Token cost tracking for DeepSeek Harness

Key features

• Prices every finalized LLM call (assistant/message usage) against a per-model pricing table you configure — cache-hit / cache-miss input, output, and optional peak-window rates.
• Cache writes are billed at the miss rate, not the hit rate.
• Publishes a per-session cost projection (total / peak / off-peak / per-model), consumed by the Web UI to show a live session cost readout.
• Flags calls that land in a configured peak window.
• Models without a pricing entry are surfaced as "unconfigured" with a placeholder, so you know to add them.