feat(cli): add Agent-friendly output flags (-q, -f json, --with-sources) to query command

[als3453]

Summary

Adds 3 new flags to the rag-python query CLI command, enabling the tool to be callable as a structured Agent tool (via subprocess + json.loads()), while remaining fully backward-compatible.

New flags

Flag	Purpose
-q / --quiet	Print the answer only — no evaluation / sources trailers. Script-friendly.
-f / --output-format {text,json,json-pretty}	Return structured JSON {text, sources, evaluation, retried} — designed for Agent tool use.
--with-sources	Print the top-5 sources trailer without the verbose evaluation section.

Example: calling as an Agent tool

import json, subprocess
out = subprocess.check_output([
    "rag-python", "query", "What is the PTO policy?",
    "-f", "json", "--quiet", "--embedding-provider", "local"
])
answer = json.loads(out)
print(answer["text"])
print([s["metadata"].get("source") for s in answer["sources"]])

Why this change (motivation)

Currently, Agent orchestrators (LangChain / CrewAI / custom ReAct loops) that want to wrap rag-python as a Tool / Function Calling payload have to:

1. Parse human-printed text (regex-heavy and brittle), or
2. Write their own wrapper script — duplicating the provider/search argument surface.

This PR turns the CLI into a first-class structured output endpoint with 0 breaking changes to the default user experience.

Testing

Added 8 hermetic unit tests under tests/test_cli_query_agent_output.py:

• A1 — --help lists all 3 new flags and their short aliases
• A2 — --quiet prints the answer only (no trailers)
• A3 — -q short alias equals --quiet
• A4 — -f json produces valid JSON with 4 keys + full sources list
• A5 — -f json-pretty produces valid, indent=2 pretty-printed JSON
• A6 — --with-sources prints sources trailer without requiring -v
• A7 — default text output + -v verbose behavior unchanged (backward compat)
• A8 — streaming with --quiet emits tokens only; no trailers after

Plus regression checks (--version, docs quickstart, docs --list) all pass.

Full TDD: RED (8/8 failed on unmodified main) → GREEN (8/8 pass after this change) → REFACTOR (0 ruff line-length / unused-import violations).

Checklist

• Code follows the project's style (uses same dataclasses.replace pattern for search, ruff-clean)
• New tests pass
• Backward-compatible (default rag-python query "..." unchanged)
• Help text updated with Agent usage examples in the query epilog
• No new dependencies (stdlib only: argparse, json)

---

Closes # — (no matching issue; this is a UX/toolability improvement)