blox-ai: populate Qwen 3 1.7B W8A8 release URL + SHA-256 (2-chunk split)

Closes the placeholder gate for the Qwen 3 model swap.

Release: https://github.com/functionland/blox-ai/releases/tag/model-qwen-3-1.7b-w8a8-v1

Assembled file:
  - qwen3-1.7b-rk3588-w8a8.rkllm
  - 2,375,021,452 bytes (~2.21 GB)
  - SHA-256: 8843d4612d42a71605c8d0b38cf9a758a5a53f65a8dde3f17f9d4549b9794e87

Two-chunk split (GitHub release assets cap at 2 GiB per file):
  - part-aa: 1,992,294,400 bytes (~1.9 GB)
  - part-ab:   382,727,052 bytes (~365 MB)

download_model.sh fetches both chunks in CHUNK_URLS order, cats them
into MODEL_FILE, then SHA-verifies the assembled result. The
SINGLE_CHUNK_IS_MODEL_FILE shortcut (line 307) correctly evaluates
false here because the chunk basenames don't equal MODEL_FILE's
basename, so the cat path runs as designed.

The .rkllm was produced from a Qwen 3 1.7B Instruct LoRA fine-tune
(2 epochs on 174 synthetic Blox AI examples, final loss 0.39),
merged + quantized W8A8 via rkllm-toolkit 1.2.3 targeting RK3588.
The on-device runtime requires librkllmrt.so v1.2.3 — the
sibling functionland/blox-ai container was rebuilt 2026-05-27
(commit 6d48a48) to bump the runtime from 1.1.4 to 1.2.3.

Release-placeholder-gate now passes (no more __SET_BEFORE_RELEASE__
strings in the plugin dir).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: ehsan6sha

docker/fxsupport/linux/plugins/blox-ai/custom/download_model.sh

@@ -30,11 +30,20 @@ set -e
 # release tag that bundles this script. The fail-fast guard further
 # down rejects any download attempt while placeholders are present
 # so a device can never silently pull an unverified blob.
+# Two-chunk release. The assembled file is 2,375,021,452 bytes (~2.21 GB);
+# GitHub's per-asset cap is 2 GiB, so the .rkllm was split into:
+#   - part-aa: 1,992,294,400 bytes (~1.9 GB)
+#   - part-ab:   382,727,052 bytes (~365 MB)
+# The download loop below fetches both, cat's them into MODEL_FILE in
+# array order, then verifies the assembled SHA-256.
 CHUNK_URLS=(
-    "https://github.com/functionland/blox-ai/releases/download/model-qwen-3-1.7b-w8a8-v1/qwen3-1.7b-rk3588-w8a8.rkllm"
+    "https://github.com/functionland/blox-ai/releases/download/model-qwen-3-1.7b-w8a8-v1/qwen3-1.7b-rk3588-w8a8.rkllm.part-aa"
+    "https://github.com/functionland/blox-ai/releases/download/model-qwen-3-1.7b-w8a8-v1/qwen3-1.7b-rk3588-w8a8.rkllm.part-ab"
 )
-DOWNLOAD_URL="https://github.com/functionland/blox-ai/releases/download/model-qwen-3-1.7b-w8a8-v1/qwen3-1.7b-rk3588-w8a8.rkllm"
-MODEL_SHA256="__SET_BEFORE_RELEASE__"
+# First chunk URL kept here for the manifest-helper fallback path (Phase 18).
+DOWNLOAD_URL="https://github.com/functionland/blox-ai/releases/download/model-qwen-3-1.7b-w8a8-v1/qwen3-1.7b-rk3588-w8a8.rkllm.part-aa"
+# SHA-256 of the ASSEMBLED file (cat part-aa part-ab), NOT of individual chunks.
+MODEL_SHA256="8843d4612d42a71605c8d0b38cf9a758a5a53f65a8dde3f17f9d4549b9794e87"
 
 MODEL_DIR="/uniondrive/blox-ai/model"
 MODEL_FILE="$MODEL_DIR/qwen3-1.7b-rk3588-w8a8.rkllm"