Fix prune calibration with cache token usage

[romy651]

Summary

Fixes prune calibration when provider usage includes prompt-cache detail fields that are not directly comparable with the active context window.

The pruner now prefers the provider-reported usageMetadata.input_tokens for calibration and skips calibration for turns where the reported input is clearly not comparable, such as Bedrock cache-artifact turns with tiny input_tokens and very large cache creation/read detail fields.

Root Cause

Cache detail fields can represent provider billing/cache accounting rather than the actual prompt window size. Folding those details into calibration could inflate the observed input and poison the running calibration ratio or resolved instruction overhead.

Impact

This keeps pruning budget estimates stable when prompt caching is active, while preserving raw token maps and existing provider-specific cache semantics.

Validation

• npx tsc --noEmit
• npx eslint src/messages/prune.ts src/specs/token-accounting-pipeline.test.ts
• git diff --check HEAD^ HEAD
• npx jest src/specs/token-accounting-pipeline.test.ts
• npx jest src/specs/summarize-prune.test.ts src/specs/token-distribution-edge-case.test.ts src/specs/thinking-prune.test.ts

[danny-avila]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: 2d39f268e5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".