Preserve DeepSeek V4 thinking history

[Ingwannu]

Summary

• add DeepSeek V4 thinking models to the built-in DeepSeek provider preset
• route deepseek-v4-pro / deepseek-v4-flash with preserveReasoningContentModels so assistant reasoning_content is replayed in multi-turn/tool-call history
• map Codex xhigh to DeepSeek max while clamping lower stale requests to DeepSeek high
• document the model-scoped decision in ADR 0003

Root Cause

DeepSeek V4 thinking mode expects prior assistant reasoning_content to be passed back in later requests. The openai-chat adapter already supported model-scoped reasoning history replay, but the built-in DeepSeek preset did not mark the V4 thinking models for that behavior, so tool/subagent continuation turns could drop the field and trigger DeepSeek 400s.

Notes

This deliberately does not enable reasoning-content replay globally, and it does not apply the V4 rule to the legacy deepseek-reasoner preset.

Addresses #58.

References:

• https://api-docs.deepseek.com/guides/thinking_mode
• https://api-docs.deepseek.com/guides/reasoning_model

Validation

• bun run typecheck
• bun test tests/reasoning-effort.test.ts tests/provider-registry-parity.test.ts tests/router.test.ts
• bun test ./tests/