feat(prefill): enforce decode-only primary mode

[FluffyAIcode]

Summary

• add remote-required prefill policy: Primary accepts only complete cache hits or completed remote worker snapshots and never computes prefill/suffix locally
• bypass cost rejection in strict mode while retaining load-aware worker selection
• map unavailable remote prefill to gRPC UNAVAILABLE instead of silent local fallback
• deploy allens as the Gemma-loaded prefill-only worker with a co-located KakeyaLattice cache on :53051
• expose all physical cache mounts through one kv:// content-addressed namespace while explicitly reporting that coherent shared memory is false

Architecture

Primary receives canonical token IDs, allens performs full prefill and stores snapshots, Primary fetches/imports once, and all autoregressive decode remains local. Partial cache hits are never extended on Primary.

Tests

• platform-neutral Python suite — 833 passed, 1 skipped
• TypeScript SDK — 54 passed, 100% coverage and typecheck
• strict no-fallback, partial-hit rejection, forced worker selection, UNAVAILABLE mapping, namespace aggregation, API, and launchd role tests pass
• Python compile, plist lint and diff checks pass

Live caveat

The selected Gemma model is ~15GB on a 16GB allens host. Previous functional tests completed but remote prefill took roughly 128–149 seconds; this deployment prioritizes role correctness, not throughput.

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.