feat(distributed): orchestrate same-model prefill workers and KV memory tier

[FluffyAIcode]

Architecture

Implements ADR 0017: one primary Mac owns RuntimeService/session/decode; peer Macs either load the exact same model and execute prefill-only jobs (PREFILL_COMPUTE) or devote RAM to immutable KV snapshots (PREFILL_CACHE). Remote state is imported once before decode; no remote RPC occurs in the autoregressive loop. DFlash/f_theta/fused speculative decode are legacy research, not product completeness criteria.

Implementation

• PrefillWorkerService: idempotent bounded jobs, status/cancel, deadlines, cancellation race guards, TTL/history cap and co-located cache leases.
• MLXPrefillComputeEngine, worker node and headless launchd installer.
• Gossip cards advertise exact model/tenant/sink/window/geometry compatibility, queue tokens/load/TPS/free RAM.
• Primary cost model selects cache import, remote compute or local prefill; rendezvous placement bounds replication.
• Mandatory local fallback for every lookup/job/lease/fetch/checksum/decompress/import failure.
• Snapshot token-boundary/hash/retained-token/continuation-logit validation; bounded chunk assembly and decompression.
• zlib framing, SHA-256, fleet PSK request HMAC, replay window and tenant-HMAC prefix hashes.
• Worker telemetry in network API; operator runbook and ADR 0017.

Runner/CI hardening discovered during validation

• Integration now uses direct path triggers instead of the non-recursive GITHUB_TOKEN label race.
• Fixed listener recovery detection, stale locks and official-service/direct fallback.
• Added a headless system LaunchDaemon watchdog + runner healthcheck.
• Bootstraps git-lfs before checkout (reused worktree retained an LFS post-checkout hook).
• Resolves the real .venv-mac interpreter and stable verifier path.
• Separates Transformers-5 MLX gate from a PEP-668-safe legacy Qwen venv.

Validation — all green

• ✅ Real Mac MLX continuation-equivalence: 1 passed (local prefill vs snapshot import preserves exact logits/argmax/retained tokens/one decode step).
• ✅ Full Mac M4 integration: pass, 2m13s.
• ✅ Linux unit + 100% coverage: pass, 2m14s.
• ✅ TypeScript SDK, proto lint/format/stub drift, package import and Docker: pass.
• ✅ Distributed modules locally: 167 passed, 1 MLX-only skipped, 100% line coverage including prefill_cache_runtime.py.
• ✅ Authenticated loopback orchestration: gossip worker -> remote prefill -> cache lease -> chunked fetch -> validated primary import.