docs: add NPC state of mind server evidence

[JOY (JOY)]

Summary

• Add Expand State of Mind and BehaviorTendencies for NPC Frames #134 evidence pack for the existing Nakama NPC state-of-mind and behavior-tendency server slice.
• Document state fields, deterministic tendency fixtures, prompt/tone effects, malformed fallback behavior, and LLM/client write boundaries.
• Keep Unity read-only debug UI and final relationship-aware tendency tuning explicitly open.

Verification

• npm.cmd ci
• npm.cmd run build
• npm.cmd test
• git diff --check
• rg -n "Not run|TODO|TBD|—|–" docs/playtests/2026-05-30-npc-state-of-mind-server-evidence returned no matches
• npx.cmd markdownlint-cli2 docs/playtests/2026-05-30-npc-state-of-mind-server-evidence/*.md

Local review fallback

• Verdict: approved for PR.
• Docs-only evidence update, no runtime behavior changed.
• Checked authority boundary: evidence maps to existing Nakama tests and does not claim Unity debug UI completion.
• Checked Unity scope: no Unity scenes, prefabs, packages, materials, or ProjectSettings changed.

Refs #134

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1d0dea4415

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Remove unwired tendency fields from implemented effects

In the inspected backend and Unity code, idle_radius_meters is only normalized/serialized and conflict_response is only exposed as prompt/voice metadata, not used by local behavior planning. Claiming both as implemented planning inputs can let #134 be closed with behavior that is not actually wired; narrow this evidence to the fields that drive decisions, or add tests/runtime use for these fields.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Don't mark malformed state as bounded

For malformed stored heartbeat data, normalizeFrameHeartbeat only clamps numeric fields like stress/fear/confidence; string fields such as mood, current_goal, and last_plan_summary are accepted with trimString and can still flow into prompt or debug summaries. This result overstates the malformed-state fallback coverage and can let the #134 evidence pass even when invalid or oversized string state is not defaulted, rejected, or truncated.

Useful? React with 👍 / 👎.