Multi-Continent Financial News Sentiment Dashboard

Fixed in review — two findings flagged critical

• A real, live-format NewsAPI key was hardcoded as a fallback default in config.py, and confirmed present in the git history of both this repo and the monorepo it was split from — meaning it was already public on GitHub. It has been removed from the current code, but that does not remove it from git history: the key should be treated as compromised and rotated at NewsAPI.org.
• The NVIDIA predictor page substituted a fake "neutral" sentiment reading — indistinguishable from a genuine one — whenever the Hugging Face API call actually failed, and fed that fabricated value straight into the prediction model's training data. Fixed to drop failed calls and surface a visible warning with the excluded count.
• A double-escaping bug that silently broke word-boundary stock-name matching for any multi-word company name — "Goldman Sachs Group Inc." never matched; only single-word names happened to work, which masked the bug.
• Ticker matching was case-insensitive, creating false-positive matches for tickers that are also common English words (C, T, MA, NOW, SPOT).
• The Model Performance page called a method that does not exist on its evaluator class, silently failed every time, and displayed meaningless placeholder text instead of real evaluation data — while also overstating how much of the reference dataset was actually used (45 hand-written sentences, not the claimed 500 from the full 4,845-sentence Financial PhraseBank).