[codex] Persist graph trace for context-only retrieval

[shuzhouzhang]

Summary

• add an opt-in persist_trace flag for context-only search and recall calls
• persist retrieved graph element IDs to the session cache without generating an LLM completion
• wire the flag through SDK/API/cloud-client paths and add focused unit coverage

Why

When only_context=True, Cognee skips completion generation. That also meant graph trace metadata was not saved, so later feedback/improve flows could not know which graph elements were used for the returned context.

Validation

• .\.venv\Scripts\python -m pytest cognee\tests\unit\modules\retrieval\graph_completion_retriever_test.py cognee\tests\unit\modules\search\test_get_retriever_output.py -q
• .\.venv\Scripts\python -m py_compile cognee\modules\retrieval\graph_completion_retriever.py cognee\modules\search\methods\get_retriever_output.py cognee\modules\search\methods\search.py cognee\api\v1\search\search.py cognee\api\v1\search\routers\get_search_router.py cognee\api\v1\recall\recall.py cognee\api\v1\recall\routers\get_recall_router.py cognee\api\v1\serve\cloud_client.py cognee\tests\unit\modules\retrieval\graph_completion_retriever_test.py cognee\tests\unit\modules\search\test_get_retriever_output.py
• git diff --check

Fixes #2910

[github-actions]

Hello @shuzhouzhang, thank you for submitting a PR! We will respond as soon as possible.