NousResearch#22660)

Two methods in MemoryStore called .fetchone()["field"] without first
checking if fetchone() returned None, causing TypeError in race conditions.

- add_fact: after sqlite3.IntegrityError, the duplicate row may have been
deleted by a concurrent operation before the fallback SELECT executes.
  → Guard with explicit None check; raise RuntimeError for this
    impossible-but-observed state.

- update_fact: when category is not provided, a second SELECT fetches the
current category. If the row was deleted between the first SELECT and this
second one, fetchone() returns None.
  → Guard with explicit None check; return False (row no longer exists).

Fixes NousResearch#22660

…search#22569)

cron.jobs.load_jobs called data.get("jobs", []) immediately after
json.load(), assuming the root is always a dict. If jobs.json is
hand-edited or written by a legacy release in a non-dict shape
(bare list, null, scalar, string), the call crashes with AttributeError.

Changes:
- Guard both json.load() and json.loads(strict=False) paths with
  isinstance(data, dict) checks.
- If data is a bare list (legacy format), auto-migrate by calling
  save_jobs() and return the list.
- If data is any other non-dict shape, log an error and return []
  instead of crashing.

Fixes NousResearch#22569

…ch#22489)

The generic API-key health-check loop sends
to every provider, but Google AI Studio (generativelanguage.googleapis.com)
rejects Bearer tokens and requires  instead. This made
 falsely report Gemini as unreachable or invalid.

Changes:
- In , detect Gemini by checking whether the
  resolved base URL host matches .
- For Gemini endpoints, set  instead of .
- All other providers continue using  unchanged.

Fixes NousResearch#22489