v3.15.14: RVF export on large brains, bounded learning tables, honest chaos results

What's New

• RVF brain export works on large learning databases. The brain checkpoint hook keeps the last good checkpoint and stops retrying a failing export every turn (#1238, #1244).
• Learning tables can stop growing without bound. Opt-in dream retention (AQE_DREAM_RETENTION=on), and aqe learning prune, which is a dry run unless --apply and backs up first (#1239, #1245).
• No overall quality score while a requested metric is unmeasured. quality_assess returns qualityScore: -1 with the reason, plus a separate subsetScore over the measured metrics (#1219, #1246).
• Chaos tools report faults that never ran as not executed. chaos_test returns status: "not-executed" and resilience: null instead of a made-up "completed, no data loss". Uninjectable faults report injected: false with what they need (#1258, #1259).
• Context compaction keeps its Tier-2 summary instead of dropping it, and the token budget matches what is kept (#1249, #1260).
• Security: zod 4.6.7 in the OpenCode plugin and fast-uri 3.1.9 (#1256).

Getting Started

npx agentic-qe init --auto

See CHANGELOG for full details.

🤖 Generated by Ruflo & AQE