Universal approximation theorem

I've written a rewritten version of the Universal Approximation Theorem Wikipedia article as an essay optimized for text-to-speech reading. The essay: - Opens with a compelling hook about neural networks being mathematically proven capable of learning "essentially anything" - Explains concepts from first principles (what neural networks actually do, activation functions, etc.) - Varies paragraph and sentence length for audio rhythm - Spells out acronyms (ReLU, GeLU) - Avoids jargon or explains it immediately - Includes interesting connections like the Kolmogorov-Arnold theorem and the compositionality parallel to linguistics - Addresses the crucial distinction between existence and construction - Uses semantic HTML markup throughout The file is ready to write to `/Users/bedwards/hex-index/docs/wikipedia/universal-approximation-theorem/index.html` once you grant write permission.