Update packaging READMEs for local Qwen usage

.github/workflows/bindings.yml

@@ -95,6 +95,7 @@ jobs:
             ${{ steps.setup-ndk.outputs.ndk-path }}/toolchains/llvm/prebuilt/linux-x86_64/bin/llvm-ranlib
         run: bash scripts/retry-command.sh make codegen-rn
       - uses: actions/upload-artifact@v4
+        if: github.event_name != 'pull_request'
         with:
           name: rn-bindings
           path: out/rn/
@@ -140,6 +141,7 @@ jobs:
       - name: Build WASM ESM package
         run: make build-wasm
       - uses: actions/upload-artifact@v4
+        if: github.event_name != 'pull_request'
         with:
           name: web-bindings
           path: out/web/
@@ -163,6 +165,7 @@ jobs:
       - name: Codegen — Flutter/Dart
         run: make codegen-flutter
       - uses: actions/upload-artifact@v4
+        if: github.event_name != 'pull_request'
         with:
           name: flutter-bindings
           path: out/flutter/

.github/workflows/release.yml

@@ -276,20 +276,8 @@ jobs:
                files: ["src/", "android/", "ios/", "README.md"],
                license: "Apache-2.0"
              }' > assembly/package.json
-      - name: Write README.md
-        run: |
-          cat > assembly/README.md <<'EOF'
-          # edge-intelligence-sdk
-
-          Generated Web and React Native bindings for the Edge Intelligence SDK.
-
-          This package contains the WASM/TypeScript browser surface, React Native
-          TypeScript bindings, and native Android/iOS libraries assembled by the
-          release pipeline.
-
-          Source, crate documentation, and release notes live in the Edge
-          Intelligence repository.
-          EOF
+      - name: Copy README.md
+        run: cp packaging/npm/README.md assembly/README.md
       - name: Assert npm package ships README.md
         run: bash scripts/assert-release-readmes.sh npm assembly
       - uses: actions/upload-artifact@v4

Cargo.toml

@@ -38,7 +38,7 @@ exclude = [
 
 [workspace.package]
 edition = "2021"
-version = "0.3.9"
+version = "0.3.10"
 license = "Apache-2.0"
 rust-version = "1.96"
 repository = "Edge-Native LLM SDK"
@@ -48,16 +48,16 @@ repository = "Edge-Native LLM SDK"
 # This is the single source of truth for internal dep version requirements;
 # member crates reference them with `{ workspace = true }` and add no inline
 # version fields. `cargo set-version --workspace <ver>` updates every entry.
-el-core               = { path = "crates/el-core", version = "0.3.9" }
-el-memory             = { path = "crates/el-memory", version = "0.3.9" }
-el-telemetry          = { path = "crates/el-telemetry", version = "0.3.9" }
-el-provenance         = { path = "crates/el-provenance", version = "0.3.9" }
-el-safety             = { path = "crates/el-safety", version = "0.3.9" }
-el-runtime            = { path = "crates/el-runtime", version = "0.3.9" }
-el-grammar            = { path = "crates/el-grammar", version = "0.3.9" }
-el-provenance-ed25519 = { path = "crates/adapters/el-provenance-ed25519", version = "0.3.9" }
-el-engine-candle      = { path = "crates/adapters/el-engine-candle", version = "0.3.9" }
-el-cloud              = { path = "crates/adapters/el-cloud", version = "0.3.9" }
+el-core               = { path = "crates/el-core", version = "0.3.10" }
+el-memory             = { path = "crates/el-memory", version = "0.3.10" }
+el-telemetry          = { path = "crates/el-telemetry", version = "0.3.10" }
+el-provenance         = { path = "crates/el-provenance", version = "0.3.10" }
+el-safety             = { path = "crates/el-safety", version = "0.3.10" }
+el-runtime            = { path = "crates/el-runtime", version = "0.3.10" }
+el-grammar            = { path = "crates/el-grammar", version = "0.3.10" }
+el-provenance-ed25519 = { path = "crates/adapters/el-provenance-ed25519", version = "0.3.10" }
+el-engine-candle      = { path = "crates/adapters/el-engine-candle", version = "0.3.10" }
+el-cloud              = { path = "crates/adapters/el-cloud", version = "0.3.10" }
 
 [profile.release]
 opt-level = 3

Makefile

@@ -83,13 +83,7 @@ codegen-flutter:
 		'  flutter_rust_bridge: ^$(FRB_VERSION)' \
 		> $(OUT)/flutter/pubspec.yaml
 	@cp LICENSE $(OUT)/flutter/LICENSE
-	@printf '%s\n' \
-		'# edge_intelligence' \
-		'' \
-		'Generated Flutter bindings for the Edge Intelligence SDK.' \
-		'' \
-		'The package contains Dart bindings generated with flutter_rust_bridge and native libraries assembled by the release pipeline.' \
-		> $(OUT)/flutter/README.md
+	@cp packaging/flutter/README.md $(OUT)/flutter/README.md
 	@printf '%s\n' \
 		'# Changelog' \
 		'' \

crates/adapters/el-cloud/README.md

@@ -46,11 +46,15 @@ a clean completion), and errors carry only parse category/position/size —
 
 ## Usage
 
+Use `CloudProvider` only as an explicit fallback or comparison path; the local
+Qwen2.5 0.5B GGUF remains on device and is not sent to the provider.
+
 ```rust
 use el_core::{ChatMessage, ChatRequest, CredentialRef, LlmProvider};
 use el_cloud::CloudProvider;
 
 let provider = CloudProvider::new();
+let _local_qwen_model = "models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
 
 // The credential is resolved at runtime from the platform keystore — never embedded.
 let req = ChatRequest::new("openai/gpt-4o", vec![ChatMessage::user("Hello!")])

crates/adapters/el-engine-candle/README.md

@@ -30,17 +30,22 @@ Mismatched shapes are rejected at load time, not silently at inference.
 
 ## Usage
 
+This is the direct Rust SDK path for a local Qwen2.5 0.5B instruct model.
+
 ```rust
 use el_core::{ChatMessage, ChatRequest, LlmProvider};
 use el_engine_candle::QwenChatProvider;
 
+const QWEN_0_5B_GGUF: &str = "models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
+const QWEN_0_5B_TOKENIZER: &str = "models/qwen2.5-0.5b-instruct.tokenizer.json";
+
 // Real on-device chat: a local Qwen2 GGUF + its tokenizer (no network egress).
-let provider = QwenChatProvider::from_paths(
-    "models/qwen2.5-0.5b-instruct-q4_k_m.gguf",
-    "models/qwen2.5-0.5b-instruct.tokenizer.json",
-)?;
+let provider = QwenChatProvider::from_paths(QWEN_0_5B_GGUF, QWEN_0_5B_TOKENIZER)?;
 
-let req = ChatRequest::new("local", vec![ChatMessage::user("What is the capital of France?")])
+let req = ChatRequest::new(
+    "local/qwen2.5-0.5b-instruct-q4_k_m",
+    vec![ChatMessage::user("Summarize edge inference in one sentence.")],
+)
     .with_max_tokens(128);
 let reply = provider.chat(&req)?;
 println!("{}", reply.content);

crates/adapters/el-ffi/README.md

@@ -34,19 +34,74 @@ No `unsafe` (`#![forbid(unsafe_code)]`). The crate is `cdylib` + `staticlib` +
 
 ## Usage (Rust side)
 
+Use the Qwen2.5 0.5B GGUF as the local model URI when constructing the facade.
+Native package bindings use the same model path after copying the file into app
+storage. For full Rust ChatML/tokenizer control, use
+`el_engine_candle::QwenChatProvider` directly.
+
 ```rust
 use el_ffi::EdgeLlm;
 
-// Empty path → deterministic toy model; exercises the full binding layer.
-let sdk = EdgeLlm::local(String::new())?;
-let reply = sdk.ask("hello".into())?;
+const QWEN_0_5B_GGUF: &str = "models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
+
+let sdk = EdgeLlm::local(QWEN_0_5B_GGUF.into())?;
+let reply = sdk.ask("Summarize edge inference in one sentence.".into())?;
 assert!(!reply.is_empty());
 
 // Streaming (Flutter / closure form):
-sdk.ask_stream("hi".into(), |fragment| print!("{fragment}"))?;
+sdk.ask_stream("Give me two deployment tips.".into(), |fragment| print!("{fragment}"))?;
 # Ok::<(), el_ffi::SdkError>(())
 ```
 
+## Usage (npm / web)
+
+The npm package exposes the wasm-bindgen browser surface. The local web path
+currently exercises the generated API shape while Candle-on-wasm is being wired;
+native React Native builds load the GGUF through the same package.
+
+```ts
+import init, { EdgeLlm } from "edge-intelligence-sdk";
+
+await init();
+
+const qwen05b = "/models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
+const sdk = new EdgeLlm(qwen05b);
+const reply = sdk.ask_wasm("Summarize edge inference in one sentence.");
+
+console.log(reply);
+```
+
+## Usage (React Native)
+
+```ts
+import { EdgeLlm } from "edge-intelligence-sdk";
+
+const qwen05b = "/data/user/0/com.example.app/files/models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
+const sdk = EdgeLlm.local(qwen05b);
+
+const reply = sdk.ask("Summarize edge inference in one sentence.");
+let streamed = "";
+sdk.ask_stream_cb("Give me two deployment tips.", {
+  on_token(token) {
+    streamed += token;
+  },
+});
+```
+
+## Usage (pub.dev / Flutter)
+
+```dart
+import "package:edge_intelligence/edge_intelligence.dart";
+
+final qwen05b = "/path/to/app/models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
+final sdk = await EdgeLlm.local(qwen05b);
+
+final reply = await sdk.ask("Summarize edge inference in one sentence.");
+await for (final token in sdk.askStream("Give me two deployment tips.")) {
+  stdout.write(token);
+}
+```
+
 ## Building the bindings
 
 The Rust binding *surfaces* compile on the host; the cross-target builds and

crates/adapters/el-grammar-llguidance/README.md

@@ -25,11 +25,16 @@ The HuggingFace `tokenizers::Tokenizer` is bridged to llguidance's `TokEnv` via
 
 ## Usage
 
+Use the same Qwen2.5 0.5B tokenizer that the chat provider uses, so schema
+masking and model decoding agree on token ids.
+
 ```rust
 use el_grammar_llguidance::LlguidanceMasker;
 use el_runtime::GrammarMasker;
 
-let tokenizer = tokenizers::Tokenizer::from_file("tokenizer.json").unwrap();
+let tokenizer = tokenizers::Tokenizer::from_file(
+    "models/qwen2.5-0.5b-instruct.tokenizer.json",
+).unwrap();
 let masker = LlguidanceMasker::from_tokenizer(&tokenizer, r#"{"type":"integer"}"#)?;
 
 // In the decode loop, wired into the session's Ports:

crates/adapters/el-provenance-ed25519/README.md

@@ -21,16 +21,24 @@ do about it (issue a `LoadPermit`, or hard-stop). Pure-Rust dependency tree, no
 
 ## Usage
 
+Use this adapter when a shipped Qwen2.5 0.5B GGUF is accompanied by a detached
+ED25519 signature and a trusted provider public key.
+
 ```rust
 use el_core::{ModelFormat, ModelId, ModelVersion};
 use el_provenance::ModelArtifact;
 use el_provenance_ed25519::Ed25519Verifier;
+use std::path::Path;
 
 let mut verifier = Ed25519Verifier::new();
 verifier.register(/* public_key_id */ 1, trusted_public_key_bytes)?;
 
+let qwen_path = Path::new("models/qwen2.5-0.5b-instruct-q4_k_m.gguf");
+let model_bytes = std::fs::read(qwen_path)?;
+let signature_bytes = std::fs::read(qwen_path.with_extension("gguf.sig"))?;
+
 let mut artifact = ModelArtifact::new(ModelId(1), ModelVersion::new(0, 1, 0), ModelFormat::Gguf);
-artifact.verify(&verifier, model_bytes, signature_bytes, 1);
+artifact.verify(&verifier, &model_bytes, &signature_bytes, 1);
 
 // Verified → a LoadPermit; tampered bytes, a forged signature, or an unknown
 // key id → a hard error with no fallback.

crates/el-core/README.md

@@ -40,6 +40,12 @@ These are enforced by the type system, not by convention:
 
 ## Usage
 
+This is the crate-local part of a real local SDK call. In an app, the request
+can be served by `el_engine_candle::QwenChatProvider::from_paths(...)` with
+`models/qwen2.5-0.5b-instruct-q4_k_m.gguf` and
+`models/qwen2.5-0.5b-instruct.tokenizer.json`; `el-core` only defines the
+backend-agnostic contract.
+
 ```rust
 use el_core::{ChatMessage, ChatRequest, LlmProvider, SessionConfig};
 
@@ -49,7 +55,10 @@ assert!(!cfg.hybrid_mode);
 
 // A backend-agnostic request. `model` is a routing hint:
 //   "local"/"" → local engine, "openai/…", "anthropic/…", "ollama/…", "gemini/…"
-let req = ChatRequest::new("local", vec![ChatMessage::user("Hello!")])
+let req = ChatRequest::new(
+    "local/qwen2.5-0.5b-instruct-q4_k_m",
+    vec![ChatMessage::user("Summarize edge inference in one sentence.")],
+)
     .with_max_tokens(256)
     .with_temperature(700); // 700 = 0.7 (milli to keep the type Eq-able)