2026-03: Replay hardening, storage lifecycle, and public mock workbench

What changed

• documented a hot / warm / cold storage strategy
• added storage envelopes and schema-aware persistence
• added bootstrap cold-start fallback and a seed bootstrap script
• hardened replay evaluation with max-hold exits, gap markers, and non-tradable hit-rate exclusions
• added merge conflict visibility for replay frames
• added stale DuckDB lock cleanup
• added a public mock replay workbench to the docs site

Why it matters

The system now explains and demonstrates how collected data survives long enough to support replay and review, instead of acting like disposable panel state.

Boundary

• the public workbench uses mock data
• cold archival still depends on deployment environment wiring
• online retraining and centralized provider quota control remain follow-up work