Fix agentic VLM image input selection

[suguanYang]

Summary

• Restrict agentic retrieval VLM image_url inputs to chunks whose type is actually image
• Keep image + table media collection unchanged for evidence rendering and table artifact hydration
• Add retrieval API contract coverage proving table artifacts remain present as evidence references but do not invoke VLM

Closes #92.

Verification

• uv run pytest apps/api/tests/contract/test_retrieval_contract.py -k table_artifacts_to_vlm -q
• uv run pytest apps/api/tests/contract/test_retrieval_contract.py -q
• uv run ruff check apps/api/tests/contract/test_retrieval_contract.py packages/shared-python/shared/services/retrieval/agentic/evidence/builder.py packages/shared-python/shared/services/retrieval/agentic/orchestrator.py
• uv run pyright
• uv run ruff check
• git diff --check
• git diff --cached --check
• codex review --uncommitted --title "Retrieval VLM table artifact contract coverage"