docs: expand role-play provider evaluation

[JOY (JOY)]

Summary

• Expand the role-play provider evaluation design with technical patterns worth studying from Alibaba-style role-play APIs.
• Clarify character-conditioned request shape, provider-session boundaries, adapter-local memory, and server-owned game state rules.
• Add benchmark cases for session behavior and character-control boundaries.

Verification

• npx.cmd --yes markdownlint-cli2 docs/design/52-llm-role-play-provider-evaluation.md: pass, 0 errors.
• git diff --check: pass.
• rg -n "—|–" docs/design/52-llm-role-play-provider-evaluation.md: pass, no generated em dash/en dash.

Local reviewer fallback

• Docs-only change; no Unity, Nakama, AGENTS.md, or .claude/CLAUDE.md changes.
• Reinforces that provider-managed memory is never canonical and provider output cannot mutate movement, combat, quests, inventory, TIME, SECOND, or relationship state directly.
• Relevant to Benchmark role-play model providers behind api.dos.ai #249.