1. When. Monday 20 April 2026. Monthly share went 2.46% (Mar) → 5.93% (Apr) → 8.03% (May) and has held at 6.9–8.0% since. At day level the break is clean and permanent on 20 April, with a smaller precursor spike on 18 April.

2. Where. Not a hostname problem. 14,785 of 14,873 Unassigned sessions (99.2%) are on www.scape.com. The volume is concentrated in (not set) / (not set) source/medium — 12,443 sessions, 84% of the segment — which is exactly what a session with no captured traffic source looks like. Monthly, that bucket goes 1,040 (Mar) → 3,281 (Apr) → 3,937 (May) → 4,326 (Jul).

3. Sessions or events. Both, and they are separable. Sessions are real (session_start present, landing pages present) but arrive with no source. Separately, 83,384 stellar_experiment_conversion events arrive with hostname (not set) and zero sessions attached — a server-side / Measurement Protocol style stream with no page context. Its start date matches the session step exactly.

Working hypothesis: an A/B testing or experimentation layer ("stellar") went live around 19–20 April 2026 and fires GA4 hits before the consent/tagging layer has established a traffic source — creating sessions with (not set) source and a large volume of context-free events. The secondary contributor is listing-partner and AI-assistant traffic (listingsite / web, chatgpt.com / (not set), livechat.com / chat) using mediums that match no channel grouping rule; that is genuine traffic and can be fixed with custom channel groups.

Next two checks (outside this build): GTM container GTM-N2RPH78D versions published 17–21 April 2026, and any CMP/consent banner release in the same window. Whichever one lands on 19–20 April is the culprit.