → 1192 passed, 3 skipped

[coveralls]

coverage: 88.589% (+0.1%) from 88.468% — filter-auto-aggregate into master