diff --git a/.github/workflows/llama-pin-check.yml b/.github/workflows/llama-pin-check.yml
index 751b75d3..78343568 100644
--- a/.github/workflows/llama-pin-check.yml
+++ b/.github/workflows/llama-pin-check.yml
@@ -55,7 +55,13 @@ jobs:
           _Auto-filed by the weekly \`llama.cpp pin freshness\` workflow._
           EOF
           )"
-          existing="$(gh issue list --state open --search "$title in:title" --json number --jq '.[0].number // empty')"
+          # Exact-title match against the open issues, NOT `--search`. The
+          # title contains "(" and ":" — GitHub's search parser reads
+          # "chore(server/cuda):" as a (bogus) qualifier and matches nothing,
+          # so every weekly run filed a fresh duplicate instead of commenting
+          # on the existing reminder (that is how #170 and #205 both existed).
+          existing="$(gh issue list --state open --limit 200 --json number,title \
+            | jq -r --arg t "$title" 'map(select(.title == $t)) | .[0].number // empty')"
           if [ -n "$existing" ]; then
             echo "Updating existing issue #$existing"
             gh issue comment "$existing" --body "$body"
diff --git a/server/Dockerfile b/server/Dockerfile
index c79e2ece..3d3a26c8 100644
--- a/server/Dockerfile
+++ b/server/Dockerfile
@@ -127,7 +127,7 @@ RUN mkdir -p /out/data/models
 # target platform. Bump this digest deliberately and re-run `make scout-cpu`.
 # Resolve a new digest with:
 #   docker buildx imagetools inspect ghcr.io/ggml-org/llama.cpp:server
-FROM ghcr.io/ggml-org/llama.cpp:server@sha256:6bc9134e3278a0ecab23d7ef2f6a46b4595740014fe9bc2f67e8ba7dca8395b4 AS llama-source
+FROM ghcr.io/ggml-org/llama.cpp:server@sha256:9e60dd363798776f451bcca97f22f5d75579480e53b0b46303d061a69ceea4df AS llama-source
 
 # ── Stage: distroless runtime ──────────────────────────────────────────────
 # gcr.io/distroless/cc-debian13:nonroot (Debian 13 trixie) — NOT static-debian12.
diff --git a/server/Dockerfile.cuda b/server/Dockerfile.cuda
index 07de417f..07eb4221 100644
--- a/server/Dockerfile.cuda
+++ b/server/Dockerfile.cuda
@@ -101,7 +101,7 @@ RUN curl -fsSL \
 # missing libllama-server-impl.so). Bump this digest deliberately and re-run
 # `make scout-cuda` to validate. Resolve a new digest with:
 #   docker buildx imagetools inspect ghcr.io/ggml-org/llama.cpp:server-cuda
-FROM ghcr.io/ggml-org/llama.cpp:server-cuda@sha256:7b3d7834fc7307cb54f24f8869b67bfff276404c416452a48d11321bc36a81be AS llama-source
+FROM ghcr.io/ggml-org/llama.cpp:server-cuda@sha256:fd68d13013141833e8214ecad6e1fbefb532db6a00b980cdecfe33603dbf2675 AS llama-source
 
 # ── Stage 3: extract CUDA shared libraries ─────────────────────────────────
 # Install the CUDA libs here, then COPY individual .so files into the
diff --git a/server/Makefile b/server/Makefile
index 4c4bda24..3edae8f8 100644
--- a/server/Makefile
+++ b/server/Makefile
@@ -5,7 +5,7 @@
 # commit — the new SHA256 must be appended to scripts/llama-checksums.txt and
 # the parity gate re-run before shipping.
 
-LLAMA_VERSION ?= b8914
+LLAMA_VERSION ?= b10238
 LLAMA_REPO    ?= ggml-org/llama.cpp
 
 # `make bundle OS=... ARCH=...` supports only darwin-arm64 in Phase 3.
diff --git a/server/scripts/fetch-llama.sh b/server/scripts/fetch-llama.sh
index daf45d56..eae1ea03 100755
--- a/server/scripts/fetch-llama.sh
+++ b/server/scripts/fetch-llama.sh
@@ -87,37 +87,19 @@ mkdir -p "$DEST_DIR"
 # Clean out any previous fetch — stale dylibs could get picked up by DYLD.
 rm -f "$DEST_DIR"/* 2>/dev/null || true
 
-# Files we ship. llama-server is the only binary we need; dylibs are its
-# runtime deps. We deliberately drop llama-cli, llama-bench, llama-quantize,
-# rpc-server, llama-server's *-debug variants, mtmd-*, etc. to keep the
-# bundle lean.
-SHIP=(
-    "llama-server"
-    "libllama.dylib"
-    "libllama-common.dylib"
-    "libmtmd.dylib"
-    "libggml.dylib"
-    "libggml-base.dylib"
-    "libggml-cpu.dylib"
-    "libggml-metal.dylib"
-    "libggml-blas.dylib"
-    "libggml-rpc.dylib"
-)
-# Versioned dylib aliases — dyld resolves these via symlink/rpath. Include
-# everything that matches the base names so @rpath lookups do not break.
-for base in "${SHIP[@]}"; do
-    # Copy the bare file if present.
-    if [[ -e "$INNER_DIR/$base" ]]; then
-        cp -p "$INNER_DIR/$base" "$DEST_DIR/"
-    fi
-    # Copy any versioned variants (libfoo.0.dylib, libfoo.0.0.1234.dylib, ...)
-    # that begin with the same stem. Loose glob: for each dylib name stem we
-    # look for "<stem>.*.dylib".
-    stem="${base%.dylib}"
-    for match in "$INNER_DIR/$stem".*.dylib; do
-        [[ -e "$match" ]] || continue
-        cp -p "$match" "$DEST_DIR/"
-    done
+# Files we ship: llama-server plus the WHOLE dylib set, not a hand-maintained
+# list. Upstream periodically refactors the shared-library layout — b10238
+# moved each tool's logic into its own libllama-<tool>-impl.dylib, so the old
+# fixed list produced a bundle whose llama-server died at dyld load time with
+# "Library not loaded: @rpath/libllama-server-impl.dylib". The same lesson is
+# already encoded in Dockerfile/Dockerfile.cuda (`COPY /app/*.so*`). The other
+# tools' impl dylibs cost ~1 MB out of a ~42 MB bundle — far cheaper than a
+# broken bundle. Standalone binaries (llama-cli, llama-bench, llama-quantize,
+# ggml-rpc-server, mtmd-*, …) are still dropped.
+cp -p "$INNER_DIR/llama-server" "$DEST_DIR/"
+for match in "$INNER_DIR"/*.dylib; do
+    [[ -e "$match" ]] || continue
+    cp -p "$match" "$DEST_DIR/"
 done
 
 # Sanity: llama-server must be present and executable.
@@ -126,6 +108,22 @@ if [[ ! -x "$DEST_DIR/llama-server" ]]; then
     exit 1
 fi
 
+# Sanity: every @rpath dependency of llama-server must have landed in
+# DEST_DIR. Without this the breakage only surfaces at runtime, on the
+# operator's machine, as a dyld abort — exactly how the b10238 layout change
+# was found. Fail the fetch instead.
+missing=()
+while read -r dep; do
+    [[ -n "$dep" ]] || continue
+    [[ -e "$DEST_DIR/$dep" ]] || missing+=("$dep")
+done < <(otool -L "$DEST_DIR/llama-server" | awk '/@rpath\//{sub(/^.*@rpath\//, "", $1); print $1}')
+if [[ ${#missing[@]} -gt 0 ]]; then
+    echo "fetch-llama: llama-server has unresolved @rpath dependencies:" >&2
+    printf '  %s\n' "${missing[@]}" >&2
+    echo "fetch-llama: upstream $LLAMA_VERSION likely changed its library layout — inspect the release archive." >&2
+    exit 1
+fi
+
 # macOS Gatekeeper quarantine can apply to downloaded binaries even via curl.
 # Strip the attribute so end users do not hit a silent kill on first run.
 if command -v xattr >/dev/null 2>&1; then
diff --git a/server/scripts/llama-checksums.txt b/server/scripts/llama-checksums.txt
index 15684e56..354d12bd 100644
--- a/server/scripts/llama-checksums.txt
+++ b/server/scripts/llama-checksums.txt
@@ -3,3 +3,4 @@
 # Bumping LLAMA_VERSION in the Makefile MUST be accompanied by a commit that
 # records the new row here. CI fails hard on SHA mismatch.
 cd70e17321b822820f757a250c2f8128c69290f43048f6e7bc79ec8822a1a5c5  llama-b8914-bin-macos-arm64.tar.gz
+1bc2c461b1c64f8f71c99deb55a21d5e43e2deeea1d398f8bb0a1969eb1306aa  llama-b10238-bin-macos-arm64.tar.gz