Add multimodal vision support (mmproj) and image attachments

CLAUDE.md

@@ -6,7 +6,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 Java bindings for [llama.cpp](https://github.com/ggerganov/llama.cpp) via JNI, providing a high-level API for LLM inference in Java. The Java layer communicates with a native C++ library through JNI.
 
-Current llama.cpp pinned version: **b9964**
+Current llama.cpp pinned version: **b9972**
 
 ## Upgrading CUDA Version
 
@@ -421,7 +421,7 @@ needs no extra step here, `build-webui` re-reads the tag and rebuilds the matchi
 ships no UI):
 ```bash
 # needs node/npm + network; embed.cpp is plain C++17 (no npm)
-git clone --depth 1 --branch b9964 https://github.com/ggml-org/llama.cpp /tmp/lc
+git clone --depth 1 --branch b9972 https://github.com/ggml-org/llama.cpp /tmp/lc
 ( cd /tmp/lc/tools/ui && npm ci && npm run build \
   && ( cd dist && find . -type f -not -path './_gzip/*' \
        | while read -r f; do mkdir -p "_gzip/$(dirname "$f")"; gzip -9 -c "$f" > "_gzip/$f"; done ) \
@@ -461,7 +461,7 @@ cache lives in **Depot Cache** over sccache's **WebDAV** backend:
 - `SCCACHE_WEBDAV_TOKEN: ${{ secrets.DEPOT_TOKEN }}` — a Depot **organization** token, stored
   as the repo secret **`DEPOT_TOKEN`**.
 
-Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b9964`), the
+Because `sccache` is **content-addressed** and llama.cpp is pinned (`GIT_TAG b9972`), the
 ~280 upstream object files are byte-identical every run, so a warm cache recompiles only the
 *changed* files. Depot's cache is **shared across all branches** (unlike GitHub's
 per-branch `actions/cache`), so every branch builds incrementally; a `b<nnnn>` version bump
@@ -1237,7 +1237,7 @@ ctest --test-dir build --output-on-failure -R "ResultsToJson"
 
 #### Upstream source location (in CMake build tree)
 
-llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b9964`.
+llama.cpp is fetched via CMake FetchContent, pinned to `GIT_TAG b9972`.
 
 **GoogleTest** is a separate `BUILD_TESTING`-only FetchContent (`GIT_TAG v1.17.0`), used solely
 by the `jllama_test` C++ unit-test binary — not by the shipped library, and not coupled to the

README.md

@@ -7,7 +7,7 @@
 **Build:**  
 ![Java 8+](https://img.shields.io/badge/Java-8%2B-informational)  
 ![Platform](https://img.shields.io/badge/Platform-Linux%20%7C%20macOS%20%7C%20Windows%20%7C%20Android-lightgrey)  
-[![llama.cpp b9964](https://img.shields.io/badge/llama.cpp-%23b9964-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b9964)  
+[![llama.cpp b9972](https://img.shields.io/badge/llama.cpp-%23b9972-informational)](https://github.com/ggml-org/llama.cpp/releases/tag/b9972)  
 [![JPMS](https://img.shields.io/badge/JPMS-modular%20JAR-25A162)](https://openjdk.org/projects/jigsaw/)  
 ![JUnit](https://img.shields.io/badge/tested%20with-JUnit6-25A162)  
 [![JSpecify](https://img.shields.io/badge/JSpecify-1.0.0%20%40NullMarked-25A162)](https://jspecify.dev)  

android-llmservice/README.md

@@ -13,7 +13,9 @@ what it says:
 1. **Pick a GGUF model** from the device's file system (Storage Access Framework — no
    storage permission, Google-Play compliant).
 2. **Chat** with it, fully **on-device**, tokens streaming into a Jetpack Compose UI.
-3. **Switch language** via a flag picker, and **Save/Load** the conversation locally.
+3. Optionally add a **vision projector (mmproj)** and attach an **image** (📎) to a
+   message — for multimodal GGUFs such as Gemma 3 4B (vision) or SmolVLM2.
+4. **Switch language** via a flag picker, and **Save/Load** the conversation locally.
 
 - **App label:** `LLM Service`
 - **applicationId / package:** `net.ladenthin.android.llmservice`

android-llmservice/TODO.md

@@ -83,7 +83,7 @@ Everything else in the brief is feasible with standard AndroidX + the existing b
 | Markdown rendering (code, lists, tables, headings) | Readable answers | M | ⬜ | add a Compose Markdown renderer (e.g. compose-markdown) |
 | Copy / regenerate / stop / clear-chat actions | Standard chat UX | M | ✅ | **stop** cancels the streaming coroutine (façade closes the source); **copy**/**regenerate** act on the last reply; **long-press any bubble** copies that message; **clear** in the top bar with a confirm |
 | Prompt shortcut chips (Summarize / Explain code / Translate / …) | Fast starts | S | ✅ | localized `SuggestionChip`s above the input (fill the draft); 3 chips × 13 languages |
-| Attachment button → image input (vision) | Multimodal chat | L | ⬜ | `ContentPart.imageFile(...)` + a **vision GGUF + mmproj**; SAF image picker |
+| Attachment button → image input (vision) | Multimodal chat | L | ✅ | SAF pickers for both a vision **mmproj** GGUF (model-picker screen) and a query **image** (📎 in the chat input, shown only once a vision projector is loaded); wired through `ContentPart.imageBytes(...)` + `ChatMessage`. Attached images are session-transient (not saved to `session.json`) |
 | Audio input (speech) | Multimodal chat | L | ⬜ | `ContentPart.audioFile(...)` + an audio-capable model + mmproj |
 | Explicit states: loading / generating / stopped / unavailable / "too large for RAM" | Transparency | S | 🔧 | basic states exist; add the rest + messages |
 
@@ -154,7 +154,7 @@ See decision #1. All rows below assume a new in-app Ktor/NanoHTTPD HTTP layer wr
 | Item | Reality today | Effort to adopt |
 |---|---|---|
 | GPU acceleration | **OpenCL/Adreno only** via the separate `net.ladenthin:llama-android-opencl` AAR (Qualcomm Snapdragon/Adreno; device must supply an OpenCL ICD). **No Vulkan** Android artifact exists. | M — swap/offer the OpenCL AAR flavor + runtime detection; non-Adreno devices stay CPU |
-| Vision / audio input | Supported by the binding (`ContentPart.imageFile`/`audioFile`), needs a multimodal GGUF + matching mmproj | L (see §3) |
+| Audio input | Supported by the binding (`ContentPart.audioFile`), needs an audio-capable GGUF + matching mmproj | L (see §3) |
 | 16 KB page size / dlopen-clean `.so` | Already guaranteed by the AAR (CI-enforced) | — |
 
 ## 8. Branding & store
@@ -175,7 +175,7 @@ See decision #1. All rows below assume a new in-app Ktor/NanoHTTPD HTTP layer wr
   icon, helper tooltips.
 - **Phase 2 — the model manager:** managed import library → `GgufInspector` model cards → RAM/compat
   badges → filters → multi-model switch/delete. Then decide on the **download manager** (`INTERNET`).
-- **Phase 3 — multimodal:** image (then audio) input with a vision/audio GGUF + mmproj.
+- **Phase 3 — multimodal:** image input ✅ shipped; audio input with an audio-capable GGUF + mmproj remains.
 - **Phase 4 — hardening/perf:** encrypted conversations, biometric lock, battery-saver + thermal
   throttling, optional Adreno/OpenCL GPU flavor.
 - **Phase 5 — the local API server (`XL`, foundational):** in-app Ktor/NanoHTTPD server wrapping the