fix: preserve reasoning in short-term memory replay

[MackDing]

Summary

• replay assistant short-term memory with reasoning parts when available
• keep plain assistant text replay unchanged when no reasoning exists
• test the actual replay helper rather than a duplicated test-only implementation

Why

Issue #24 reports provider failure: The reasoning_content in the thinking mode must be passed back to the API. The agent already stores result.reasoning, but replaying short-term memory previously sent only assistant text back to the model. Providers that require prior thinking content can reject the next turn when that context is dropped.

Relationship to #29

#29 enables thinking for DeepSeek models routed through OpenAI-compatible providers. This PR is complementary: it replays captured reasoning on later turns. The two PRs cover different stages and are not duplicates.

Validation

• npm run typecheck
• npm test (69 tests passed)
• npm run build

Closes #24

[MackDing]

Quick follow-up: this PR appears mergeable from my side, and I do not see a failing required check. Happy to make any maintainer-preferred tweak if needed.