Fix Responses-API cache token double-count in context occupancy

[TheGreatAxios]

Summary

• Responses-API adapters (codex, grok, openai-responses) split cached_tokens out of input_tokens so TokenUsage fields stay non-overlapping and context occupancy sums equal real prompt tokens
• The context meter and compaction governor no longer read up to ~2x inflated occupancy on high cache-hit sessions
• Faremeter bills cached tokens at the cache rate exactly once for Responses-API usage

Why: on the OpenAI/xAI Responses API, input_tokens_details.cached_tokens is a subset of input_tokens. Summing the raw fields double-counts every cached token, so the meter reads 70-80% while real occupancy is ~40% — and compaction either fires in early bursts or appears stuck.

Verification

• bun run typecheck, bun run build, and bun run test pass
• Adapter tests pin input + cacheRead == wire prompt_tokens, the omitted-details edge, and a clamp for inconsistent gateways reporting cached_tokens > input_tokens
• New faremeter tests pin cost math against normalized usage: uncached input at the input rate, cached reads at the cache rate, each token billed once

Fixes CL-7075

[linear-code]

CL-7075

[TheGreatAxios]

Review · Approve

Responses-API usage parsing splits cached_tokens out of input so occupancy and faremeter sums count each prompt token once.

No findings.

GitHub rejected --approve on this PR (author cannot approve their own pull request); verdict is recorded here as a comment.