feat: implement dynamic model mapping for Claude background models via configuration overrides

[fajarhide]

PR Auto Describe

🚀 Summary

Adds configurable per-model mapping for Claude Haiku / Sonnet / Opus requests. Allows users to route Claude Code's hardcoded background calls to cheaper / supported upstream models instead of failing on unrecognised model names. Fully backwards compatible, no changes required for existing deployments.

---

🔑 Key Changes

✅ 3 new optional environment variables for model alias mapping
✅ Graceful fallback logic preserves existing behaviour
✅ All supporting docs, config, docker and tests updated
✅ Zero breaking changes for existing users

---

📋 Detailed Breakdown

Build & Ignore

• Added compiled binary outputs server, heimsense, heimsense_* to .gitignore

Documentation & Config Templates

• Updated README environment variable table, clarified default model description
• Added commented example mappings with explanation to env.example
• Added model mapping template to auto-generated .env on first run
• Exposed new env vars to container runtime in docker-compose.yaml

Core Implementation

• Added ModelMapHaiku, ModelMapSonnet, ModelMapOpus fields to root Config struct
• Implemented normalizeModel() resolution function with strict precedence:
  1. Force model override always takes priority
  2. Match requested model name against haiku/sonnet/opus substrings
  3. Fallback to default custom model if mapping not configured
  4. Final fallback: pass original requested model unmodified
• Updated ToOpenAIRequest() signature to accept config and return resolution errors
• Updated all existing unit tests for modified function interface
• Added proper error handling in request handler, returns clean 400 error for model resolution failures

---

🧠 Notes

Resolves the very common pain point where Claude Code hardcodes Haiku for background operations, causing 400 errors or unexpected costs for users running against non-Anthropic upstream providers.

---

⚠️ Breaking Changes

None. All new functionality is optional. All existing configurations and behaviour remain completely unchanged.