LIMONATA PROVIDER 0.8.0 — SHARED RESIDENT MODEL

Windows and Linux now offer shared-model contribution as the default mode.
The initial target is the pinned Apache-2.0 Qwen3-32B Q4_K_M model, divided
into two resident portions. Every shared answer executes both portions.

The first portion contains layers 1–48 (14.99 GB download); the second contains
layers 49–64 (5.86 GB download). Admission currently requires approximately
18.2 GiB and 8.2 GiB of free GPU memory, respectively, followed by actual full
GPU inference checks. These estimates are not measured hardware guarantees.
Only a complete verified route opens the shared chat. The page shows loaded
layer coverage, rather than treating a sum of advertised VRAM as a ready model.

Weights remain in each contributor's GPU memory between requests and network
reconnections. Additional contributors prepare replicas independently; they do
not reload active routes. Stopping the client releases its GPU memory. A full
process restart still needs disk verification and a GPU reload.

This release uses a fixed two-portion layout. It does not automatically repartition
the model for arbitrary GPU sizes or replace it with a larger model as capacity
grows. Smaller GPUs can still use the explicit independent-work mode.

Upgrading Linux requires the COMPLETE 0.8.0 archive: it includes the new native
stage runtime. Replacing only the small provider executable from an older release
is insufficient. Stop the old process, verify the complete archive, and extract
into /opt/limonata using tar --no-same-owner. Preserve the existing --data-dir
and its private identity; start with the same public payout address. On Runpod,
use a local filesystem for executable files and back up the private identity
securely before replacing a Pod.

Windows users extract the new archive and start Limonata-Provider.exe. The saved
Windows-account identity is reused. Shared mode detects the GPU, obtains an
assignment, downloads the pinned portion and verifies real GPU execution before
announcing readiness. No public inbound GPU port or wallet private key is needed.

Shared requests in this release are sponsored and pay no LIMO. The dashboard
attributes accepted participation to the actual route and shows whole-group
historical throughput; idle replicas do not receive completed-work credit.
Existing paid independent requests retain their escrow/reward behavior.

Validation includes backend/protocol tests, CPU numerical parity, real two-stage
CPU inference through the production job API and Go transport, concurrent chats,
resident reconnects, standby enrollment and desktop/mobile UI checks. These do
not establish 32B physical GPU fit, Windows Vulkan execution or WAN performance.
The first real Windows/Linux GPU acceptance run remains necessary.

See swarm/SHARED_RELEASE.txt for operating limits and AI_ECONOMICS.txt for the
implemented payment flow and the remaining shared reward distribution work.