fix(proxy): preserve Responses context accounting

[LPFchan]

Summary / 概要

Preserve OpenAI Responses input/cache usage in the translated Anthropic stream so Claude Code can track context growth and auto-compact before the provider limit. Surface terminal context overflow as one invalid_request_error instead of silently completing the failed stream or allowing the original error to be masked by a non-streaming retry.

Related Issues / 关联 Issue

Closes #10.

Related to #5 and #6, which address the separate HTTP request-body limit encountered during large sessions.

Root Cause

The Responses translator emitted input_tokens: 0 in message_start, retained only upstream output usage, and sent only output_tokens in the final message_delta. Claude Code therefore recorded zero input usage throughout translated sessions and could not reliably reach its proactive auto-compaction threshold.

The translator also ignored response.failed, then emitted a normal end_turn. A backend context_length_exceeded event was consequently unavailable to Claude Code's reactive compaction path. In error-aware downstream variants, treating the event as a generic API failure caused Claude Code to try a non-streaming request, which ChatGPT's Codex endpoint rejects with Stream must be set to true, obscuring the overflow.

Changes

• Propagate terminal Responses input, cache-read, cache-creation, and output usage through message_delta.usage.
• Split cached input from the Responses total with defensive clamping, avoiding double-counting in Anthropic's additive usage model.
• Translate context_length_exceeded to invalid_request_error; preserve unrelated failures as api_error.
• End failed streams after the error event without fabricating block closure, usage deltas, or message_stop.
• Add translated-SSE regression tests for ordinary usage, cache accounting, malformed fields, context overflow, and unrelated errors.

The Anthropic TypeScript SDK explicitly merges late input/cache fields from message_delta.usage, so this preserves immediate streaming and requires no buffering or token estimator.

Type of Change / 变更类型

• Bug fix / 缺陷修复
• New feature / 新功能
• New provider support / 新增提供商
• Breaking change / 破坏性变更
• Documentation / 文档
• Refactor / 重构

Test Plan / 测试计划

• cargo fmt --check
• cargo clippy --all-targets — change introduces no warnings; 12 pre-existing warnings remain on upstream main
• cargo test responses_stream — 8 passed
• cargo test — 346 unit tests and 1 integration test passed
• git diff --check

Checklist / 检查清单

• cargo fmt --check passes
• cargo clippy — zero warnings (12 pre-existing upstream warnings remain)
• cargo test — all tests pass
• Documentation updated (not applicable: internal wire-compatibility fix with no new configuration or public contract)
• No sensitive data (API keys, tokens) committed