The full `just test-all` first run surfaced state leakage from prior
suites (tasks tests) into SV-02/SV-03: another spec had left the
collective row's feature_flags populated, so by the time the SV suite
ran the realtime test was hitting a noisy baseline.
Add a symmetric beforeEach + afterEach that resets BOTH user.feature_flags
(Ana + Borja, each via their own context to satisfy RLS) and
collectives.feature_flags (seed collective, via Ana). Cost: ~6s per
spec in setup overhead; benefit: 5 sequential `just test-all` runs
green.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>