Record language model token usage

// Configure once in Terminal: aner secrets set openrouter
// Checking makes no provider request. Running sends one bounded request and may incur charges.
import aner.llm
import aner.secrets

fn report_tokens(label: String, count: Int64) {
    if count < 0 {
        print(label + ": unavailable")
    } else {
        print(label + ": " + String(count))
    }
}

let model = llm.client(provider: "openrouter", model: "openai/gpt-4.1-mini",
    key: secrets.stored("openrouter"))
let answer = model.generate("Explain the purpose of validation data.", max_tokens: 256)
print("Provider: " + model.provider())
print("Requested model: " + model.model())
print("Returned model: " + answer.model())
print("Finish reason: " + answer.finish_reason())
report_tokens("Input tokens", answer.input_tokens())
report_tokens("Output tokens", answer.output_tokens())
report_tokens("Total tokens", answer.total_tokens())
report_tokens("Cached input tokens (included in input)", answer.cached_input_tokens())
report_tokens("Cache-write input tokens (included in input)", answer.cache_write_input_tokens())
report_tokens("Reasoning tokens (included in output)", answer.reasoning_tokens())
// Details are subsets: do not add them again to input, output or total.
// Provider-reported tokens use that model's tokenizer; they do not verify money charged.
// Store these per-request counts with provider/model identifiers in your experiment log.