feat: v2.5.0 - models, tools, pricing from open-sourced Claude Code

[ttlequals0]

Summary

Updates the wrapper with authoritative data extracted from Anthropic's open-sourced Claude Code CLI source:

• Model accuracy: Added claude-sonnet-4-6, re-enabled 3.x models, per-model context windows/output limits/pricing
• Tool registry: Expanded from 15 to 33 tools matching Claude Code's actual inventory (Agent, TaskCreate, EnterPlanMode, worktree tools, cron tools, etc.)
• Retry logic: Exponential backoff with jitter, automatic Opus-to-Sonnet fallback after 3 consecutive 529s
• Cost tracking: Per-request and per-session cost estimation using authoritative pricing, with auto-cleanup
• Parameter support: X-Claude-Effort, X-Claude-Thinking headers, model-specific max_tokens validation
• Code quality: Extracted shared options-building helper, deduplicated constants, asyncio.Lock
• Docs: README rewritten (774 -> 280 lines), migration docs marked historical

Version

2.5.0

Docker

• ttlequals0/claude-code-openai-wrapper:2.5.0
• ttlequals0/claude-code-openai-wrapper:latest

Test plan

• All 566 tests pass (35 new for retry + cost tracker)
• Docker image built for linux/amd64 and pushed
• Model/tool/pricing consistency verified programmatically
• Security review: 0 findings above threshold
• pyproject.toml and src/init.py versions match