From 06a86e9d67f7ad8482ab9b4788d3f17069385003 Mon Sep 17 00:00:00 2001
From: Tovli <Dekel@tovli.co.il>
Date: Tue, 23 Jun 2026 12:56:36 +0300
Subject: [PATCH 1/3] Update packaged READMEs for local Qwen usage

---
 .github/workflows/release.yml                 | 16 +----
 Makefile                                      |  8 +--
 crates/adapters/el-cloud/README.md            |  4 ++
 crates/adapters/el-engine-candle/README.md    | 15 +++--
 crates/adapters/el-ffi/README.md              | 63 +++++++++++++++++--
 .../adapters/el-grammar-llguidance/README.md  |  7 ++-
 .../adapters/el-provenance-ed25519/README.md  | 10 ++-
 crates/el-core/README.md                      | 11 +++-
 crates/el-grammar/README.md                   |  4 ++
 crates/el-memory/README.md                    |  7 +++
 crates/el-provenance/README.md                | 12 +++-
 crates/el-runtime/README.md                   |  5 ++
 crates/el-safety/README.md                    |  9 ++-
 crates/el-telemetry/README.md                 |  5 ++
 packaging/flutter/README.md                   | 24 +++++++
 packaging/npm/README.md                       | 46 ++++++++++++++
 16 files changed, 209 insertions(+), 37 deletions(-)
 create mode 100644 packaging/flutter/README.md
 create mode 100644 packaging/npm/README.md

diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 8186034..31d5dbe 100644
--- a/.github/workflows/release.yml
+++ b/.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
diff --git a/Makefile b/Makefile
index 2031b88..086c46b 100644
--- a/Makefile
+++ b/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' \
 		'' \
diff --git a/crates/adapters/el-cloud/README.md b/crates/adapters/el-cloud/README.md
index 82fd053..8949bfa 100644
--- a/crates/adapters/el-cloud/README.md
+++ b/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!")])
diff --git a/crates/adapters/el-engine-candle/README.md b/crates/adapters/el-engine-candle/README.md
index e8f2527..7cc6bb0 100644
--- a/crates/adapters/el-engine-candle/README.md
+++ b/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);
diff --git a/crates/adapters/el-ffi/README.md b/crates/adapters/el-ffi/README.md
index 1172d87..0e9f17f 100644
--- a/crates/adapters/el-ffi/README.md
+++ b/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
diff --git a/crates/adapters/el-grammar-llguidance/README.md b/crates/adapters/el-grammar-llguidance/README.md
index b280cc6..b8f44d9 100644
--- a/crates/adapters/el-grammar-llguidance/README.md
+++ b/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:
diff --git a/crates/adapters/el-provenance-ed25519/README.md b/crates/adapters/el-provenance-ed25519/README.md
index 489f2d8..772214b 100644
--- a/crates/adapters/el-provenance-ed25519/README.md
+++ b/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.
diff --git a/crates/el-core/README.md b/crates/el-core/README.md
index 67cf956..80b0948 100644
--- a/crates/el-core/README.md
+++ b/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)
 
diff --git a/crates/el-grammar/README.md b/crates/el-grammar/README.md
index 377f2d2..35a5864 100644
--- a/crates/el-grammar/README.md
+++ b/crates/el-grammar/README.md
@@ -31,6 +31,10 @@ is wired.
 
 ## Usage
 
+For Qwen2.5 0.5B, build the DFA over token ids from
+`models/qwen2.5-0.5b-instruct.tokenizer.json`, then wire the masker into the
+same runtime ports used by `QwenChatProvider`.
+
 ```rust
 use el_grammar::{Dfa, DfaMasker};
 use el_runtime::GrammarMasker;
diff --git a/crates/el-memory/README.md b/crates/el-memory/README.md
index d2ea895..dcd813f 100644
--- a/crates/el-memory/README.md
+++ b/crates/el-memory/README.md
@@ -29,9 +29,16 @@ Depends only on `el-core`. No `unsafe` (`#![forbid(unsafe_code)]`).
 
 ## Usage
 
+In a Qwen2.5 0.5B local chat run, `el-engine-candle` loads the GGUF/tokenizer
+and the runtime uses memory planning and KV descriptors like these during
+prefill and decode.
+
 ```rust
 use el_memory::{Arena, BufferLifetime, KvRegion, MemoryTier, StaticMemoryPlanner, TensorSpec};
 
+let _qwen_model = "models/qwen2.5-0.5b-instruct-q4_k_m.gguf";
+let _qwen_tokenizer = "models/qwen2.5-0.5b-instruct.tokenizer.json";
+
 // Two same-tier tensors with disjoint lifetimes share one offset, so the tier
 // needs max(size), not the sum.
 let tensors = [
diff --git a/crates/el-provenance/README.md b/crates/el-provenance/README.md
index 25c8fa8..a1644a7 100644
--- a/crates/el-provenance/README.md
+++ b/crates/el-provenance/README.md
@@ -27,9 +27,13 @@ Depends only on `el-core`. No `unsafe` (`#![forbid(unsafe_code)]`).
 
 ## Usage
 
+For a packaged Qwen2.5 0.5B model, verify the local GGUF bytes and detached
+signature before the runtime can construct a session.
+
 ```rust
 use el_core::{ModelFormat, ModelId, ModelVersion};
 use el_provenance::{ModelArtifact, SignatureVerifier};
+use std::path::Path;
 
 // Plug in a real verifier (el-provenance-ed25519) or a test double.
 struct AlwaysOk;
@@ -37,12 +41,16 @@ impl SignatureVerifier for AlwaysOk {
     fn verify(&self, _bytes: &[u8], _sig: &[u8], _key_id: u32) -> bool { true }
 }
 
+let qwen_path = Path::new("models/qwen2.5-0.5b-instruct-q4_k_m.gguf");
+let qwen_bytes = std::fs::read(qwen_path)?;
+let qwen_signature = 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(&AlwaysOk, b"<model-bytes>", b"<signature>", /* public_key_id */ 7);
+artifact.verify(&AlwaysOk, &qwen_bytes, &qwen_signature, /* public_key_id */ 7);
 
 // No verified signature → no permit → no session.
 let permit = artifact.ensure_loadable()?; // Err(UnverifiedModel | SignatureRejected) otherwise
-# Ok::<(), el_core::EdgeError>(())
+# Ok::<(), Box<dyn std::error::Error>>(())
 ```
 
 ## Place in the workspace
diff --git a/crates/el-runtime/README.md b/crates/el-runtime/README.md
index 727eb12..06e9e9f 100644
--- a/crates/el-runtime/README.md
+++ b/crates/el-runtime/README.md
@@ -44,6 +44,11 @@ and covered by tests.
 
 ## Usage
 
+The snippet uses `NullEngine` so the crate stays self-contained. In the real
+Qwen2.5 0.5B path, replace it with the Candle `QwenEngine`/`QwenChatProvider`
+loaded from `models/qwen2.5-0.5b-instruct-q4_k_m.gguf` plus the matching
+tokenizer.
+
 ```rust
 use el_core::{SessionConfig, SessionId, StopReason};
 use el_provenance::{ModelArtifact, SignatureVerifier};
diff --git a/crates/el-safety/README.md b/crates/el-safety/README.md
index 8b572fc..e8bdb34 100644
--- a/crates/el-safety/README.md
+++ b/crates/el-safety/README.md
@@ -39,6 +39,10 @@ Depends only on `el-core`. No `unsafe` (`#![forbid(unsafe_code)]`).
 
 ## Usage
 
+`QwenChatProvider` resolves guard words through the Qwen2.5 0.5B
+`tokenizer.json`; this crate receives the resulting token ids and applies the
+same deterministic steering regardless of the model backend.
+
 ```rust
 use el_core::{DeviceTarget, SafetyMode};
 use el_safety::{LightweightFilter, SafetyModeSelector, SafetySteerer};
@@ -47,8 +51,9 @@ use el_safety::{LightweightFilter, SafetyModeSelector, SafetySteerer};
 let mode = SafetyModeSelector::resolve(SafetyMode::SecDecoding, DeviceTarget::MidRange);
 assert_eq!(mode, SafetyMode::Lightweight);
 
-// Lightweight bans specific token ids outright.
-let filter = LightweightFilter::new(vec![42, 99]);
+// Lightweight bans specific Qwen tokenizer ids outright.
+let qwen_guard_token_ids = vec![42, 99];
+let filter = LightweightFilter::new(qwen_guard_token_ids);
 let adj = filter.adjust(&[]);
 assert_eq!(adj.delta_for(42), LightweightFilter::HARD_BAN);
 assert_eq!(adj.delta_for(7), 0);
diff --git a/crates/el-telemetry/README.md b/crates/el-telemetry/README.md
index e8b7723..0790e22 100644
--- a/crates/el-telemetry/README.md
+++ b/crates/el-telemetry/README.md
@@ -22,12 +22,17 @@ events it observes.
 
 ## Usage
 
+The collector sees only numeric events from a session, whether that session is
+the local Qwen2.5 0.5B path or another `LlmProvider`; prompts and responses
+never enter telemetry.
+
 ```rust
 use el_core::{DomainEvent, EventEnvelope, SessionId};
 use el_telemetry::MetricsCollector;
 
 let mut collector = MetricsCollector::new();
 
+// Example events from a local Qwen2.5 0.5B prefill/decode pass.
 collector.observe(&EventEnvelope::new(
     SessionId(1),
     0,
diff --git a/packaging/flutter/README.md b/packaging/flutter/README.md
new file mode 100644
index 0000000..85864cc
--- /dev/null
+++ b/packaging/flutter/README.md
@@ -0,0 +1,24 @@
+# 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.
+
+## Usage
+
+Copy the Qwen2.5 0.5B GGUF into app storage, then open the local SDK facade with
+that model path.
+
+```dart
+import "dart:io";
+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);
+}
+```
diff --git a/packaging/npm/README.md b/packaging/npm/README.md
new file mode 100644
index 0000000..a1de7a2
--- /dev/null
+++ b/packaging/npm/README.md
@@ -0,0 +1,46 @@
+# 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.
+
+## Usage
+
+Copy the Qwen2.5 0.5B GGUF into app storage and pass that local path to the SDK
+facade.
+
+Browser / WASM:
+
+```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);
+```
+
+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;
+  },
+});
+```
+
+Source, crate documentation, and release notes live in the Edge Intelligence
+repository.

From 736945ea7eb430fc8480dabdf29361d3a86ba252 Mon Sep 17 00:00:00 2001
From: Tovli <Dekel@tovli.co.il>
Date: Tue, 23 Jun 2026 13:15:56 +0300
Subject: [PATCH 2/3] bump version to 0.3.10

---
 Cargo.toml | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/Cargo.toml b/Cargo.toml
index 2a2d296..2fa495e 100644
--- a/Cargo.toml
+++ b/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

From 07d3ab1df99c3b4159d6da714b869254c43b301b Mon Sep 17 00:00:00 2001
From: Tovli <Dekel@tovli.co.il>
Date: Tue, 23 Jun 2026 14:02:43 +0300
Subject: [PATCH 3/3] Gate binding artifact uploads on pull requests

---
 .github/workflows/bindings.yml                |  3 +++
 Cargo.lock                                    | 26 +++++++++----------
 ...gistry-release-ci-crates-io-npm-pub-dev.md |  7 +++--
 tests/check_ci_workflows.py                   | 10 +++++++
 4 files changed, 31 insertions(+), 15 deletions(-)

diff --git a/.github/workflows/bindings.yml b/.github/workflows/bindings.yml
index 0cad49f..fee636c 100644
--- a/.github/workflows/bindings.yml
+++ b/.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/
diff --git a/Cargo.lock b/Cargo.lock
index 0d95d38..bd18a65 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -793,7 +793,7 @@ checksum = "91622ff5e7162018101f2fea40d6ebf4a78bbe5a49736a2020649edf9693679e"
 
 [[package]]
 name = "el-bench"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
  "el-engine-candle",
@@ -803,7 +803,7 @@ dependencies = [
 
 [[package]]
 name = "el-chat"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
  "el-engine-candle",
@@ -811,7 +811,7 @@ dependencies = [
 
 [[package]]
 name = "el-cloud"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
  "reqwest",
@@ -822,11 +822,11 @@ dependencies = [
 
 [[package]]
 name = "el-core"
-version = "0.3.9"
+version = "0.3.10"
 
 [[package]]
 name = "el-engine-candle"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "candle-core",
  "candle-transformers",
@@ -838,7 +838,7 @@ dependencies = [
 
 [[package]]
 name = "el-ffi"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-cloud",
  "el-core",
@@ -852,7 +852,7 @@ dependencies = [
 
 [[package]]
 name = "el-grammar"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
  "el-provenance",
@@ -861,21 +861,21 @@ dependencies = [
 
 [[package]]
 name = "el-memory"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
 ]
 
 [[package]]
 name = "el-provenance"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
 ]
 
 [[package]]
 name = "el-provenance-ed25519"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "ed25519-dalek",
  "el-core",
@@ -884,7 +884,7 @@ dependencies = [
 
 [[package]]
 name = "el-runtime"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
  "el-memory",
@@ -894,14 +894,14 @@ dependencies = [
 
 [[package]]
 name = "el-safety"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
 ]
 
 [[package]]
 name = "el-telemetry"
-version = "0.3.9"
+version = "0.3.10"
 dependencies = [
  "el-core",
 ]
diff --git a/docs/adr/ADR-011-multi-registry-release-ci-crates-io-npm-pub-dev.md b/docs/adr/ADR-011-multi-registry-release-ci-crates-io-npm-pub-dev.md
index 8ab3343..d1ad819 100644
--- a/docs/adr/ADR-011-multi-registry-release-ci-crates-io-npm-pub-dev.md
+++ b/docs/adr/ADR-011-multi-registry-release-ci-crates-io-npm-pub-dev.md
@@ -66,8 +66,11 @@ Run `make codegen-rn` (Android + RN), `make build-ios`, `make build-wasm`,
 index propagation before the next dependent is submitted.
 
 The `bindings.yml` CI (ADR-011, triggered on every PR/push to affected paths)
-provides the pre-release confidence gate. `release.yml` only runs on explicit
-version tags.
+provides the pre-release confidence gate. Pull requests validate that binding
+codegen still succeeds without depending on nonessential artifact upload
+finalization; pushes to `master` retain the generated binding artifacts for
+inspection. `release.yml` only runs on explicit version tags and still uploads
+artifacts because later assembly jobs consume them.
 
 ## Consequences
 
diff --git a/tests/check_ci_workflows.py b/tests/check_ci_workflows.py
index 374ea28..df88ec1 100644
--- a/tests/check_ci_workflows.py
+++ b/tests/check_ci_workflows.py
@@ -12,6 +12,7 @@
 
 RN_RETRY_COMMAND = "bash scripts/retry-command.sh make codegen-rn"
 WASM_RETRY_COMMAND = "bash scripts/retry-command.sh curl -fsSL"
+BINDINGS_UPLOAD_IF = "if: github.event_name != 'pull_request'"
 
 REQUIRED_INSTALLS = [
     (
@@ -47,6 +48,15 @@ def check_workflow(path: Path) -> list[str]:
     if "Install wasm-pack" in text and WASM_RETRY_COMMAND not in text:
         errors.append(f"{path}: wasm-pack download must run through retry wrapper")
 
+    if path.name == "bindings.yml":
+        step_blocks = text.split("\n      - ")
+        for block in step_blocks:
+            if "uses: actions/upload-artifact@v4" in block and BINDINGS_UPLOAD_IF not in block:
+                errors.append(
+                    f"{path}: PR validation artifact uploads must be gated with "
+                    f"{BINDINGS_UPLOAD_IF!r}"
+                )
+
     return errors