fix(llm): LiteLLM proxy responses record $0.00 cost when x-litellm-response-cost header is omitted #51958

Bug

When using OpenCode with a LiteLLM proxy (e.g. litellm/gpt-6-luna), some responses are recorded with $0.00 cost even though the proxy correctly billed the request.

Root cause

LiteLLM 1.98+ omits the x-litellm-response-cost response header when its value equals x-litellm-response-cost-original (a deduplication optimisation). However, OpenCode does not read any LiteLLM response headers at all — cost is estimated purely from token counts and a local price table.

For models without a price entry in the local config, this produces $0.00. For models with a price entry but where the header is absent, the local estimate may differ from the proxy's actual billing.

Steps to reproduce

1. Configure OpenCode with a LiteLLM provider (e.g. a self-hosted proxy or prophet-router)
2. Send two identical consecutive requests with the same model
3. Check opencode stats — one or both may show $0.00 for that model

Observed vs expected

• Observed: opencode stats shows $0.00 for LiteLLM-routed models
• Expected: cost matches x-litellm-response-cost (or x-litellm-response-cost-original) from the proxy response headers

Evidence

Two manual curl calls to the proxy with identical payloads:

• Request 1: x-litellm-response-cost: 0.000434 present
• Request 2: x-litellm-response-cost absent, x-litellm-response-cost-original: 0.000434 present
• Both billed: x-litellm-key-spend incremented on both