feat: upgrade llama.cpp to b10976, guard the JNI exception boundary, smoke the aarch64 fat jars

.github/models.csv

@@ -19,3 +19,4 @@ SmolVLM-500M-Instruct-Q8_0.gguf,https://huggingface.co/ggml-org/SmolVLM-500M-Ins
 mmproj-SmolVLM-500M-Instruct-Q8_0.gguf,https://huggingface.co/ggml-org/SmolVLM-500M-Instruct-GGUF/resolve/main/mmproj-SmolVLM-500M-Instruct-Q8_0.gguf
 Qwen3-TTS-12Hz-1.7B-Base-Q4_K_M.gguf,https://huggingface.co/ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF/resolve/main/Qwen3-TTS-12Hz-1.7B-Base-Q4_K_M.gguf
 mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf,https://huggingface.co/ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF/resolve/main/mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf
+stories260K.gguf,https://huggingface.co/ggml-org/tiny-llamas/resolve/main/stories260K.gguf

.github/verify-patches-applied.sh

@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+
+# SPDX-FileCopyrightText: 2026 Bernard Ladenthin <[email protected]>
+#
+# SPDX-License-Identifier: MIT OR Apache-2.0
+
+# Asserts that every llama/patches/*.patch really reached the fetched llama.cpp tree.
+#
+# WHY THIS EXISTS. The patch applier (llama/cmake/apply-llama-patches.cmake) is fail-loud on
+# "does not apply", so a *stale* patch cannot ship silently. What it cannot detect is a patch
+# that stops having an effect while still applying, and most patches do not need this check
+# because they have a runnable guard that reds CI on every platform if they go missing:
+#
+#   0003, 0006, 0007, 0008  -> jllama.cpp / native_server.cpp call the symbols they add,
+#                              so dropping one is a compile or link error.
+#   0011                    -> the ContentOnlyParseUtf8 tests in src/test/cpp/test_utils.cpp.
+#   0012                    -> src/test/cpp/test_model_split.cpp.
+#   0001, 0002              -> model-gated Java jobs (Windows argv, LoadProgressCallbackTest).
+#
+# `0010` is the exception and the reason for this script. It casts one enum to int inside
+# upstream's `get_res_model_info()`, which is `static` in server-context.cpp and therefore
+# unreachable from jllama_test; reverting it leaves `ctest` completely green. Its only guard is
+# NativeServerAttachIntegrationTest.models_reportNumericVocabType, which is model-gated — so the
+# day a platform stops downloading models, the regression ships. This check runs in the
+# always-on `C++ Tests` job, needs no model, and costs milliseconds.
+#
+# Usage: .github/verify-patches-applied.sh [<llama.cpp-src-dir>]
+# Exit codes: 0 all good, 1 a check failed.
+
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+SRC="${1:-$ROOT/llama/build/_deps/llama.cpp-src}"
+PATCH_DIR="$ROOT/llama/patches"
+STAMP="$SRC/.jllama-patches-applied"
+
+fail() {
+    echo "ERROR: $*" >&2
+    exit 1
+}
+
+[ -d "$SRC" ] || fail "llama.cpp source dir not found: $SRC (configure the build first)"
+[ -f "$STAMP" ] || fail "patch stamp not found: $STAMP — the applier never ran, so the tree is unpatched"
+
+# --- 1. every patch on disk is named in the stamp -----------------------------------------------
+# Self-maintaining on purpose: adding a patch file needs no edit here. The stamp's first line is
+# the checked-out llama.cpp commit; every other line is "<patch filename> <sha256>".
+on_disk=0
+for p in "$PATCH_DIR"/*.patch; do
+    [ -e "$p" ] || fail "no *.patch files in $PATCH_DIR"
+    on_disk=$((on_disk + 1))
+    name="$(basename "$p")"
+    grep -qF "$name" "$STAMP" || fail "patch '$name' is on disk but absent from the stamp $STAMP"
+done
+
+in_stamp="$(($(wc -l < "$STAMP") - 1))"
+[ "$in_stamp" -eq "$on_disk" ] \
+    || fail "stamp lists $in_stamp patch(es) but $on_disk are on disk — the build dir is stale; configure into a fresh one"
+
+# --- 2. the tree is actually modified ------------------------------------------------------------
+# A valid stamp over a clean tree means the patches were reverted after the fact.
+if git -C "$SRC" rev-parse --git-dir >/dev/null 2>&1; then
+    if git -C "$SRC" diff --quiet; then
+        fail "stamp says $on_disk patch(es) applied but '$SRC' is clean — the patched files were reverted"
+    fi
+fi
+
+# --- 3. the one patch with no runnable guard ------------------------------------------------------
+VOCAB_CAST='(int) meta.model_vocab_type'
+SERVER_CONTEXT="$SRC/tools/server/server-context.cpp"
+[ -f "$SERVER_CONTEXT" ] || fail "not found: $SERVER_CONTEXT"
+grep -qF "$VOCAB_CAST" "$SERVER_CONTEXT" \
+    || fail "patches/0010 is not present in $SERVER_CONTEXT: expected '$VOCAB_CAST'.
+         Without the cast, common_json binds the unscoped enum to its bool constructor and
+         GET /models + GET /v1/models report vocab_type as true/false instead of a number.
+         If upstream added the cast themselves, DROP patch 0010 and update this check."
+
+echo "patches verified: $on_disk applied, tree dirty, patches/0010 cast present"

.github/verify-test-counts.sh

@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+
+# SPDX-FileCopyrightText: 2026 Bernard Ladenthin <[email protected]>
+#
+# SPDX-License-Identifier: MIT OR Apache-2.0
+
+# Fails a Java test job when the suite silently stopped running tests.
+#
+# WHY THIS EXISTS. Surefire's working directory is the module basedir while the shared GGUF cache
+# restores to the reactor root, so every model path resolved one directory too deep and EVERY
+# model-gated class aborted in its @BeforeAll assumption — on every test-java-* job, for months,
+# while the pipeline stayed green. Several stale assertions rode along unnoticed.
+#
+# The shape is what defeats the obvious guard: a CLASS-LEVEL assumption failure makes Surefire
+# record tests="0" errors="0" failures="0" skipped="0" for that class. It contributes NO test
+# entries at all, so "did this run skip anything?" is structurally blind to it — a skipped test is
+# still a reported test. Two checks catch it:
+#
+#   1. Any testsuite reporting tests="0". This is the exact signature above, it is precise, and it
+#      is platform-independent: a class that runs nowhere is a bug on every OS.
+#   2. A floor on the total number of tests executed. The backstop for a whole class file going
+#      missing from the run rather than reporting zero.
+#
+# Check 1 is the sensitive one; check 2 is deliberately slack so it never fails spuriously on a
+# platform that legitimately runs fewer tests. Tighten --min-total once real per-platform numbers
+# are known from a green run.
+#
+# Usage: .github/verify-test-counts.sh <surefire-reports-dir> [--min-total N]
+# Exit codes: 0 all good, 1 a check failed, 2 nothing to scan.
+
+set -euo pipefail
+
+REPORT_DIR="${1:-}"
+MIN_TOTAL=0
+shift || true
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --min-total) MIN_TOTAL="$2"; shift 2 ;;
+        *) echo "unknown argument: $1" >&2; exit 1 ;;
+    esac
+done
+
+[ -n "$REPORT_DIR" ] || { echo "usage: $0 <surefire-reports-dir> [--min-total N]" >&2; exit 1; }
+[ -d "$REPORT_DIR" ] || { echo "ERROR: no such directory: $REPORT_DIR" >&2; exit 2; }
+
+shopt -s nullglob
+reports=("$REPORT_DIR"/TEST-*.xml)
+shopt -u nullglob
+
+# An empty input is a failure, never a pass — the same rule verify-bytecode-version.sh follows,
+# for the same reason: a job that produced no reports at all has not proved anything.
+if [ "${#reports[@]}" -eq 0 ]; then
+    echo "ERROR: no TEST-*.xml under $REPORT_DIR — the suite did not run" >&2
+    exit 2
+fi
+
+total=0
+empty_suites=()
+for f in "${reports[@]}"; do
+    # The count lives on the <testsuite> element; take the first match so a nested element
+    # carrying the same attribute name cannot shift the number.
+    n="$(grep -o 'tests="[0-9]*"' "$f" | head -1 | grep -o '[0-9]*' || true)"
+    [ -n "$n" ] || n=0
+    total=$((total + n))
+    if [ "$n" -eq 0 ]; then
+        empty_suites+=("$(basename "$f")")
+    fi
+done
+
+status=0
+
+if [ "${#empty_suites[@]}" -gt 0 ]; then
+    echo "ERROR: ${#empty_suites[@]} test class(es) contributed ZERO test entries:" >&2
+    printf '  %s\n' "${empty_suites[@]}" >&2
+    echo "  A class-level @BeforeAll assumption that fails looks exactly like this. It is NOT a" >&2
+    echo "  skip — the class reports no tests at all, so it is invisible to a skip check. If a" >&2
+    echo "  model is missing, validate-models should have failed the job before this point." >&2
+    status=1
+fi
+
+if [ "$total" -lt "$MIN_TOTAL" ]; then
+    echo "ERROR: only $total test(s) executed, below the floor of $MIN_TOTAL" >&2
+    status=1
+fi
+
+if [ "$status" -eq 0 ]; then
+    echo "test counts verified: $total test(s) across ${#reports[@]} class(es), none empty (floor $MIN_TOTAL)"
+fi
+exit "$status"