feat: Claude enhancements - prompt caching, batch API, and translation improvements

[seidnerj]

Summary

Advanced translation features specifically for the Claude (Anthropic) engine, including cost-saving optimizations and reliability improvements.

Translation Refusal Detection & Retry

• When prompt caching provides full book context, Claude may refuse to translate content it identifies as copyrighted
• Enhanced system prompt with anti-refusal language for both standard and cached modes
• Two-stage refusal detection: fast heuristic pre-filter (pattern matching) followed by LLM classification to avoid false positives
• Automatic retry (up to 3 attempts) when refusal is detected in translate_paragraph
• Handles edge cases like translating copyright-related text to English without false positives

Prompt Caching (90% Cost Reduction)

• Enable parallel section translation with full book context
• Cache full book text as system message
• 90% input token cost reduction on subsequent sections
• Works with existing merge_length for section sizing
• UI checkbox in Claude settings (default: disabled)

Batch Translation API (50% Cost Reduction)

• New asynchronous bulk translation engine: ClaudeBatchTranslate
• Process up to 100,000 messages per batch
• 50% cost reduction compared to standard API
• Combined savings: Up to 84% total cost with prompt caching
• Trade-off: Processing time up to 24 hours (most complete <1 hour)
• Automatic polling with progress logging

Dynamic Token Management

• Automatic max_tokens scaling based on input length
• Model-specific output limits (4K-128K depending on model)
• Pre-translation warning if merge length exceeds model capacity
• Prevents partial/truncated translations
• Language-aware token estimation helper

Extended Output/Context (Beta Features)

• Extended output: 128K tokens (Claude 3.7 Sonnet)
• Extended context: 1M token window (Claude Sonnet 4.0/4.5)
• Conditional UI visibility based on selected model
• UI checkboxes for opt-in beta features

Dynamic Timeout (Opt-in)

• Content-based timeout scaling for large translations
• Example: 50K chars → 6.5min, 200K chars → 23min
• Prevents timeout failures on very large merged sections
• UI checkbox (default: disabled)

Streaming Improvements

• Immediate stop button response during streaming
• Preserve scroll position when adding streamed text
• Prevent text corruption on clicks during streaming
• Qt5/Qt6 compatibility

Testing

• Prompt caching with full context
• Batch API submission and retrieval
• Dynamic max_tokens prevents truncation
• Extended output/context settings
• Streaming stop button works immediately
• Timeout scaling handles large content
• Refusal detection with copyrighted ebook content
• Refusal retry loop completes successfully
• No false positives when translating copyright-related text

Version

Bumped to v2.4.2 with comprehensive CHANGELOG