P6 — Agent analysis tooling & ops polish

[foae]

Goal: longitudinal analysis is possible for an agent on request; ops loose ends closed. Spec: design/plays-plan.md §8, design/plays-product.md §5.

• design/plays-analysis.md: schema map, canonical SQL (category cohorts, mark trajectories, resolution rates), mandatory selection-bias caveats (invariant P5) an analyzing agent must repeat
• plays-export CLI (--from/--to/--format json|csv → data/exports/)
• Reprocess path: separate queue mode that never touches status (row stays published; pointers advance on commit); idempotent per (play, stage, version); explicit budget-override flag
• deploy/v2/README.md + runbook: budget knob, queue-depth, media sizing (~500 MB/month) + retention/prune knob, healthcheck notes for the plays loops
• Media retention default confirmed: keep forever, prune knob available

Gate: an agent can run a correct longitudinal analysis from the export + how-to doc alone, caveats included.