Build a small `cache_decision(hit_ratio, origin_cost_ms, cache_cost_ms)` that returns either `"cache"` or `"origin"`: - Expected latency for "cache" = `hit_ratio * cache_cost_ms + (1 - hit_ratio) * (cache_cost_ms + origin_cost_ms)` - Expected latency for "origin" = `origin_cost_ms` - Pick whichever is lower; break ties in favour of `"origin"` (simpler architecture wins on tie). Inputs: - `hit_ratio` is a float 0.0 to 1.0 - costs are positive ints (ms)