aiohttp is only declared under the optional `async` extra, not the dev
dependency group, so the black-box repros failed with ModuleNotFoundError
under the default `uv sync --dev` setup before exercising anything.
- Async repro now uses pytest.importorskip("aiohttp"), skipping cleanly
when the extra is absent (it exercises the aiohttp code path).
- Sync repro forces the requests fallback by making `import aiohttp` fail
via sys.modules, so it runs in the default dev setup and actually covers
the requests path production uses when aiohttp is not installed.
Verified both ways: with aiohttp -> 35 passed, 11 skipped; without aiohttp
-> 34 passed, 12 skipped (async repro skips, sync repro passes via requests).
SupermemoryOpenAIWrapper wraps client.chat.completions.create(). With
default configuration, every call triggers a network request to
api.supermemory.ai via add_system_prompt() -> supermemory_profile_search().
That call chain had no error handling, so any Supermemory outage,
timeout, or network hiccup crashed the entire chat completion, even
though the underlying LLM call was completely unaffected.
The memory-write path (add_memory_tool) already failed open correctly
in this same file; the memory-search path (which runs on every call
by default) had no equivalent protection. The original test for this
scenario even documented the intended behavior in its own comment
('Should not raise exception, should fall back gracefully') while its
assertion checked the opposite.
Both async and sync paths now catch SupermemoryNetworkError and
SupermemoryAPIError (warn-level log) plus any other exception
(error-level log), and fall through to the original unmodified
messages instead of propagating the failure.
11 new tests covering both paths: expected Supermemory errors, an
unexpected exception, the happy-path regression (memory injection
still works normally), and black-box repros matching the filed issue.
35 passed, 11 skipped (pre-existing, gated on a live API key).
