[zwcf5200]

Summary

• preserve semantic_rank, bm25_rank, graph_rank, and temporal_rank in recall traces
• accept both the flattened metadata currently emitted by the recall pipeline and the existing nested source_ranks shape
• add a regression test for the flattened metadata path

SearchTracer.add_rrf_merged() currently reads rrf_meta["source_ranks"], while the recall pipeline expands MergedCandidate.source_ranks into the metadata top level. The trace therefore reports an empty source-rank map even though fusion computed the ranks correctly.

This change only fixes trace serialization; it does not change fusion, reranking, strategy boosts, or result ordering.

Validation

• uv run ruff check hindsight-api-slim/hindsight_api/engine/search/tracer.py hindsight-api-slim/tests/test_search_trace.py
• uv run pytest hindsight-api-slim/tests/test_search_trace.py::test_rrf_trace_preserves_flattened_source_ranks -q
• uv run pytest hindsight-api-slim/tests/test_interleave_fusion.py hindsight-api-slim/tests/test_fusion_cap.py hindsight-api-slim/tests/test_recall_boost.py -q (29 passed)
• pre-commit hooks: docs generation and lint passed

[nicoloboschi]

LGTM. Confirmed the bug: the recall pipeline flattens source_ranks into the top level of rrf_meta (memory_engine.py:4142/4246 via **mc.source_ranks), so the old rrf_meta.get("source_ranks", {}) always returned empty. The fallback correctly harvests the *_rank keys. Fast deterministic regression test is the right call for pure serialization logic. Minimal scope — no fusion/ranking/ordering changes.