Compact MCP evidence query responses

[mahikaw]

PR description:

Summary

• Compact format="evidence" MCP query responses by trimming verbose evidence text
• Preserve all evidence items and citation metadata so agents can still see top-k results
• Add coverage that verifies long evidence chunks are compacted without dropping hits

Why

MCP evidence responses can exceed the agent-visible tool output budget because each hit includes full chunk text. In Harbor/
Codex evals this caused function_call_output truncation, so the agent sometimes saw only a subset of the retrieved top-k
candidates even though the backend returned all 10.

This change keeps citation/source/page/score metadata intact and caps evidence text length server-side, including when an
agent asks for a larger text budget.

Testing

• uvx pre-commit run --files nemo_retriever/src/nemo_retriever/service/mcp_server.py nemo_retriever/tests/ test_service_mcp.py
• uv run --project nemo_retriever --no-dev --extra service --extra dev pytest nemo_retriever/tests/test_service_mcp.py -q

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.