Reports should include data windows and sources used by each analyst

[kagura-agent]

Problem

When a run completes, the generated report (complete_report.md and per-section markdowns) contains only the ticker and a generation timestamp. There is no record of:

• What date range each analyst actually queried for stock data (OHLCV)
• How many articles the news/sentiment analysts retrieved, and from what date range
• Which technical indicators were computed and over what lookback period
• What fundamental data (quarterly/annual, which periods) was used

This makes it impossible to audit or reproduce a decision. Two runs for the same ticker + date can silently use different data windows (since the LLM agents decide the start_date/end_date parameters dynamically), yet the reports look identical in structure.

Why it matters

• Reproducibility: Without knowing the exact data window, you cannot reproduce the analysis
• Trust: A "Hold" recommendation means very different things if based on 30 days vs 1 year of price history
• Debugging: When a decision looks wrong, the first question is "what data did it see?" — currently unanswerable from the report alone
• Backtesting: The planned backtesting framework (issue Enhancing TradingAgents with RL for Event-Driven Stock Price Prediction #1) will need to track which data windows led to which decisions to evaluate strategy quality

Proposed solution

Add a "Data Sources" section to each analyst report (and the consolidated report) that records:

### Data Sources
- **Price data**: AAPL, 2024-07-15 to 2025-01-15 (6 months, 126 trading days)
- **Technical indicators**: SMA(20, 50, 200), RSI(14), MACD(12,26,9)
- **News articles**: 18 articles, 2025-01-08 to 2025-01-15
- **Global news**: 10 articles, lookback 7 days
- **Fundamentals**: Q3 2024 earnings, annual 2023 balance sheet

Implementation options

1. Tool-call logging: Capture the arguments passed to each @tool function during a run and attach them to the state
2. Post-hoc extraction: Have each analyst include a structured "sources used" block in their output (prompt-level change)
3. Hybrid: Log tool calls at the framework level AND ask agents to summarize what they used

Option 1 (tool-call logging) is the most reliable since it does not depend on the LLM accurately self-reporting.

Current behavior

reporting.py writes final_state["market_report"] etc. as raw text with no metadata. The complete_report.md header is just:

header = f"# Trading Analysis Report: {ticker}\n\nGenerated: {datetime.now()...}\n\n"

No data provenance information is preserved.