fix: agent-framework drops profile search memories

[noQbot]

What & why

In the Agent Framework integration, middleware.py and context_provider.py both pass list(response.search_results.results) into deduplicate_memories(), which routes each item through extract_memory_text():

def extract_memory_text(item):
    if isinstance(item, dict):
        memory = item.get("memory")
        ...
    if isinstance(item, str):
        ...
    return None   # <- SDK models land here

But the Supermemory SDK returns response.search_results.results as Pydantic model objects, not dicts or strings. A model matches neither branch, so extract_memory_text returns None and the memory is silently dropped — in query/full mode, no search memories get injected into the context. Unlike the Cartesia/Pipecat path this doesn't crash; it just quietly loses retrieved context, which is harder to notice.

Why existing tests didn't catch it: the current fixtures pass dicts ({"memory": "..."}), a shape production never sends. (The sibling openai-sdk-python package already model_dump()s the same results, confirming the model shape.)

The fix

Add a model branch to extract_memory_text — read "memory" from model_dump(by_alias=True), falling back to attribute access — while leaving the existing dict and str handling untouched:

if isinstance(item, str):
    ...
if isinstance(item, dict):
    memory = item.get("memory")
else:
    model_dump = getattr(item, "model_dump", None)
    if callable(model_dump):
        memory = model_dump(by_alias=True).get("memory")
    else:
        memory = getattr(item, "memory", None)

Tests

Added tests/test_search_result_models.py feeding dict-less model stand-ins through the dedup path — asserts models are extracted and deduped, the attribute fallback works for models without model_dump, and dict inputs still work.

Note: the repo's PR CI runs only TypeScript type-checks + Biome, not the Python suites, so these tests won't run in CI here — I verified them locally.

Compatibility

Backward compatible — dict and str inputs behave exactly as before. No API or behavior change for callers already passing dicts.