advisor tool goes from confirmed-live to unavailable mid-session, no warning

Summary: The advisor tool (backed by Fable 5) intermittently loses its binding partway through a session. Reproduced twice on 2026-07-01:

• Session 1: /advisor reported "Advisor set to Fable 5," but every subsequent advisor() call errored as unavailable for the entire session — the binding never came up at all.
• Session 2 (same day): /advisor again reported success, and the first advisor() call worked correctly (returned a real critique). The very next advisor() call a few tool-turns later — same session, no /advisor re-invocation in between — errored "advisor is unavailable" and instructed not to retry.

Why this looks like a binding/session issue, not model instability: it fails as a hard tool-unavailable error, not a degraded or wrong answer, and the failure boundary is time/turns-in-session rather than anything about the query itself.

Impact: breaks workflows that rely on advisor() holding for a whole task (e.g. iterative draft-critique loops), since there's no way to detect the drop except by it happening.

Environment: Claude Code CLI, Windows 11, session-based advisor tool with Fable 5 backing.