[KunalSiyag]

Summary

Fixes "chat with AI not working". In a city nobody lives in yet the chat stayed silent: chatter rounds only ran with residents, and a player line had nobody to answer it. Now the rival mayors answer and argue there, and when no reply can come the chat says so instead of going quiet.

What changes in the game

• Empty city, live replies working: the rival mayors answer each player line (the mayor the line names first, else the player's district's mayor, then a rival, up to 3 lines) and argue among themselves once a minute (3-5 lines from 2-3 mayors). Their lines skip the per-author limit, so a mayor can answer back within a round.
• No reply possible (offline, or the live call failed): a status card says "The mayors are busy fighting right now. Chat later."
• Live replies down with nobody living there: the card says "Live replies are slow right now." without promising neighbours who do not exist.
• Face to face: a reply still out after 8 s of feed time gets the resident's busy line ("can't think straight with the mayors fighting") instead of a made-up offline answer, and the late reply still shows when it arrives. Busy lines are not turns, quotes or mood changes. Up to two talk calls may be out at once, and the free-text box keeps what was typed while a reply is pending.

Implementation

• scripts/ai/chatter_minds.gd: run_mayors_round, _mayors_answer (used by reply_to_player when no residents can answer), mayor_panel, _mayors_live; validate_lines accepts extra mayors; talk() caps parallel calls at 2 and falls back to the LRU, then talk_busy (source busy).
• scripts/ai/gmi_client.gd: TEMPERATURE and a public request_body(). A model that refuses 0.7 with an HTTP 400 naming the temperature (Kimi-K3 takes only 1) is retried once at 1 and keeps 1 for the session (temperature_of). Every failed live call prints one scrubbed GMI_FAIL role/agent on model: error line (never the prompt; a repeat of the last line is skipped), so an editor run shows why replies fell back.
• scripts/fp/chat/chat_feature.gd: rounds run with residents or while live replies work; _run_round(mayors_only); the busy and empty-city status cards.
• scripts/fp/chat/talk_dialog.gd: the 8 s busy line, late replies accepted, _answered gates the end-of-conversation effects.

Tests

• fp_chat_live_test (new, 27 checks): the empty city (mayors answer, the named mayor first, rounds once a minute, failures and offline give the busy card), then a grown city with a slow talk reply (busy line at 8 s, late reply shown, the parallel cap, the conversation remembered).
• fp_ops_test: the temperature fallback and the request body.
• fp_chat_test unchanged and passing.

Docs

docs/fp/slices/C-chat.md (§7, the budget table, tests, risks), docs/fp/ARCHITECTURE.md (C row), README (GMI section), docs/runbooks/gmi-outage.md (the new symptoms and the temperature 400), regenerated API docs.

Validation

• GRIDFALL_OFFLINE=1 GATE_RETRIES=1 tools/gate.sh full: PASS (42 steps: import, parse, 37 unit tests, the orbit and first-person smokes), no retries.
• pre-commit run --all-files and python tools/checks/run_all.py: PASS, no new findings.

[github-actions]

Automated review

No new findings on the 13 changed files (gdlint, complexity, TODO policy, dead code).

Run python3 tools/checks/run_all.py locally for the full set of checks.