The valley keeps rendering

The original native collector continued after the recovery life-controller stopped, without a reset. The final same-run snapshot preserves the earlier failed result and its intermediate telemetry. The remaining tail kept life paused and verified the saved world, call ledger and original player identity unchanged. Camera exploration, captures, local overview work and offline development tests remained part of the observed workload; this is not an isolated performance benchmark or a clean continuous-life acceptance. Measured snapshot: 1,800.12 wall seconds, 106,889 valid retained Update intervals, all-frame p95 of 17.575 ms, and thinking-labelled p95 of 25.386 ms. Source: private endurance-20261015-030718/render-only-tail/performance-final.json, snapshot 20261015T010718435Z-069aa2ee-0003-20261015T013718617Z-c5cb49, captured 15 October 2026 at 01:37:18.617 UTC. The loaded native source remained unity-source-975d21e7292f-20261015005815. Full distributions, focus/window conditions, thresholds and the separate failed trial counts are retained in docs/local-first-refinement.md. Thinking is a label from the latest resident session poll, not an exact measurement of GPU inference; connection checks and unrelated development work can occur outside that label. The report measures frame intervals rather than model latency. The approved sunset frame was captured during a paused interval within the same window, with its original date and native build preserved. The producing game session approved it individually, and the support session inspected its exact pixels and sidecar. A still image illustrates the rendered environment; it does not establish the measured distribution. Next: build the later target-grounding correction and use only the remaining authorized attempts for its separate follow-up. The old-build rendering result does not establish new-source model reliability.