feat: add reproducible zero-swarm comparison harness

[christopherjnelson]

Summary

• Add an offline, seeded three-way comparison of legacy multi-agent, zero-swarm v1, and a deterministic worker baseline through the real SimulationService and world engine.
• Report per-tick territory, capture and survival, provider attempts, Zero plans, reflex confidence/distribution, stalls, replan signals, token usage, and latency, with a clear disclaimer for synthetic provider telemetry and unavailable billed cost.
• Make deterministic baseline workers choose only from engine-compiled legal candidates without reflex provider calls.
• Preserve legacy observations when message participants are captured and removed from the active roster.
• Document a provisional decision to retain legacy mode until live-provider trials establish better gameplay and cost.

Validation

• pnpm validate — passed locally.
• pnpm test:e2e — passed locally (1 Playwright test).
• pnpm --silent compare:offline --ticks 12 --seeds worker-capture-spawn,comparison-seed-b,comparison-seed-c — passed and repeated byte-for-byte. Report SHA-256: 43ee6680e2b88e679856bbf6b7d7577919da0fa5fce73b253da0bf5dd737965a.

No provider credentials or network calls are used by the comparison harness.