Count the Common N-grams

Before embeddings, and still underneath most search, text is counted rather than understood. N-grams are the smallest useful unit of that.

Write ngram_counts(text, n, top_k) returning a list of (ngram, count) pairs.

1. Tokenise: lowercase, split on any run of non-alphanumeric characters, drop empty pieces
2. Build n-grams: every window of n consecutive tokens, joined with a single space
3. Count and rank: most frequent first; ties broken alphabetically; return at most top_k

The tie-break is not decoration. Without it the same input returns different results on different runs, and a ranking nobody can reproduce is a ranking nobody can debug.