test(metrics): add overlapping-chunk regression fixtures for ContextualRecallMetric (closes #2788)

[Ruthwik-Data]

Summary

Adds tests/test_metrics/test_contextual_recall_overlapping_chunks.py — regression test fixtures for ContextualRecallMetric overlapping-chunk behaviour.

Closes #2788.

This is the symmetric complement to PR #2787 (which addresses the same issue for ContextualPrecisionMetric). The tests document the failure mode and will serve as regression targets once the source-grouping fix is applied to ContextualRecallMetric.

Motivation

PR #2743 added _group_retrieval_contexts() (source-grouping deduplication) to ContextualPrecisionMetric to fix issue #2594. The same fix was not applied symmetrically to ContextualRecallMetric, leaving an identical failure mode:

• RAG pipelines with 10–20% sliding-window chunk overlap (standard for dense financial documents) produce chunks where the same information appears in adjacent chunks.
• ContextualRecallMetric scores each chunk independently. If the LLM judge returns yes for the first chunk and no for the second (partial redundancy), the recall score is halved — even though the expected output is fully covered.
• Redundancy ≠ Missing coverage.

What this PR adds

Fixtures

Fixture	Scenario
overlapping_revenue_chunks	Two same-source overlapping chunks from 10-K MD&A
multi_statement_expected_output	Three-statement expected output spanning two sources
retrieval_context_multi_source	Two overlapping 10-K chunks + one earnings-call chunk

Tests

Test	Guards against
test_same_source_overlap_does_not_lower_recall	Recall penalised for same-source chunk redundancy
test_multi_source_recall_not_inflated_by_overlap	Overlapping chunks distorting multi-statement recall
test_increasing_overlap_does_not_decrease_recall	Monotonicity: more context from same source → recall stays stable

Design notes

• Tests use RetrievedContextData(content=..., source=...) — post-feat(contextual-precision): add RetrievedContextData source grouping and fix weighted precision score #2743 API.
• threshold=0.0 in the monotonicity test isolates scoring logic from pass/fail cutoffs.
• Tests are written as regression targets — they document the expected post-fix behaviour.
• No new dependencies.

Type of change

• Test (adding missing tests or correcting existing tests)

[vercel]

@Ruthwik-Data is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.