rrsi: preserve default checkpoint evidence and prevent prose-only structural novelty #337

Follow-up to merged #333 and the existing two-P2 review, related to umbrella #221. Both are limited to the opt-in RRSI example bundle, not core default acceptance.

Evidence boundary: source-confirmed control/data flow at main a7fbb7ccd3825aeb718888db65b3e405a5735243; proposed reproductions and regression tests below have NOT been executed locally. No claim of an observed production incident. Current local project materialization/testing remains paused. These merges are newer than release v0.6.0.

Two bounded failure paths

1. checkpoint/restore omit history even when history_path=None uses the in-memory backend. A fresh resumed bundle restores S*/round/progress but loses accepted component counts, pruning and proposer evidence.
2. upstream_components feeds raw key/new text to normalize, while text_only expects +/- changed-line markers. A SingleSlot prompt such as “Do not mention ToolRegistry in the final answer” can be classified client_tool and receive structural novelty despite no machinery being added.

Proposed tests / acceptance

• Round-trip default in-memory state with accepted/rejected records; preserve records/counts/pruning, and match uninterrupted next-candidate acceptance. A previously accepted skill must not become novel after resume.
• Generate the prompt example via SingleSlot.to_diff; require prompt classification and no novelty-based acceptance at equal score/cost. Cover comment/string-only edits too.
• Preserve explicit durable-history resume and the pinned selection/governance rules; represent actual edit semantics instead of keyword-only replacement text.

Dependencies/scope: #333, umbrella #221. Limit changes to RRSI adapter/history integration and regression coverage; do not weaken core gates or reopen unrelated port work.