feat: add first visible LLM invasion

[christopherjnelson]

What changed

Server-owned simulation

• adds the deterministic 61-cell Toledo development world with six fixed named agents
• runs one agent per turn in stable round-robin order with Start, Pause, Single Turn, and Reset controls
• keeps a monotonic completed-turn count independent of the newest 120 retained turn records
• retains only the newest 120 world events while preserving the newest eight relevant observation events in chronological order
• validates complete accepted or rejected records before atomically committing world state, turn history, count, and cursor
• records sanitized provider failures while allowing unexpected engine, schema, and internal failures to propagate without partial state changes

Visible World Lab

• renders all 61 H3 cells through MapLibre with visible open, infected, border, and selected-cell styles
• keeps all six agent markers visible, clickable, and above the H3 overlay
• verifies readiness from unique MapLibre-rendered features instead of React source counts
• exposes rendered cell and infection diagnostics for deterministic Playwright assertions
• updates infection and reset data without recreating the map
• configures a same-origin MapLibre worker path so GeoJSON processing works under Next.js

Real OpenRouter provider

• targets google/gemini-3.7-flash with no fallback model
• makes exactly one provider request per agent turn
• uses max_tokens: 1024, low reasoning effort, and exclude: true
• preserves provider.require_parameters: true and stream: false
• uses a strict root-object response schema with a nested anyOf action union and single-value discriminator enums
• leaves H3 formatting, summary bounds, action validation, adjacency, infection state, and consequences authoritative in Zod and the world engine
• keeps the abort timeout active through body reading, decoding, extraction, and validation
• reads non-success bodies only to a fixed bound and retains only sanitized status, code, message, request ID, and selected-model diagnostics for the opt-in CLI
• keeps those diagnostics out of turn records, snapshots, browser responses, and normal logs

Environment and deterministic seams

• loads provider configuration from the repository-root .env for both normal Game API startup and pnpm smoke:openrouter
• makes .env provider values authoritative over stale exported values while loading only the provider variables in the shared override helper
• keeps .env.example as the committed template and never exposes the API key to browser code
• keeps the scripted provider explicit and limited to deterministic tests and Playwright; it is never an OpenRouter fallback

Why

This is Roadmap PR 2: the first browser-observable LLM invasion. It delivers a genuine model-backed vertical slice while keeping world authority deterministic and deferring persistence, autonomous scheduling, player systems, and later-roadmap social mechanics.

The correction passes also addressed retained-history turn-number reuse, non-atomic world mutation, unbounded world events, response-body timeout coverage, invisible H3 expressions, false overlay readiness, Next.js MapLibre worker loading, Gemini parameter/schema compatibility, safe provider diagnostics, and inconsistent root environment loading.

Validation

The repository owner reported the complete local validation sequence green, including formatting, lint, type checking, unit/integration tests, production build, and Playwright. The owner also ran the explicitly opted-in pnpm smoke:openrouter call successfully against google/gemini-3.7-flash and confirmed normal pnpm dev startup with repository-root .env configuration.

The real-provider smoke remains separate from default tests, Playwright, and CI because it incurs third-party cost.

Security and operating notes

• OPENROUTER_API_KEY remains server-only.
• Raw prompts, observations, request/response bodies, secrets, and private reasoning are not recorded or exposed.
• Provider failures are public-safe; bounded detailed diagnostics are available only to the local opt-in smoke command.
• The cost-incurring development API remains loopback-bound and is not suitable for unauthenticated public deployment.
• API process restart resets the in-memory simulation.