fix(provider): preserve reasoning_content when merging consecutive assistant messages

[chengyongru]

Summary

• Fix DeepSeek thinking mode 400 error caused by _enforce_role_alternation() silently dropping reasoning_content when merging consecutive assistant messages
• Add _carry_over_reasoning() helper to preserve and concatenate reasoning_content across all 4 merge branches in _enforce_role_alternation()
• Strip assistant-only fields (reasoning_content, thinking_blocks, tool_calls) when recovering trailing assistant messages as user role

Root Cause

DeepSeek thinking mode requires reasoning_content to be passed back in all subsequent API requests when tool calls were involved. The _enforce_role_alternation() method merged consecutive assistant messages by concatenating content but silently discarded reasoning_content, triggering:

Error: The reasoning_content in the thinking mode must be passed back to the API.

This could occur after history truncation (_snip_history) or checkpoint restoration created consecutive assistant messages.

Changes

• nanobot/providers/base.py — New _carry_over_reasoning(src, dst) helper + integration into all merge branches + recovery path cleanup
• tests/providers/test_enforce_role_alternation.py — 8 new test cases covering reasoning_content preservation

Test plan

• All 8 new tests pass (reasoning carry-over in all merge scenarios)
• All 167 existing provider/agent tests pass with no regression
• Manual test with DeepSeek-R1/V4 thinking mode + tool calls in multi-turn conversation