feat(swarm): add zero-swarm reflex execution seam

[christopherjnelson]

PR A: Zero-swarm reflex seam

Introduces an isolated zero-swarm-v1 reflex path while preserving legacy-multi-agent ticks. Production zero-swarm ticks fail explicitly until PR B adds the Agent Zero planner.

• Versioned cognition mode in scenarios, snapshots, and exports; historical data defaults to legacy.
• Engine-derived legal physical actions and server-held opaque candidate mapping; Jev chooses only a candidate ID and the engine validates the mapped action.
• Separate scripted and direct TypeSafe ReflexProvider implementations, pinned to jev-1.13.0; outbound state omits H3 IDs, tick numbers, chat, diplomacy, goals, and prose memory.
• Runtime response validation, bounded 429/529 retry, cancellation and deadline handling, and an independent ledger record per HTTP POST.
• Safe bounded probabilities, confidence, directive attribution, latency, model, and factual token usage; no invented monetary cost, raw payloads, or key in exports.

Validation

The repository owner confirmed in this session that pnpm validate and pnpm test:e2e passed. Focused offline tests cover the full directive-to-engine path, response rejection, retry accounting, cancellation, and rollback-safe attempt retention.

A one-off live call through TypeSafeJevReflexProvider also succeeded via authorized network access outside the default sandbox. TypeSafe returned model jev-1.13.0, a legal opaque candidate, confidence, probabilities for both offered candidates, and factual token usage. The temporary key was supplied only to the process and was not written to the repository or PR.

PR B will add Agent Zero planning and atomic two-stage production ticks.