reharness 0.5.0 — the repeated-request loop, per-task adoption, two figures

First public release. A self-improving agent harness: a Revision Session reads the agent's own returned run, revises the Harness Package (tools, hooks, skills, prompts) and proves the change with a Live Probe before Per-Task Adoption, under a Frozen Core.

• Results: Claude Opus 5.5 at 42.47% on the AutomationBench leaderboard; with reharness, 68% at steady state on the benchmark's held-out finance tasks; up to +13 points over the same model without it across sealed finals; no passing task lost.
• Feedback tiers: none / outcome / attempts / items — what the session may be told is a field of the Sealed Protocol.
• Two figures: components and architecture; how one improvement happens.
• Records: docs/RESULTS.md (every final and run, with costs), docs/JOURNAL.md, docs/DESIGN.md.