XY-845: [ELF benchmark suite] Add retrieval quality and hierarchical routing cases

[yvette-carlisle]

Summary

• add real-world retrieval fixtures for alternate phrasing, distractors, multi-hop routing, current-vs-obsolete selection, minimal sufficient context, and stage-level explainability
• extend real_world_job reports with expected evidence recall, irrelevant context ratio, latency/cost, and trace-stage attribution while preserving memory-evolution and work-resume report counters
• add cargo make real-world-memory-retrieval plus benchmark docs/spec routing that keeps qmd/OpenViking reference-only unless adapters run

Validation

• cargo test -p elf-eval --test real_world_job_benchmark
• cargo make real-world-memory
• cargo make real-world-memory-retrieval
• cargo make fmt
• cargo make lint-fix
• cargo make checks
• ELF_BASELINE_PROJECTS=ELF cargo make baseline-live-docker