fix: gh api tag creation syntax + remove --depth=1 in Dockerfile [skip-version]

- Fix gh api to use lightweight tags (POST to /git/refs/{ref} with sha body)
  instead of annotated tags (which require message field)
- Handle 422 (already exists) gracefully by checking if tag exists via GET
- Remove --depth=1 from pi git clone in Dockerfile (breaks tag fetches)

Author: localpibox

.github/workflows/build-and-publish.yml

@@ -154,11 +154,18 @@ jobs:
             echo "Tagging $repo@$VERSION (from $branch)"
             sha=$(git ls-remote "https://github.com/$repo.git" "refs/heads/$branch" | awk '{print $1}')
             if [ -n "$sha" ]; then
-              gh api repos/$repo/git/refs/tags \
-                --method POST \
-                -f ref="refs/tags/$VERSION" \
-                -f sha="$sha" \
-                2>/dev/null && echo "  ✅ $repo@$VERSION" || echo "  ⚠️  $repo tag skipped (may exist)"
+              # Create lightweight tags: POST to /git/refs/{ref} with sha body
+              # Handle 422 (already exists) gracefully by checking if tag exists
+              if gh api "repos/$repo/git/refs/refs/tags/$VERSION" \
+                  --method POST \
+                  -f sha="$sha" \
+                  2>/dev/null; then
+                echo "  ✅ $repo@$VERSION"
+              elif gh api "repos/$repo/git/ref/tags/$VERSION" 2>/dev/null > /dev/null; then
+                echo "  ✅ $repo@$VERSION (already exists)"
+              else
+                echo "  ⚠️  $repo tag creation failed (tag doesn't exist)"
+              fi
             else
               echo "  ⚠️  $repo:$branch not found, skipping tag"
             fi

.pi/memory-user.md

@@ -0,0 +1,9 @@
+User Preferences (Memory):
+
+1. **Subprocess Model Routing Verification**: User prefers verifying subprocess model routing by having it perform a visible task (tell a story, do computation) so they can watch NPU/monitoring tools to confirm the right model hardware is active, rather than just checking log output. <!-- created=2026-08-15 -->
+
+2. **Configurable Timeouts & Thresholds**: User explicitly stated timeouts should be configurable, not hardcoded — important for fine-tuning operations. This applies to all timeout/threshold values in subprocess invocations (NPU model calls). <!-- created=2026-08-15 -->
+
+3. **Upstream Fixes Over Downstream Workarounds**: Prefers proper root-cause fixes over workarounds/band-aids. Calls out when a solution masks a deeper design problem (e.g., "this was a workaround not a proper solution"). Expects upstream prevention (fix prompts/architecture) rather than downstream filtering (regex patterns, meta-entry cleanup). <!-- created=2026-08-15 -->
+
+4. **lpb-memory NPU Model**: User wants the lpb-memory extension to use the NPU model (a separate model from the main Qwen3.6-35B-A3B-MTP-GGUF session model) for background memory operations. <!-- created=2026-08-15 -->

Dockerfile

@@ -124,7 +124,7 @@ RUN --mount=type=cache,target=/home/lpb/.npm \
 RUN set -eux; \
     export PATH="/home/lpb/.npm-global/bin:${PATH}"; \
     mkdir -p /opt/pi-src && cd /opt/pi-src; \
-    git clone --depth=1 --branch lpb-dev ${PI_FORK} .; \
+    git clone --branch lpb-dev ${PI_FORK} .; \
     git fetch origin ${PI_REF}; \
     git checkout ${PI_REF}; \
     git config user.email "build@localpibox.dev"; \