omni-dispatcher — Asolaria federation single-parent dispatcher

The omnidispatcher is the federation's single-parent dispatcher: one process that holds a 1000-slot PID-table in memory and routes FEDENV-v1 envelopes to the right downstream worker — the “one type-blind spawner, addresses cheap, bodies only on crank” engine pattern. This repo is the engine source, carve-out clean.

Published by acer for liris bilateral attack-verify. GitHub = mediator. The engine only moves when envelopes are fed and the operator gate is open.

2026-07-11 verified recovery/storage connection

The dispatch layer now links explicitly to the measured exact-recovery plane:

• Path 1: route a small authenticated address to a node that already retains the object;
• Path 2: route jointly sufficient CRT shadow lanes to a node that reconstructs with no retained original;
• DBBH→DBWH: re-project the candidate and require SHA/shadow/shell equality before emission;
• storage tier: HDD/SSD holds bodies, cubes, shadows, receipts, queues, and cold state;
• neural sidecar: trained GNN/LLM inference remains an optional CPU/GPU route.

Full component-specific analysis and verification:

PATH2-STORAGE-BACKED-DISPATCH-2026-07-11.md

What it is — measured from source

• Single loopback HTTP ingress :4950.
• 1,000-slot PID table loaded from HBP/cold manifest into bySlotId / byHCoord / byPid indices.
• Four priority queues drained by a worker-thread pool.
• Lazy ports :4951–:5950, allocate-on-first-call, LRU eviction, and 300-second idle DARK sweep.
• Passive bus subscriber and heartbeat emissions.
• Per-slot filesystem inboxes, row-hash deduplication, unlink-on-handled.
• FEDENV-v1 validator with required fields, target whitelist, cube/glyph/hash checks, payload ceiling, optional row-hash verification, and cosign-window check.
• Route table for local/CLI/device/citizen/bus/model surfaces.
• Best-effort HBP rejection tee that never changes dispatch control flow.
• Graceful drain and HBP PID-table snapshot on shutdown.

Boot: node omnidispatcher.mjs --boot [--workers=N]
Preflight: node omnidispatcher.mjs --preflight

Why the dispatcher enables low-GPU computers

A storage-rich or CPU-only machine can be a real fabric participant:

persistent state -> HDD/SSD
active routes     -> compact PID table + bounded queues in RAM
exact recovery    -> SHA/BEHCS/CRT/receipts on CPU
neural inference  -> optional remote/local accelerator sidecar

The dispatcher does not require every possible agent body or model to be resident. It materializes only the addressed worker/slot and returns idle slots to DARK.

This does not mean a hard drive executes neural matrix multiplication. It means durable state, routing, proof, and recovery are no longer forced to live in GPU VRAM.

Pre-Asolaria GNN route lineage

The neural sidecar route originates in Jesse's AI healthcare assistant. Four model files were copied byte-for-byte into the Asolaria GNN sidecar:

EdgeLevelGNN    510f78890ec94b113f0610afbade8bafe6ca20e0
PrototypeGNN    99e3087a10ee58e90c0935f5ab63b72fd3cdd07e
ContrastiveGNN  56329e61eb3e6ddb3ee97b46f997dd8dd8c6b39f
GSLGNN          886b3b0c0cdbddba983fa8c3ae083c4520d38f0e

BigPickle later orchestrates L0 :4792, L4 :4793, G1/G2/G3/G4, OmniShannon, SHA fallback, Fischer, and Hookwall. The healthcare comparison is repository-reported training evidence; later trained checkpoints/manifests live in the trained-GNN repository.

Standalone verification from a clone

npm run check
npm test
npm run verify:sha256

• syntax checks every published module;
• smoke tests cover validator rejection, port allocation/eviction, route stubs, response envelopes, and the HBP reject tee;
• SHA verification checks working-tree bytes;
• LF pinning prevents cross-platform line-ending hash drift.

Independent recovery verification — 2026-07-11

• MEASURED_CLAUDE_FABLE5_THIRD_SEAT, operator supplied: Path 1 rustc 1.97 19/19, Path 2 rustc 1.97 30/30.
• AUDITED_GPT_5_6_PRO: complete healthcare-GNN, BigPickle/Fischer, trained-GNN, Hookwall/Shannon, Q-PRISM, white-room, cube-mint, Dispatcher, HyperHermes, reductions, algorithms, and N-Nest audit.
• MEASURED_GPT_DIRECTED_GITHUB_ACTIONS: Rust 1.97.0 runs 29134408321, 29134413119, and 29134419389 all completed successfully.

These Rust receipts validate the exact recovery substrate; they do not replace this repository's own Node tests or claim a new live dispatch benchmark.

Live status — measured 2026-06-25, acer

Running on acer as the :4950 engine. Anchor ASOLARIA-OMNIDISPATCHER-SPEC-2026-05-22.

Relation to the real 100B run

The dispatcher embodies the single-parent/type-blind-spawner discipline. The 100B runner is a sibling tool, not this file. The 100B counters are tallies, not 100B resident agents. Both share the same law: possibility is cheap; bodies are materialized only when cranked.

Carve-out / boundary

• Engine source only; private manifests, corpora, checkpoints, keys, and PII are excluded.
• Acer-local paths require porter overrides.
• The published cosign-window check is structural, not the real cryptographic authority surface.
• Some tests depend on external private harness files; use clone-safe smoke tests where documented.
• Live Hilbra traversal between physical Path-2 poles remains unverified.

Bilateral attack-verify

Liris clone-verified the published engine, added LF pinning, standalone smoke tests, and SHA verification. Formal receipts:

• LIRIS-ATTACK-VERIFY-2026-06-25.md
• LIRIS-ATTACK-VERIFY-2026-06-25.hbp

---

carve-out clean · no cutover · auto_fire_allowed=false · GitHub = mediator.