ccusage hermes reports $0 cost when Hermes session uses subscription billing #1112

What happened?

Every row of ccusage hermes daily shows $0.00 even though token counts are clearly non‑zero. In my case ccusage sums to 10.2M input / 790K output / 267M cache‑read tokens (model gpt-5.5) but the cost column is $0 for every day.

Tracing it down: ccusage reads estimated_cost_usd / actual_cost_usd from Hermes' state.db. When Hermes is pointed at a subscription endpoint (in my config: provider: openai-codex → base_url: https://chatgpt.com/backend-api/codex), Hermes intentionally writes estimated_cost_usd = 0.0, actual_cost_usd = NULL, billing_mode = subscription_included, because those tokens are covered by the user's ChatGPT Plus/Pro subscription rather than metered.

calculate_hermes_cost in rust/crates/ccusage/src/adapter/hermes.rs short-circuits on Some(cost) regardless of whether the recorded cost is zero, so it returns 0.0 and never falls through to pricing-based calculation. The codex adapter doesn't have this short‑circuit and prices the same gpt-5.5 model correctly — ccusage codex daily shows a real dollar amount for the same model on my machine, while ccusage hermes daily shows $0.00.

Steps to reproduce

1. Run Hermes against the ChatGPT subscription endpoint (provider: openai-codex, base_url: https://chatgpt.com/backend-api/codex) so it writes estimated_cost_usd = 0.0 with billing_mode = subscription_included into ~/.hermes/state.db.
2. Accumulate non‑zero token usage on a model that exists in ccusage's pricing table, e.g. gpt-5.5.
3. ccusage hermes daily --offline. Every row shows $0.00.
4. Compare with ccusage codex daily for the same gpt-5.5 model — it shows a real cost.

Expected behavior

If a Hermes session's recorded cost is zero but it has non‑zero token usage, ccusage should fall back to pricing‑based calculation (the same path the codex adapter uses), so the user still sees a meaningful cost number. A recorded positive cost should still be trusted as is.

Version

20.0.1

I have a fix ready (very small change to calculate_hermes_cost + two regression tests; full crate cargo test passes 219/219). Happy to open the PR once approved.