Pi: report one-hour cache writes in usage
Our Pi stream set cacheWrite but not Pi's native cacheWrite1h, so Pi priced every cache write as a five-minute write. It now sets cacheWrite1h from usage.cache_creation.ephemeral_1h_input_tokens, as Pi's own Anthropic provider does. Agent Runner bills Pi forks from this field.
Summary by cubic
Pi now prices one-hour prompt-cache writes at the one-hour rate instead of the five-minute rate.
Previously stream.ts set cacheWrite from cache_creation_input_tokens but never set cacheWrite1h, so every Pi cache write was billed as a five-minute write. The stream now reads usage.cache_creation.ephemeral_1h_input_tokens and sets cacheWrite1h, matching how Pi's own Anthropic provider reports it. Agent Runner bills Pi forks from this field, so fork costs now reflect actual usage.