fix: don't subtract cached tokens from input_tokens in OpenAI→Claude translation

[skrabe]

Summary

extractOpenAIUsage in the OpenAI→Claude response translator subtracts cached_tokens from prompt_tokens before reporting input_tokens to downstream clients. This causes clients that rely on input_tokens for context window tracking (Claude Code, Factory Droid, etc.) to see near-zero values on cache hits, breaking compaction triggers.

Before: prompt_tokens=150000, cached_tokens=149900 → input_tokens=100
After: prompt_tokens=150000, cached_tokens=149900 → input_tokens=150000

The cache_read_input_tokens field is still set separately for clients that need the breakdown.

Context

• Related to token under-reporting when translating from Anthropic to OpenAI CLIProxyAPI#2281 (cache token fix for Chat Completions path)
• This fix addresses the same issue in the OpenAI→Anthropic translation path (internal/translator/openai/claude/openai_claude_response.go)
• Affects BYOK setups where Claude Code talks to non-Claude upstreams through the proxy

Test plan

• Verify input_tokens in Anthropic-format responses includes cached tokens
• Verify cache_read_input_tokens is still reported correctly
• Confirm Claude Code compaction triggers at expected thresholds with BYOK

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the .gitignore file, simplifies token usage extraction in the Claude translator by removing manual subtraction of cached tokens, and introduces an update-and-build.sh script for automated deployment. Feedback was provided to enhance the automation script by dynamically detecting the current branch and ensuring the installation directory exists before copying the binary.

[gemini-code-assist]

The branch name is hardcoded to a specific feature branch (fix/byok-compaction-cached-tokens). This makes the script unusable for future updates or other branches. It is better to dynamically detect the current branch so the script remains functional across different tasks.

Suggested change

	FIX_BRANCH="fix/byok-compaction-cached-tokens"
	FIX_BRANCH=$(git rev-parse --abbrev-ref HEAD)

[gemini-code-assist]

The script assumes that the destination directory ~/.local/bin/ already exists. If it does not, the cp command will fail. It is safer to ensure the directory exists before attempting to install the binary.

Suggested change

	cp "$BINARY_NAME" "$INSTALL_PATH"
	mkdir -p "$(dirname "$INSTALL_PATH")"
	cp "$BINARY_NAME" "$INSTALL_PATH"