Feat/m12 devshardd integration

.github/workflows/ci.yml

@@ -17,7 +17,7 @@ jobs:
       - name: Set up Go
         uses: actions/setup-go@v5
         with:
-          go-version: "1.25.8"
+          go-version: "1.25.10"
 
       - name: Download dependencies
         run: go mod download
@@ -48,7 +48,7 @@ jobs:
       - name: Set up Go
         uses: actions/setup-go@v5
         with:
-          go-version: "1.25.8"
+          go-version: "1.25.10"
 
       - name: golangci-lint
         uses: golangci/golangci-lint-action@v7

.github/workflows/release.yml

@@ -21,7 +21,7 @@ jobs:
       - name: Set up Go
         uses: actions/setup-go@v5
         with:
-          go-version: "1.25.8"
+          go-version: "1.25.10"
 
       - name: Install Cosign
         uses: sigstore/cosign-installer@v3

.github/workflows/security.yml

@@ -24,7 +24,7 @@ jobs:
       - name: Set up Go
         uses: actions/setup-go@v5
         with:
-          go-version: "1.25.8"
+          go-version: "1.25.10"
 
       - name: Install govulncheck
         run: go install golang.org/x/vuln/cmd/govulncheck@latest

README.md

@@ -7,17 +7,6 @@
 
 One-command CLI for deploying GPU-accelerated Gonka AI inference nodes. Auto-detects GPUs, configures NVIDIA runtime, generates optimized configs, and deploys containers with security hardening.
 
-## What's New (v0.2.1-rc5)
-
-### Blackwell GPU support improvements
-- **Auto-detect latest Blackwell image**: queries GHCR registry for the latest `-blackwell` tag (e.g., `3.0.12-post6-blackwell`) instead of using a hardcoded version. Falls back to suffix convention if registry is unreachable.
-- **`--mlnode-image` flag**: override the MLNode Docker image entirely for custom builds (e.g., `ghcr.io/segovchik/gonka-b300-image:3.0.13-b300-tp1`). Takes highest priority over all auto-detection.
-- **`--attention-backend` flag**: choose between `FLASHINFER` (default) and `FLASH_ATTN` (lower memory footprint for constrained setups like 2×B200 where FlashInfer workspace causes OOM).
-- **Interactive prompts**: setup wizard now asks for custom image and attention backend after showing recommended config. Skipped in `--yes` mode when flags are provided.
-
-### Bug fix
-- **`gonka-nop status` broken as root**: when `state.json` loaded successfully but `AdminURL` was empty, all HTTP requests silently failed. Non-root users weren't affected because they couldn't read `state.json` (0600 permissions), triggering the default-URL fallback path. Fixed by always initializing `StatusConfig` from defaults before overriding with state values.
-
 ## Demo
 
 [![Gonka NOP Demo](https://img.youtube.com/vi/0w6bIEROxUQ/maxresdefault.jpg)](https://youtu.be/0w6bIEROxUQ?si=FJywggqlVax90Ohn)
@@ -158,138 +147,6 @@ gonka-nop ml-node add
 gonka-nop ml-node list
 ```
 
-## GPU-Specific Deployment Guides
-
-NOP auto-detects GPU architecture and selects optimal settings. These guides document real-world tested configurations and known issues per hardware class.
-
-### 8× A100 SXM4 80GB (Ampere, sm_80)
-
-Standard configuration. Works out of the box.
-
-```bash
-gonka-nop setup
-```
-
-| Setting | Value |
-|---------|-------|
-| Image | `mlnode:3.0.12-post6` (auto-selected) |
-| TP | 4 (auto, NOP calculates PP=2 for 8 GPUs) |
-| Backend | FLASHINFER |
-| gpu-memory-utilization | 0.90 |
-| Weight (observed) | ~860 per ML node |
-
-### 2× B200 (Blackwell, sm_100)
-
-Blackwell image auto-detected. **Use FLASH_ATTN** to avoid FlashInfer workspace OOM with chain-enforced `max_model_len=240000`.
-
-```bash
-gonka-nop setup --attention-backend FLASH_ATTN
-```
-
-| Setting | Value |
-|---------|-------|
-| Image | `mlnode:3.0.12-post6-blackwell` (auto from GHCR) |
-| TP | 2 |
-| Backend | **FLASH_ATTN** (FLASHINFER causes OOM on 2×B200) |
-| gpu-memory-utilization | 0.88 |
-| Weight (observed) | ~920 with [T,T] timeslots |
-
-**Known issue:** The chain enforces `--max-model-len 240000` which pre-allocates a large KV cache. With TP=2 on 2×B200, FlashInfer's workspace buffer (~285MB) doesn't fit in the remaining free memory. Switching to FLASH_ATTN eliminates the workspace allocation entirely.
-
-### 4× B200 (Blackwell, sm_100)
-
-Works with alpha4 image or blackwell image. FLASHINFER is OK because TP=4 means less model weight per GPU → more headroom.
-
-```bash
-gonka-nop setup
-```
-
-| Setting | Value |
-|---------|-------|
-| Image | `mlnode:3.0.13-alpha4` or `3.0.12-post6-blackwell` |
-| TP | 4 |
-| Backend | FLASHINFER (enough headroom at TP=4) |
-| gpu-memory-utilization | 0.90 |
-| Weight (observed) | ~2,178 |
-
-### 8× B300 SXM6 AC (Blackwell Ultra, sm_103a)
-
-Requires custom image — standard and blackwell images lack sm_103a CUTLASS kernels. Must use `vllm/vllm-openai:v0.15.1-cu130` as base.
-
-```bash
-gonka-nop setup \
-  --mlnode-image ghcr.io/segovchik/gonka-b300-image:3.0.13-b300-tp1
-```
-
-| Setting | Value |
-|---------|-------|
-| Image | **Custom** (`ghcr.io/segovchik/gonka-b300-image:3.0.13-b300-tp1`) |
-| TP | 1 (8 independent instances, one per GPU) |
-| Backend | FLASHINFER |
-| gpu-memory-utilization | 0.95 |
-| max-model-len | 131072 |
-| max-num-seqs | 128 |
-| Weight (observed) | ~7,700–8,300 |
-| PoC throughput | ~8,700 nonces/min |
-
-**Why custom image?** B300 (sm_103a) needs:
-1. CUDA 13.0 base (`vllm/vllm-openai:v0.15.1-cu130`) for CUTLASS kernel compatibility
-2. Triton ptxas 13.0 (replacing bundled 12.8 that doesn't know sm_103a)
-3. Runner patches: TP=1 for maximum PoC throughput (8 instances vs 2 with TP=4)
-
-**Host requirement:** `cuda-compat-13-0` package must be installed if host CUDA toolkit < 13.0. Mount compat libs into container:
-```yaml
-# docker-compose.mlnode.yml
-volumes:
-  - /usr/local/cuda-13.0/compat:/usr/local/cuda/compat:ro
-environment:
-  - LD_LIBRARY_PATH=/usr/local/cuda/compat
-```
-
-See [B300 deployment report](docs/b300-mlnode-deployment.md) for the full build process.
-
-### 8× H100/H200 SXM 80GB (Hopper, sm_90)
-
-Standard configuration similar to A100.
-
-```bash
-gonka-nop setup
-```
-
-| Setting | Value |
-|---------|-------|
-| Image | `mlnode:3.0.12-post6` (auto-selected) |
-| TP | 4 or 8 (NVLink full mesh enables TP=8) |
-| Backend | FLASHINFER |
-| gpu-memory-utilization | 0.90 |
-
-### Changing Image on a Running Node
-
-Use `ml-node set-image` to swap the MLNode image without full re-setup:
-
-```bash
-# Switch to blackwell image
-gonka-nop ml-node set-image ghcr.io/product-science/mlnode:3.0.12-post6-blackwell
-
-# Switch to custom B300 image
-gonka-nop ml-node set-image ghcr.io/segovchik/gonka-b300-image:3.0.13-b300-tp1
-```
-
-This performs a safe rollout: disable → update compose → pull → recreate → enable.
-
-### Spot Instance Recovery
-
-If a spot/preemptible instance is killed and reprovisioned with the old data disk attached:
-
-1. Mount old disk: `mount /dev/vdb4 /mnt`
-2. Move Docker storage: set `data-root` in `/etc/docker/daemon.json` to mounted disk
-3. Symlink gonka-node: `ln -sf /mnt/root/gonka-node /root/gonka-node`
-4. Fix HF cache path: `ln -sf /mnt/mnt/shared/huggingface /mnt/shared/huggingface`
-5. Install CUDA compat if needed: `apt-get install cuda-compat-13-0`
-6. Start: `set -a && source config.env && set +a && docker compose up -d`
-
-Keys, chain data, and model cache survive on the persistent disk. No re-registration needed if IP stays the same.
-
 ## Commands
 
 | Command | Description |
@@ -307,11 +164,22 @@ Keys, chain data, and model cache survive on the persistent disk. No re-registra
 | `ml-node status` | Detailed ML node status |
 | `ml-node enable/disable` | Enable or disable an ML node |
 | `ml-node set-image` | Change MLNode Docker image and restart (safe rollout) |
+| `govmodel list` | Show this node's per-model PoC opt-in state |
+| `govmodel candidates <model>` | Rank on-chain hosts running `<model>` for delegation |
+| `govmodel delegate <model> <addr>` | Delegate PoC weight for `<model>` to `<addr>` |
+| `govmodel refuse <model>` | Refuse participation in `<model>`'s PoC |
+| `govmodel declare-intent <model>` | Bootstrap-only signal of intent to run `<model>` |
+| `collateral show [addr]` | Show on-chain collateral balance |
+| `collateral deposit <amount>ngonka` | Deposit collateral to back PoC weight |
+| `collateral withdraw <amount>ngonka` | Withdraw collateral (1-epoch unbonding, still slashable) |
+| `repair-permissions` | Re-grant missing ml-ops permissions to the warm key |
 | `download-model` | Pre-download model weights before setup |
 | `reset` | Stop containers and clean up |
 | `cleanup` | Recover disk space |
 | `version` | Print version info |
 
+See [`docs/nop-guide.md`](docs/nop-guide.md) for full end-to-end walkthroughs per topology, including the post-epoch-180 collateral + governance opt-in flow.
+
 ## Setup Flags
 
 | Flag | Description | Used in |