feat(android): add on-device RAG and guard pipeline

.gitignore

@@ -79,6 +79,7 @@ build_tmp/
 CLAUDE.md
 .claude/
 .cursor/
+.learnings/
 
 # Prebuilt llama.xcframework: 必须从 ./llama.cpp-omni 子模块构建出来再放到这里，
 # 仓库不再追踪二进制本身 (~189 MB)，详见 README "构建 llama.xcframework" 一节

MiniCPM-V-demo-Android/.gitattributes

@@ -0,0 +1,4 @@
+.gitattributes text eol=lf
+models/rag-guard-v4-2-e5/*.md text eol=lf
+models/rag-guard-v4-2-e5/*.json text eol=lf
+models/rag-guard-v4-2-e5/*.onnx filter=lfs diff=lfs merge=lfs -text

MiniCPM-V-demo-Android/.gitignore

@@ -18,17 +18,44 @@
 .externalNativeBuild
 .cxx
 local.properties
+signing.local.properties
+environment.local.bat
 
 app/build/
 app/.cxx/
 .kotlin/
+.gradle-user-home/
+.android/
+.android-user-home/
+.gradle-local/
+.android-local/
+
+# Persist Graphify's reusable graph products, not machine-specific or rebuildable intermediates.
+graphify-out/.graphify_python
+graphify-out/.graphify_root
+graphify-out/.graphify_analysis.json
+graphify-out/.graphify_ast.json
+graphify-out/.graphify_detect.json
+graphify-out/.graphify_extract.json
+graphify-out/.graphify_semantic.json
+graphify-out/.graphify_incremental.json
+graphify-out/cache/
+graphify-out/20??-??-??/
+
+# Local Codex hooks contain machine-specific executable paths.
+.codex/
 
 # Native libs produced by the buildGgmlCpu_v86 task (built per-machine).
 app/src/main/jniLibs/
 
 *.gguf
 *.bin
 
+# Python tooling caches.
+__pycache__/
+*.py[cod]
+tools/rag_guard/data/generated/
+
 app/src/main/cpp/build-llama/
 app/src/main/cpp/build-mtmd/
 

MiniCPM-V-demo-Android/.graphifyignore

@@ -0,0 +1,27 @@
+# Generated outputs and local build state.
+graphify-out/
+app/build/
+build/
+.gradle/
+.gradle-user-home/
+.android/
+.android-user-home/
+.cxx/
+app/.cxx/
+
+# Binary UI variants and reference audio do not carry architecture knowledge.
+app/src/main/res/mipmap-*/ic_launcher*.webp
+app/src/main/assets/ref_audios/
+
+# Pure configuration/schema JSON produces no AST nodes; its meaning is captured
+# by AGENTS.md, Room migration code, manifests, and project documentation.
+.codex/hooks.json
+app/schemas/**/*.json
+tools/rag_guard/data/dataset_sources.json
+
+# Generated model/data artifacts are represented by manifests and source code.
+*.gguf
+*.onnx
+*.bin
+*.apk
+tools/rag_guard/data/generated/

MiniCPM-V-demo-Android/AGENTS.md

@@ -0,0 +1,34 @@
+# Android build and installation rules
+
+## Stable application signing
+
+- Never install `com.example.minicpm_v_demo` with Gradle's generated debug key.
+- Every device install and connected Android test must first run `verifyInstallationSigning`.
+- Use the canonical certificate pinned in `app/build.gradle.kts`; keep the keystore and credentials outside Git via `signing.local.properties` or Gradle properties.
+- Before changing the pinned certificate, compare it with the installed package certificate and obtain explicit approval for any uninstall that could erase application data.
+- On `INSTALL_FAILED_UPDATE_INCOMPATIBLE`, stop. Do not uninstall automatically and do not generate another key.
+- Do not run Gradle `connected*AndroidTest` tasks: AGP cleanup can uninstall the target package and erase app-private models, conversations, and knowledge bases even when tests pass.
+- For device tests, build and verify signing first, install both APKs with `adb install -r`, then invoke the selected test with `adb shell am instrument`; never uninstall the target package as test cleanup.
+
+## Canonical Windows Android environment
+
+- Run Gradle through `gradlew.bat`; it loads `android-env.bat` and the ignored machine-specific `environment.local.bat`.
+- Keep the single Gradle user home at the workspace root `.gradle-user-home` and the single Android user home at `.android`.
+- Do not create `.gradle-user-home`, `.gradle-local`, `.android-local`, `.android`, or `.android-user-home` inside `MiniCPM-V-demo-Android`.
+- Use JDK 21 and `D:\Android\Sdk`; do not rely on the older outer `D:\Android\platform-tools` PATH entry.
+
+## graphify
+
+This project has a knowledge graph at graphify-out/ with god nodes, community structure, and cross-file relationships.
+
+When the user types `/graphify`, use the installed graphify skill or instructions before doing anything else.
+
+Rules:
+- For codebase questions, first run `graphify query "<question>"` when graphify-out/graph.json exists. Use `graphify path "<A>" "<B>"` for relationships and `graphify explain "<concept>"` for focused concepts. These return a scoped subgraph, usually much smaller than GRAPH_REPORT.md or raw grep output.
+- Dirty graphify-out/ files are expected after hooks or incremental updates; dirty graph files are not a reason to skip graphify. Only skip graphify if the task is about stale or incorrect graph output, or the user explicitly says not to use it.
+- If graphify-out/wiki/index.md exists, use it for broad navigation instead of raw source browsing.
+- Read graphify-out/GRAPH_REPORT.md only for broad architecture review or when query/path/explain do not surface enough context.
+- After modifying code, run `graphify update .` to keep the graph current (AST-only, no API cost).
+- After modifying plans, ADRs, threat models, READMEs, or other documents, use the installed graphify skill to refresh semantic extraction before finishing the task; `graphify update .` alone is insufficient because it only refreshes code AST.
+- Before reporting a local task complete, run `graphify check-update .`. If semantic changes are pending, update them with an available configured LLM backend or an explicitly approved semantic-extraction sub-agent; never stamp a failed or omitted document as current.
+- Keep `graphify-out/graph.json`, `graphify-out/GRAPH_REPORT.md`, `graphify-out/graph.html`, labels, manifest, and cost audit in the workspace as persistent project artifacts.