fix(executor): honor explicit-zero Claude stream usage fields

[rome-xi]

Problem

Claude stream usage is accumulated across message_start and later message_delta events. Parsing used gjson.Int(), which maps both a missing field and an explicit 0 to zero. The shared stream merge then kept any prior positive value whenever the update was 0.

A full cache hit therefore billed incorrectly: message_start can report positive input_tokens / cache-creation, then the final message_delta reports input_tokens: 0 (and often cache_creation_input_tokens: 0) with the real cache_read_input_tokens and output_tokens. The start input was kept, so uncached input and cache-hit rate were wrong.

Omitted-only deltas such as {"output_tokens": 15} must still preserve start input and cache fields.

Fix

Claude metering now tracks field presence (Exists() && Type != Null) and merges on that basis:

• field present, including explicit 0 → overwrite
• field missing or JSON null → keep prior

This is scoped to the Claude executor/plugin observe path (ParseClaudeStreamUsage / parseClaudeUsageNode / ObserveClaudeStream / parseClaudeStreamLine). MergeStreamUsageDetail is unchanged for OpenAI and other protocols that still keep prior values on zero.

Tests

• full cache hit: positive start input/cache, then message_delta with explicit zeros → final detail has input=0, reported cache read, reported output
• missing input_tokens preserves start
• input_tokens: null preserves start
• later SSE line with no usage preserves prior
• late positive input_tokens replaces start
• plugin Claude stream path has the same explicit-zero overwrite

go test ./internal/runtime/executor/helps/ -count=1

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-14T13:45:28.568297Z	80cda53	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.