feat: rank and filter logic for biolink:statistical_significance_qualifier

[SkyeAv]

Adds ranking and KG-level filtering support for the biolink:statistical_significance_qualifier (biolink/biolink-model#1766), which categorizes an edge's evidence into coarse significance bands (very_strongly_significant through not_significant) as a companion to the authoritative numeric p value / adjusted p value slots. Closes #2858.

Design note: the statistical_significance_qualifier is is_a: statement_qualifier (a descendant of qualifier in biolink-model), so BMT routes it into edge.qualifiers[]. It is treated only as a qualifier here — never as an edge attribute. (See the decision documented in #2858.)

Ranker

• Additive scoring signal: edge_attribute_score_combiner() in ARAX_ranker.py now injects a significance-band score into the existing loop algorithm (W_r = W_r + (1 - W_r) * W_i) alongside numeric p-value and other attributes — the qualifier can only boost, never penalize.
• Conservative trust weight: Held in a standalone _significance_trust_weight constant at 0.5, with band scores ranging 0.0–0.70 (max effective contribution 0.35). TODO comment marks this for revisit once adoption is widespread.
• Categorical bypass: Added _get_significance_qualifier_value(self: Self, edge: Edge) -> Optional[str] to look up the qualifier before the float() conversion in edge_attribute_score_normalizer() (which would otherwise ValueError on enum strings and silently return 0).
• Qualifier-only lookup: Reads the value exclusively from edge.qualifiers (the TRAPI path for biolink qualifier descendants), stripping any biolink: prefix from the value.

KG Filter

• New action: remove_edges_by_statistical_significance in ARAX_filter_kg.py + Filter_KG/remove_edges.py — an ordinal threshold filter, not exact-match. minimum_significance=significant removes suggestive and not_significant edges while keeping significant and above.
• DSL: filter_kg(action=remove_edges_by_statistical_significance, minimum_significance=suggestive) removes only not_significant; supports remove_connected_nodes, qedge_keys, and qnode_keys.
• Registered: Added to allowable_actions, command_definitions, and parameter descriptors matching the existing remove_edges_by_discrete_attribute pattern.

Design

• Qualifier, not attribute. The qualifier is a biolink qualifier descendant, so BMT routes it into edge.qualifiers[]. It is read only from there — it is not registered as an edge attribute in the decorator and is not read from edge.attributes anywhere.
• Edges without the qualifier are never penalized. The ranker only boosts positive bands; the filter only removes edges that explicitly carry a below-threshold value. This is intentional during rollout — penalizing would unfairly downrank sources that haven't adopted the qualifier yet.
• Deferred: Result-level filtering (ARAX_filter.py) — edge-level KG filtering is the right granularity. Significance-aware pathfinding is a candidate for future work.

Testing

• Unit verification of ARAX_ranker.py — score mapping, qualifier lookup (edge.qualifiers/prefixed), edge cases (multiple qualifiers, empty lists, unrecognized values), additive-with-pValue → all PASS
• Unit verification of Filter_KG/remove_edges.py — ordinal removal, threshold boundaries (very_strongly_significant removes 4/5, not_significant removes 0/5), qualifier-less edges kept, command_definitions registration → all PASS
• Slow integration tests (@pytest.mark.slow) included that query live infores:retriever end-to-end; these are forward-looking until BioPack-team/retriever#197 ships.