fix: improve NPC dialog retry and UI readability

[JOY (JOY)]

Summary

• Add bounded retry for direct player-to-NPC chat when DOS.AI returns transient timeout or 5xx fallback reasons.
• Keep deterministic fallback speech suppressed in Unity dialogue so only model-authored NPC lines appear as spoken dialogue.
• Improve prototype speech bubble and UGUI chat readability on high-resolution displays.
• Lower the default Nakama DOS.AI decision timeout from 120 seconds to 30 seconds, with env override still supported.
• Update ROADMAP, CHANGELOG, and Nakama README.

Local code review

Verdict: APPROVED WITH SUGGESTIONS.

Notes:

• No authority regression found. LLM output still cannot mutate authoritative game state.
• No client key exposure found. Unity still calls Nakama, not api.dos.ai directly.
• Suggestion for a later PR: add an explicit NPC response lifecycle event so the chat panel can show model-response completion or failure instead of only the initial society queue status.

Verification

• npm.cmd install in backend/nakama
• npm.cmd run build in backend/nakama
• npm.cmd test in backend/nakama
• dotnet build Unity\SecondSpawn.AI.csproj
• dotnet build Unity\SecondSpawn.UI.csproj
• dotnet build Unity\SecondSpawn.Networking.csproj via worktree compiler setup
• npx.cmd markdownlint-cli2 ROADMAP.md CHANGELOG.md backend/nakama/README.md
• git diff --check
• em dash scan on changed files

Review fallback

Codex Cloud and Gemini reviews are best-effort per repo policy. If cloud review quota or connector limits prevent them from running, this PR is ready for merge after green CI plus the local review fallback above.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c8787e421d

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]

Limit retry loop to player-originated chat stimuli

The new retry loop runs for every pending speech stimulus, but this coroutine is also used for NPC-to-NPC speech (NotifyNearbyNpcSpeech sets fromNpc=true before starting PlayerChatResponseLoop). Because the retry path here does not check _pendingMessageFromNpc, transient DOS.AI failures now cause up to three model calls per NPC-heard line, which can amplify load and consume model budget during outages even though the feature is scoped to direct player chat. Restrict retries to player-originated messages before entering the retry attempts.

Useful? React with 👍 / 👎.