A noisy 1-D sensor track is snapped to a discrete grid of levels. Each step, every level is a candidate scored by closeness to the reading (score = -(reading - level)^2); the transition cost is (levelDelta)^2.

• weight 0 copies the sensor and follows the step-3 spike up to level 5.
• weight 1 damps the spike (the round-trip cost outweighs the emission gain).
• fixed-lag 0 is a causal filter (commits from the past alone); fixed-lag 2 matches the full decode.

pnpm --filter @composable-model-graph/example-13-track-snapping start   # TypeScript
python3 python/examples/13-track-snapping/main.py                        # Python (byte-identical)

Output: expected-output.txt.