Flare state diagnostics evidence

This public-safe evidence set closes the 2026-10-23 Flare state diagnostics follow-up. It retains interrupted attempts, failures, repairs, recovery, complete protected A/B/B/A observations, native state diagnostics, top-20 observations, and paired scheduler traces.

The decision is no promotion. The trace review accepts complete scheduler evidence only as a bounded diagnostic: A/B graph-shape mixtures differ and HTTP requests cannot be mapped to scheduler slots or generations. No causal bottleneck, numerical repair, protected-performance qualification, tuning choice, or promotion is proven.

Start with summary.json, report.md, coverage-and-gaps.md, and the final artifact manifest. The source manifest and provenance map public files to sanitised source evidence.

Raw evidence is archived by campaign family. The failed native executable is deliberately excluded; its provenance and exclusion reason are retained.