perf: avoid eager job chunk hydration during retrieval

[suguanYang]

Summary

• select only JobResult.job_id during final reference hydration
• avoid implicitly loading the unused JobResult.chunks relationship
• preserve revision-pinned hydration and public job_id projection

This removes unnecessary ORM work from the final retrieval stage without changing prompts, ranking, or evidence selection.

Validation

• uv run pytest -q apps/api/tests/contract/test_retrieval_snapshot_consistency_contract.py apps/api/tests/contract/test_retrieval_mapnav_session_contract.py apps/api/tests/contract/test_retrieval_contract.py — 15 passed
• uv run ruff check packages/shared-python/shared/services/retrieval/hydration/reference.py
• uv run ruff format --check packages/shared-python/shared/services/retrieval/hydration/reference.py
• uv run pyright packages/shared-python/shared/services/retrieval/hydration/reference.py
• Local hydration harness against the production read-only database completed successfully; no production data was modified.

Notes

The map-nav LLM episode remains the dominant end-to-end latency source and is outside this change.