輝達2026財年預測：端到端測試流程

You are running a source-grounded MadCowork Office E2E workflow.
Communicate final user-facing conclusions in Traditional Chinese, but all tool planning and tool arguments should be concise English.

Topic: NVIDIA fiscal 2026 forecast (輝達2026財年預測).

Hard requirements:
1. Do not rely on model memory. Use web_search, web_fetch, and web_crawl as needed before making factual or numerical claims.
2. Gather evidence from multiple sources, including NVIDIA investor relations or SEC filings / earnings releases and analyst or market forecast sources.
3. Create CSV files in this session workspace:
   - nvda_sources.csv with columns: source_type,title,url,published_date,used_for.
   - nvda_financials.csv with columns: company,metric,period,value,unit,source_url,note.
   - nvda_analyst_forecasts.csv with columns: provider,metric,period,forecast,unit,source_url,note.
   - session_metadata.csv with columns: artifact,type,path,created_at,source_inputs.
4. Every numeric claim used in the deck must have a source_url in one of the CSV files.
5. Generate at least two chart-ready CSVs or chart JSON snippets from the CSV facts. Charts must be derived from CSV values, not invented.
6. Use xdeck_generate to create a self-contained web presentation at nvda-fy2026-web-deck/deck.html.
7. The deck should be 7 slides, Traditional Chinese, investor / strategy briefing style, with speaker_note on each slide.
8. Run xdeck_audit and xdeck_screenshots. If the audit finds layout issues, run one xdeck_revise pass.
9. Final answer: list generated artifact paths and briefly explain the evidence chain.

Important: if a source cannot be fetched, record the failed URL in nvda_sources.csv with used_for="attempted but unavailable" and continue with fetchable sources.