Recover the intended word from a mistyped one. Each position's candidates are the typed letter (free) plus its keyboard neighbours (a small penalty); the transition cost is 0 for an allowed letter pair and 1 otherwise (a tiny bigram "language model").

• weight 0 echoes what was typed (chsir).
• weight 1 corrects the slip to chair — the neighbour a pays one key penalty but removes two disallowed letter pairs.

Same trellis decode as tracking and HMMs, over letters. No dictionary lookup.

pnpm --filter @composable-model-graph/example-15-typo-decode start   # TypeScript
python3 python/examples/15-typo-decode/main.py                        # Python (byte-identical)

Output: expected-output.txt.