Two unrelated test-quality fixes batched. 1. Remove the "ollama must never come back" guards. Per founder: if we ever reintroduce ollama it will be a deliberate decision, so a test that fails on its return is wrong -- it encodes a point-in-time policy, not an invariant of every valid state (the same principle that retired the changelog test). Deleted tests/test_docs.py entirely (both tests were ollama-absence guards) and dropped the two `assert "ollama" not in PROVIDER_DEFAULTS/REGISTRY` lines from test_local.py, keeping its positive local-provider invariants. 2. Fix the test_voice_brain xdist flake (root-caused, not silenced). test_schedule_refresh_updates_instruction asserted on app.voice_brain_instruction immediately after future.result(). That attribute is set by the future's done-callback, but a concurrent.futures.Future notifies result() waiters BEFORE invoking done-callbacks -- so under xdist CPU starvation the runtime thread can lag and the assert reads the stale value. Reliable solo, flaky under load (the "1 failed, passes on rerun" we saw). Proven deterministically by lagging the callback 0.2s: result() returns ('session-3','New voice') while the immediate read is still '' (old assert FAILS), and a poll reads 'New voice' (new assert PASSES). Fix: poll for the side effect with a 1s deadline instead of racing it. make ci green (8255 passed). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>