fix(prefill): keep allens as cache-only storage

[FluffyAIcode]

Summary

• make the two-Mac production profile route head gossip/lookups/publishes to allens cache-only :52051
• deploy allens with no model process, 8 GiB RAM cache, fleet auth, exact sink/window geometry, and KakeyaLattice v3 storage
• remove worker-placement forcing from the head plist
• document Worker A/B/C as separate future fleet roles that consume canonical token IDs
• make the E2E verifier accept remote-cache reuse by default and require remote_jobs only under --require-worker

Why

The previous deployment incorrectly combined Prefill Compute Worker and cache-only RAM node on allens. In the current two-Mac topology, the primary is the snapshot producer/fallback and allens is storage only; separate model-loaded workers belong on additional machines.

Tests

• topology and acceptance contract tests — 5 passed
• bridge/network/prefill-cache regression selection — 67 passed
• both launchd plists pass plutil -lint
• Python compile and diff checks pass

Post-merge live flow

1. stop ai.kakeya.prefill-worker :53051 on allens
2. install cache-only ai.kakeya.prefill-network :52051
3. restart head against :52051
4. primary computes and publishes a deterministic 512-token snapshot
5. restart primary, replay, and require accepted_path=remote_cache, remote_hits +1, tokens_reused >=512, remote_jobs=0

Made with Cursor