diff --git a/Package.swift b/Package.swift
new file mode 100644
index 0000000..f48fa8f
--- /dev/null
+++ b/Package.swift
@@ -0,0 +1,38 @@
+// swift-tools-version: 6.2
+//
+// AIFabric — S2 LLM 패브릭 (Document Server 멀티디바이스 앱).
+//
+// 이 매니페스트는 `Sources/AI/` (S2 소유) 만 라이브러리 product 로 노출한다.
+// S3 앱(별 트랙)은 이 `AIFabric` product 에 **로컬 SwiftPM 의존**(소비모델 b)으로 붙어
+// `Sources/AI` 를 앱 타깃에 직접 포함하지 않는다 → 소스 이중소유/중복심볼 방지(plan S2-0b/S2-Fb).
+//
+// 경계: S2 는 이 Package.swift + Sources/AI/** + Tests/AITests/** + contract/fixtures/** 만 소유.
+//       앱 셸·뷰·구체 DS client = S3.
+import PackageDescription
+
+let package = Package(
+    name: "AIFabric",
+    platforms: [
+        .macOS(.v26),   // FoundationModels (OnDeviceProvider) = macOS 26 SDK
+    ],
+    products: [
+        .library(name: "AIFabric", targets: ["AIFabric"]),
+    ],
+    targets: [
+        .target(
+            name: "AIFabric",
+            path: "Sources/AI",
+            swiftSettings: [
+                .swiftLanguageMode(.v6),   // strict concurrency complete (동결 인터페이스 = swift6 PASS)
+            ]
+        ),
+        .testTarget(
+            name: "AITests",
+            dependencies: ["AIFabric"],
+            path: "Tests/AITests",
+            swiftSettings: [
+                .swiftLanguageMode(.v6),
+            ]
+        ),
+    ]
+)
diff --git a/Sources/AI/Composition.swift b/Sources/AI/Composition.swift
new file mode 100644
index 0000000..6330670
--- /dev/null
+++ b/Sources/AI/Composition.swift
@@ -0,0 +1,82 @@
+// Composition.swift — S2 → S3 통합 진입점 (소비모델 b) + 엔드포인트 단일소스 config.
+//
+// INTEGRATION (소비모델 b): 앱(S3)은 이 패키지의 `AIFabric` product 에 **로컬 SwiftPM 의존**으로 붙고,
+//   Sources/AI 를 앱 타깃에 직접 포함하지 않는다(소스 이중소유/중복심볼 방지). S3 는 makeDefaultRouter(...)
+//   하나로 실 라우터를 와이어링하고 MockAIProvider 를 대체한다. 구체 DSAskClient(HTTP) = S3 소유.
+//
+// 엔드포인트 단일소스(S2-Fa): raw URL 산재 금지 — 부주의한 편집의 침묵 엔드포인트 swap 방지
+// (2026-05-17 Hermes incident 선례). env override → 검증된 기본값. ([[feedback_hermes_config_single_source_envvar]])
+import Foundation
+import os
+
+public struct AIProviderConfiguration: Sendable {
+    /// 맥미니 llm-router base (trailing slash 없는 base; 경로는 provider 가 append).
+    public var localMLXBaseURL: URL
+    /// llm-router 모델 별칭(라이브 캡처로 확정 필요 — provisional 'gemma-macmini').
+    public var localMLXModel: String
+    /// DS API base — S3 의 DSAskClient 가 사용. 공인 https://document.hyungi.net/api · 내부 http://100.110.63.63:8000/api.
+    /// 주의: DS `/search/ask` 는 **trailing slash 필수**(경로 결합 시 S3 client 가 보장).
+    public var dsBaseURL: URL
+    public var requestTimeout: TimeInterval
+    public var probeTimeout: TimeInterval
+
+    public init(
+        localMLXBaseURL: URL,
+        localMLXModel: String = "gemma-macmini",
+        dsBaseURL: URL,
+        requestTimeout: TimeInterval = 60,
+        probeTimeout: TimeInterval = 2
+    ) {
+        self.localMLXBaseURL = localMLXBaseURL
+        self.localMLXModel = localMLXModel
+        self.dsBaseURL = dsBaseURL
+        self.requestTimeout = requestTimeout
+        self.probeTimeout = probeTimeout
+    }
+
+    /// 환경변수 override → 검증된 기본값(단일 source). 키 부재 시 기본값.
+    public static func resolved(
+        environment: [String: String] = ProcessInfo.processInfo.environment
+    ) -> AIProviderConfiguration {
+        let localMLX = environment["AIFABRIC_LOCALMLX_URL"].flatMap(URL.init(string:))
+            ?? URL(string: "http://100.76.254.116:8890")!
+        let model = environment["AIFABRIC_LOCALMLX_MODEL"] ?? "gemma-macmini"
+        let ds = environment["AIFABRIC_DS_URL"].flatMap(URL.init(string:))
+            ?? URL(string: "https://document.hyungi.net/api")!
+        return AIProviderConfiguration(localMLXBaseURL: localMLX, localMLXModel: model, dsBaseURL: ds)
+    }
+}
+
+/// 기본 OSLog 라우팅 훅 — 폴백/스킵을 가시화(silent 금지). S3 도 참조 가능(public).
+public enum AIFabricLog {
+    static let router = Logger(subsystem: "ds-app.AIFabric", category: "AIRouter")
+    public static let routerHook: @Sendable (String) -> Void = { msg in
+        router.info("\(msg, privacy: .public)")
+    }
+}
+
+/// S3 → S2 단일 진입점. 4 provider 전부 등록(vision 체인 가시 폴백 보장) + 기본 정책 + log 훅.
+/// - client: S3 가 주입하는 구체 DS ask client(HTTP).
+/// - config: 엔드포인트 단일소스(기본 = env override → 검증 기본값).
+/// - session: LocalMLX URLSession(기본 .shared; 테스트는 mock 주입).
+public func makeDefaultRouter(
+    client: DSAskClient,
+    config: AIProviderConfiguration = .resolved(),
+    session: URLSession = .shared,
+    policy: AIRoutingPolicy = .default,
+    log: @escaping @Sendable (String) -> Void = AIFabricLog.routerHook
+) -> AIRouter {
+    let providers: [AIProviderID: any AIProvider] = [
+        .remoteDS: RemoteDSProvider(client: client),
+        .localMLX: LocalMLXProvider(
+            baseURL: config.localMLXBaseURL,
+            model: config.localMLXModel,
+            session: session,
+            requestTimeout: config.requestTimeout,
+            probeTimeout: config.probeTimeout
+        ),
+        .onDevice: OnDeviceProvider(),
+        .specialized: SpecializedProvider(),   // scaffold(불가) — vision 폴백 가시화
+    ]
+    return AIRouter(providers: providers, policy: policy, log: log)
+}
diff --git a/Sources/AI/Providers/LocalMLXProvider.swift b/Sources/AI/Providers/LocalMLXProvider.swift
index 244959d..1f8c521 100644
--- a/Sources/AI/Providers/LocalMLXProvider.swift
+++ b/Sources/AI/Providers/LocalMLXProvider.swift
@@ -1,33 +1,161 @@
-// LocalMLXProvider.swift — S2 구현 스켈레톤 (맥미니 메인 로컬 LLM 허브).
+// LocalMLXProvider.swift — S2 구현 (맥미니 메인 로컬 LLM 허브).
 //
-// 실제 구현: 맥미니 Gemma 4 26B (MLX) OpenAI 호환 엔드포인트 호출.
-//   - 엔드포인트: llm-router :8890 (권장) 또는 MLX :8801 (Tailscale 100.76.254.116)
-//   - isAvailable = 짧은 health 핑 (도달 + 모델 로드)
-//   - complete   = POST /v1/chat/completions (messages: system/user 분리, call-shape 고정)
-// 인터페이스 동결 단계에서는 스텁.
-
+// 맥미니 Gemma 4 26B, llm-router :8890 (OpenAI 호환, wake-on-call). 결정 #4: raw MLX :8801 폐기.
+//   - isAvailable = GET /v1/models 경량 probe(짧은 timeout, wake 미트리거 → '가용인데 콜드' 정상)
+//   - complete   = POST /v1/chat/completions, messages system/user 분리(call-shape 고정)
+//
+// ⚠️ fixture(llm-router-chat.*.json) = PROVISIONAL_SYNTHETIC (맥미니 offline 으로 라이브 캡처 불가, 2026-06-04).
+//    OpenAI 표준 스펙 기반. 맥미니 복귀 시 S2-2a 라이브 재캡처로 교체(S2-Ff). 코드 자체는 라이브 동작.
 import Foundation
 
 public struct LocalMLXProvider: AIProvider {
     public let id: AIProviderID = .localMLX
 
-    /// 맥미니 허브 베이스 URL (S2 가 설정/Keychain 에서 주입).
+    /// 맥미니 허브 베이스 URL (S2-Fa config 에서 주입). trailing slash 없는 base, 경로는 appendingPathComponent.
     public let baseURL: URL
+    let model: String
+    let session: URLSession
+    let requestTimeout: TimeInterval
+    let probeTimeout: TimeInterval
 
-    public init(baseURL: URL) {
+    public init(
+        baseURL: URL,
+        model: String = "gemma-macmini",
+        session: URLSession = .shared,
+        requestTimeout: TimeInterval = 60,
+        probeTimeout: TimeInterval = 2
+    ) {
         self.baseURL = baseURL
+        self.model = model
+        self.session = session
+        self.requestTimeout = requestTimeout
+        self.probeTimeout = probeTimeout
     }
 
+    // MARK: isAvailable — 경량 health probe (wake 미트리거)
+
     public var isAvailable: Bool {
         get async {
-            // S2: GET /v1/models 또는 경량 health 핑으로 교체.
-            false
+            var req = URLRequest(url: baseURL.appendingPathComponent("v1/models"))
+            req.httpMethod = "GET"
+            req.timeoutInterval = probeTimeout
+            do {
+                let (_, resp) = try await session.data(for: req)
+                guard let http = resp as? HTTPURLResponse else { return false }
+                return (200..<300).contains(http.statusCode)
+            } catch {
+                // timeout/연결오류 → false(throw 아님). probe 는 wake/모델로드 유발 X.
+                // '가용=응답가능 ≠ 즉답' — 콜드 모델은 첫 complete() 가 로드지연 흡수.
+                return false
+            }
         }
     }
 
+    // MARK: complete — OpenAI 호환 chat/completions
+
     public func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
-        // S2: OpenAI 호환 chat/completions 호출 → AICompletionResponse(providerUsed: .localMLX).
-        // messages 구조(system/user 분리)는 production 호출과 단일 source-of-truth.
-        throw AIProviderError.notImplemented(id)
+        try Task.checkCancellation()
+
+        var req = URLRequest(url: baseURL.appendingPathComponent("v1/chat/completions"))
+        req.httpMethod = "POST"
+        req.setValue("application/json", forHTTPHeaderField: "Content-Type")
+        req.timeoutInterval = requestTimeout          // S2-Fe: 행 걸린 생성 무한정지 방지
+        req.httpBody = try Self.encodeRequest(request, model: model)
+
+        let started = Date()
+        let data: Data
+        let resp: URLResponse
+        do {
+            // URLSession async 는 Task 취소를 자동 honor → CancellationError 전파(S2-Fe).
+            (data, resp) = try await session.data(for: req)
+        } catch let e as URLError where e.code == .timedOut {
+            throw AIProviderError.backendError(id, status: -1, reason: "request timed out after \(Int(requestTimeout))s")
+        }
+
+        guard let http = resp as? HTTPURLResponse else {
+            throw AIProviderError.backendError(id, status: -1, reason: "non-HTTP response")
+        }
+        guard (200..<300).contains(http.statusCode) else {
+            // non-200 → backendError (빈 text 침묵 금지).
+            let reason = String(data: data, encoding: .utf8).map { String($0.prefix(300)) }
+            throw AIProviderError.backendError(id, status: http.statusCode, reason: reason)
+        }
+
+        let decoded = try JSONDecoder().decode(OpenAIChatResponse.self, from: data)
+        guard let choice = decoded.choices.first else {
+            throw AIProviderError.backendError(id, status: http.statusCode, reason: "no choices in response")
+        }
+        return AICompletionResponse(
+            text: choice.message.content,
+            providerUsed: .localMLX,
+            finishReason: Self.finishReason(choice.finishReason),
+            citations: [],                                  // 로컬 생성 — 코퍼스 인용 없음
+            confidence: nil,
+            latencyMs: Date().timeIntervalSince(started) * 1000,
+            routingNote: nil                                // fallback note 는 라우터가 채움
+        )
+    }
+
+    // MARK: 매핑
+
+    static func finishReason(_ openAI: String?) -> AIFinishReason {
+        switch openAI {
+        case "stop":   return .completed
+        case "length": return .completed   // max_tokens 절단 — 정상 완료로 취급
+        default:       return .completed
+        }
+    }
+
+    /// AICompletionRequest → OpenAI chat/completions body. messages system/user 분리(fixture 와 단일 source-of-truth).
+    /// system.content = systemPrompt ?? "" (plan S2-2c). temperature 는 AICompletionRequest 에 없음(동결) → 미설정(서버 기본).
+    static func encodeRequest(_ request: AICompletionRequest, model: String) throws -> Data {
+        let body = OpenAIChatRequest(
+            model: model,
+            messages: [
+                OpenAIChatRequest.Message(role: "system", content: request.systemPrompt ?? ""),
+                OpenAIChatRequest.Message(role: "user", content: request.prompt),
+            ],
+            maxTokens: request.maxTokens,
+            stream: false
+        )
+        let enc = JSONEncoder()
+        enc.outputFormatting = [.sortedKeys]
+        return try enc.encode(body)
     }
 }
+
+// MARK: - OpenAI 호환 wire 타입 (내부)
+
+struct OpenAIChatRequest: Encodable, Sendable {
+    struct Message: Encodable, Sendable {
+        let role: String
+        let content: String
+    }
+    let model: String
+    let messages: [Message]
+    let maxTokens: Int?
+    let stream: Bool
+
+    enum CodingKeys: String, CodingKey {
+        case model, messages, stream
+        case maxTokens = "max_tokens"
+    }
+}
+
+struct OpenAIChatResponse: Decodable, Sendable {
+    struct Choice: Decodable, Sendable {
+        struct Message: Decodable, Sendable {
+            let role: String
+            let content: String
+        }
+        let index: Int?
+        let message: Message
+        let finishReason: String?
+
+        enum CodingKeys: String, CodingKey {
+            case index, message
+            case finishReason = "finish_reason"
+        }
+    }
+    let choices: [Choice]
+}
diff --git a/Sources/AI/Providers/OnDeviceProvider.swift b/Sources/AI/Providers/OnDeviceProvider.swift
index bf0a4d5..5b21885 100644
--- a/Sources/AI/Providers/OnDeviceProvider.swift
+++ b/Sources/AI/Providers/OnDeviceProvider.swift
@@ -1,26 +1,196 @@
-// OnDeviceProvider.swift — S2 구현 스켈레톤 (맥북·아이폰 온디바이스).
+// OnDeviceProvider.swift — S2 구현 (맥북·아이폰 온디바이스, Apple FoundationModels).
 //
-// 실제 구현: `import FoundationModels` 후 SystemLanguageModel / LanguageModelSession 사용.
-//   - isAvailable = SystemLanguageModel.default.availability == .available
-//   - complete   = LanguageModelSession 으로 prompt 응답 (citations 없음 — 로컬 생성)
-// 인터페이스 동결 단계에서는 Foundation-only 스텁(notImplemented).
-
+// 라이브 경로(결정 #3): SystemLanguageModel.default.availability(동기) + LanguageModelSession.respond().
+// 모델 ~3B/2-bit QAT → quickSummarize/classify 적합, corpusAsk 무리(라우팅이 corpusAsk 를 RemoteDS 로만 보냄).
+//
+// 테스트 가능성: 모델 backend 를 seam(OnDeviceModelBackend)으로 주입 → CI/비-AI Mac 은 mock 으로
+// 에러매핑/라우팅 단위테스트, M5 Max 는 라이브 통합테스트(아래 default = FoundationModels).
+//
+// 실측 동결(S2-3a, M5 Max 2026-06-04): availability=available · respond()→Response<String>.content ·
+// GenerationError 9 case · 컨텍스트 오버플로→exceededContextWindowSize · 취소=COOPERATIVE(respond() 가
+// mid-flight Task.cancel 협조 → checkCancellation 은 belt-and-suspenders, streamResponse 불필요).
 import Foundation
+import os
+
+// MARK: - 중립 타입 (FoundationModels 비의존 — 매핑 로직을 HW 없이 단위테스트)
+
+public enum OnDeviceAvailability: Sendable, Equatable {
+    case available
+    case unavailable(reason: String)
+}
+
+/// FoundationModels.LanguageModelSession.GenerationError 의 중립 미러(9 case + unknown).
+/// 라이브 backend 가 SDK 에러를 이걸로 번역 → provider 가 finishReason/throw 로 매핑(순수·테스트 가능).
+public enum OnDeviceGenerationError: Error, Sendable, Equatable {
+    case guardrailViolation
+    case refusal
+    case exceededContextWindowSize
+    case rateLimited
+    case concurrentRequests
+    case unsupportedLanguageOrLocale
+    case unsupportedGuide
+    case decodingFailure
+    case assetsUnavailable
+    case unknown(String)
+}
+
+/// 온디바이스 모델 backend seam. 라이브 = FoundationModelsBackend, 테스트 = mock.
+protocol OnDeviceModelBackend: Sendable {
+    var availability: OnDeviceAvailability { get }
+    /// 실패는 OnDeviceGenerationError 또는 CancellationError 로 throw.
+    func generate(prompt: String, systemPrompt: String?, maxTokens: Int?) async throws -> String
+}
+
+// MARK: - Provider
 
 public struct OnDeviceProvider: AIProvider {
     public let id: AIProviderID = .onDevice
+    private let backend: OnDeviceModelBackend
+    private let log: @Sendable (String) -> Void
 
-    public init() {}
+    public init() {
+        self.backend = Self.makeLiveBackend()
+        let logger = Logger(subsystem: "ds-app.AIFabric", category: "OnDeviceProvider")
+        self.log = { msg in logger.warning("\(msg, privacy: .public)") }
+    }
+
+    /// 테스트 seam — backend/log 주입(HW 비의존 단위테스트).
+    init(backend: OnDeviceModelBackend, log: @escaping @Sendable (String) -> Void = { _ in }) {
+        self.backend = backend
+        self.log = log
+    }
 
     public var isAvailable: Bool {
-        get async {
-            // S2: FoundationModels 가용성 프로브로 교체.
-            false
-        }
+        get async { backend.availability == .available }
     }
 
     public func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
-        // S2: LanguageModelSession(.default) 호출 → AICompletionResponse(providerUsed: .onDevice).
-        throw AIProviderError.notImplemented(id)
+        // belt-and-suspenders: 실제 mid-flight 취소는 respond() 내부가 협조(S2-3a 실측).
+        try Task.checkCancellation()
+        let started = Date()
+        do {
+            let text = try await backend.generate(
+                prompt: request.prompt,
+                systemPrompt: request.systemPrompt,
+                maxTokens: request.maxTokens
+            )
+            return AICompletionResponse(
+                text: text,
+                providerUsed: .onDevice,
+                finishReason: .completed,
+                latencyMs: Date().timeIntervalSince(started) * 1000
+            )
+        } catch let e as OnDeviceGenerationError {
+            return try mapError(e)
+        }
+        // CancellationError 등은 자연 전파.
+    }
+
+    /// GenerationError 매핑(S2-3c). 거부 = 답변의 일종 → 반환(폴백 X). 그 외 = provider 불가 → throw(라우터가 가시 폴백).
+    private func mapError(_ e: OnDeviceGenerationError) throws -> AICompletionResponse {
+        switch e {
+        case .guardrailViolation, .refusal:
+            return AICompletionResponse(
+                text: "",
+                providerUsed: .onDevice,
+                finishReason: .refused,
+                routingNote: "on-device refused (guardrail/refusal)"
+            )
+        case .rateLimited:
+            // stateless-per-request 인데 발생 = 세션 재사용 버그 신호(가림 금지 → loud + 불가).
+            log("UNEXPECTED onDevice rateLimited on stateless session — 세션 재사용 버그 의심")
+            throw AIProviderError.unavailable(id)
+        case .concurrentRequests:
+            log("UNEXPECTED onDevice concurrentRequests on stateless session — 세션 재사용 버그 의심")
+            throw AIProviderError.unavailable(id)
+        case .exceededContextWindowSize:
+            log("onDevice context window(4096) exceeded — 라우터가 localMLX 로 폴백")
+            throw AIProviderError.unavailable(id)
+        case .unsupportedLanguageOrLocale:
+            log("onDevice unsupported language/locale — 폴백")
+            throw AIProviderError.unavailable(id)
+        case .unsupportedGuide, .decodingFailure, .assetsUnavailable:
+            throw AIProviderError.unavailable(id)
+        case .unknown(let detail):
+            log("onDevice unknown generation error: \(detail)")
+            throw AIProviderError.unavailable(id)
+        }
+    }
+
+    static func makeLiveBackend() -> OnDeviceModelBackend {
+        #if canImport(FoundationModels)
+        return FoundationModelsBackend()
+        #else
+        return UnavailableBackend(reason: "FoundationModels not importable on this platform")
+        #endif
     }
 }
+
+/// FoundationModels 미가용 플랫폼/SDK 폴백.
+struct UnavailableBackend: OnDeviceModelBackend {
+    let reason: String
+    var availability: OnDeviceAvailability { .unavailable(reason: reason) }
+    func generate(prompt: String, systemPrompt: String?, maxTokens: Int?) async throws -> String {
+        throw OnDeviceGenerationError.unknown("backend unavailable: \(reason)")
+    }
+}
+
+// MARK: - 라이브 FoundationModels backend (M5 Max / Apple Intelligence)
+
+#if canImport(FoundationModels)
+import FoundationModels
+
+struct FoundationModelsBackend: OnDeviceModelBackend {
+
+    var availability: OnDeviceAvailability {
+        switch SystemLanguageModel.default.availability {
+        case .available:
+            return .available
+        case .unavailable(let reason):
+            switch reason {
+            case .deviceNotEligible:           return .unavailable(reason: "deviceNotEligible")
+            case .appleIntelligenceNotEnabled: return .unavailable(reason: "appleIntelligenceNotEnabled")
+            case .modelNotReady:               return .unavailable(reason: "modelNotReady")
+            @unknown default:                  return .unavailable(reason: "unknownReason")
+            }
+        @unknown default:
+            return .unavailable(reason: "unknown")
+        }
+    }
+
+    func generate(prompt: String, systemPrompt: String?, maxTokens: Int?) async throws -> String {
+        // instructions = init 시점. systemPrompt nil 이면 생략(빈 문자열 금지 — S2-3c, LocalMLX 의 ?? '' 와 구분).
+        let session: LanguageModelSession
+        if let systemPrompt {
+            session = LanguageModelSession(model: .default, instructions: systemPrompt)
+        } else {
+            session = LanguageModelSession(model: .default)
+        }
+        session.prewarm()   // 동기 반환(~1.3ms), 백그라운드 워밍(S2-3a)
+        // temperature 는 AICompletionRequest 에 없음(동결) → 미설정(모델 기본). LocalMLX 와 동일 정책.
+        let options = GenerationOptions(maximumResponseTokens: maxTokens)
+        do {
+            let response = try await session.respond(to: prompt, options: options)
+            return response.content   // Response<String>.content : String
+        } catch let g as LanguageModelSession.GenerationError {
+            throw Self.translate(g)
+        }
+    }
+
+    /// SDK GenerationError(9 case) → 중립 OnDeviceGenerationError. exhaustive + @unknown default.
+    static func translate(_ g: LanguageModelSession.GenerationError) -> OnDeviceGenerationError {
+        switch g {
+        case .guardrailViolation:          return .guardrailViolation
+        case .refusal:                     return .refusal
+        case .exceededContextWindowSize:   return .exceededContextWindowSize
+        case .rateLimited:                 return .rateLimited
+        case .concurrentRequests:          return .concurrentRequests
+        case .unsupportedLanguageOrLocale: return .unsupportedLanguageOrLocale
+        case .unsupportedGuide:            return .unsupportedGuide
+        case .decodingFailure:             return .decodingFailure
+        case .assetsUnavailable:           return .assetsUnavailable
+        @unknown default:                  return .unknown("\(g)")
+        }
+    }
+}
+#endif
diff --git a/Sources/AI/Providers/RemoteDSProvider.swift b/Sources/AI/Providers/RemoteDSProvider.swift
index dcda7d5..1744e8d 100644
--- a/Sources/AI/Providers/RemoteDSProvider.swift
+++ b/Sources/AI/Providers/RemoteDSProvider.swift
@@ -1,22 +1,99 @@
-// RemoteDSProvider.swift — S2 구현 스켈레톤 (원격 DS 코퍼스 RAG).
+// RemoteDSProvider.swift — S2 구현 (원격 DS 코퍼스 RAG).
 //
-// 이 provider 가 S1 계약과 만나는 다리:
-//   complete(corpusAsk) → GET /search/ask?q=&backend=  (CONTRACT.md §4, AskResponse)
-//   AskResponse.citations → [AICitation] 매핑
+// S1 계약과 만나는 다리 (CONTRACT.md §4 / AI-ROUTING.md §4):
+//   complete(corpusAsk) → DSAskClient.ask(query:backend:) → AskResponse → AICompletionResponse
+//   AskResponse.ai_answer        → text
+//   AskResponse.citations[]      → [AICitation]
 //   AskResponse.synthesis_status → AIFinishReason
-//   AskResponse.backend_used → routingNote (어느 LLM 이 응답했는지)
-// backend 인자: nil(=mac-mini-default) 또는 explicitProvider 매핑(localMLX→gemma-macmini 등).
-// 인터페이스 동결 단계에서는 스텁(S3 의 DS API client 주입 후 S2 가 결선).
-
+//   AskResponse.confidence       → AIConfidence
+//   AskResponse.backend_used     → routingNote (어느 LLM 이 응답했는지)
+//
+// HTTP 는 S3 의 구체 client(LiveDSClient)가 소유 — S2 는 DSAskClient 프로토콜 seam + 매핑만.
+// 인터페이스 동결: AIProvider 프로토콜은 불변. RemoteDSProvider.init(client:) 은 S2 가 채우는 구현부.
 import Foundation
 
+// MARK: - S2 가 소유하는 DS ask seam (구체 impl = S3)
+
+/// DS `GET /search/ask?q=&backend=` 호출 추상화. S3 의 LiveDSClient 가 conform,
+/// S2 는 mock 으로 단위테스트(라이브 네트워크 0). HTTP 실패는 conformer 가 throw
+/// (권장: `AIProviderError.backendError(.remoteDS, status:, reason:)`) — 침묵 폴백 금지.
+public protocol DSAskClient: Sendable {
+    func ask(query: String, backend: String) async throws -> AskResponse
+}
+
+// MARK: - DS /search/ask 응답 (부분 미러, 디코딩 전용)
+//
+// 명시 CodingKeys — convertFromSnakeCase 금지(S3 모델 규약과 일관). fixture: contract/fixtures/ask.json.
+
+public struct AskResponse: Decodable, Sendable {
+    public let aiAnswer: String
+    public let citations: [AskCitation]
+    public let synthesisStatus: String
+    public let synthesisMs: Double?
+    public let confidence: String?
+    public let backendUsed: String?
+    public let refused: Bool?
+
+    enum CodingKeys: String, CodingKey {
+        case aiAnswer = "ai_answer"
+        case citations
+        case synthesisStatus = "synthesis_status"
+        case synthesisMs = "synthesis_ms"
+        case confidence
+        case backendUsed = "backend_used"
+        case refused
+    }
+
+    public init(aiAnswer: String, citations: [AskCitation], synthesisStatus: String,
+                synthesisMs: Double? = nil, confidence: String? = nil,
+                backendUsed: String? = nil, refused: Bool? = nil) {
+        self.aiAnswer = aiAnswer
+        self.citations = citations
+        self.synthesisStatus = synthesisStatus
+        self.synthesisMs = synthesisMs
+        self.confidence = confidence
+        self.backendUsed = backendUsed
+        self.refused = refused
+    }
+}
+
+public struct AskCitation: Decodable, Sendable {
+    public let n: Int
+    public let docId: Int
+    public let title: String?
+    public let sectionTitle: String?
+    public let spanText: String
+
+    enum CodingKeys: String, CodingKey {
+        case n
+        case docId = "doc_id"
+        case title
+        case sectionTitle = "section_title"
+        case spanText = "span_text"
+    }
+
+    public init(n: Int, docId: Int, title: String?, sectionTitle: String?, spanText: String) {
+        self.n = n
+        self.docId = docId
+        self.title = title
+        self.sectionTitle = sectionTitle
+        self.spanText = spanText
+    }
+}
+
+// MARK: - Provider
+
 public struct RemoteDSProvider: AIProvider {
     public let id: AIProviderID = .remoteDS
+    private let client: DSAskClient
 
-    public init() {}
+    public init(client: DSAskClient) {
+        self.client = client
+    }
 
+    /// 원격 코퍼스는 항상 후보(라우팅 시맨틱). 실제 도달 실패는 complete 에서 표면화.
     public var isAvailable: Bool {
-        get async { true }   // 원격 코퍼스는 항상 후보 (실제 도달 실패는 complete 에서 backendError).
+        get async { true }
     }
 
     public func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
@@ -24,14 +101,31 @@ public struct RemoteDSProvider: AIProvider {
         guard request.task == .corpusAsk else {
             throw AIProviderError.notImplemented(id)
         }
-        // S2: DS API client.ask(q:) 호출 → AskResponse 디코딩 → 아래 매핑.
-        //   let r = try await dsClient.ask(q: request.prompt, backend: mappedBackend(request.explicitProvider))
-        //   return Self.map(r)
-        throw AIProviderError.notImplemented(id)
+        try Task.checkCancellation()
+        let backend = Self.dsBackend(for: request.explicitProvider)
+        // HTTP 실패(503 등)는 client 가 throw → 그대로 전파(자동 로컬 폴백 금지).
+        let response = try await client.ask(query: request.prompt, backend: backend)
+        return Self.map(response)
+    }
+
+    // MARK: 매핑 (AI-ROUTING.md §4, 고정)
+
+    static func map(_ r: AskResponse) -> AICompletionResponse {
+        let citations = r.citations.map {
+            AICitation(n: $0.n, docId: $0.docId, title: $0.title,
+                       sectionTitle: $0.sectionTitle, spanText: $0.spanText)
+        }
+        return AICompletionResponse(
+            text: r.aiAnswer,
+            providerUsed: .remoteDS,
+            finishReason: finishReason(fromSynthesisStatus: r.synthesisStatus),
+            citations: citations,
+            confidence: r.confidence.flatMap(AIConfidence.init(rawValue:)),
+            latencyMs: r.synthesisMs,            // latency 는 synthesis_ms 만 기록(하드 게이트 없음)
+            routingNote: r.backendUsed           // 어느 LLM 이 응답했는지
+        )
     }
 
-    /// AskResponse(JSON) → AICompletionResponse 매핑 규칙(고정). S2 가 이 형태로 결선.
-    /// 시그니처만 동결 — 실제 호출은 S3 DS client 와 결합.
     static func finishReason(fromSynthesisStatus status: String) -> AIFinishReason {
         switch status {
         case "completed": return .completed
@@ -41,4 +135,21 @@ public struct RemoteDSProvider: AIProvider {
         default: return .refused
         }
     }
+
+    /// explicitProvider → DS 합성 backend (AI-ROUTING.md §4, 고정).
+    /// **dict 아닌 exhaustive switch** — 미래 AIProviderID 추가 시 컴파일러가 backend 결정을 강제
+    /// (미매핑 provider → nil → 미정의 backend → 404 침묵실패를 컴파일 타임에 차단).
+    static func dsBackend(for explicit: AIProviderID?) -> String {
+        guard let explicit else { return "mac-mini-default" }   // 미지정 → DS 기본
+        switch explicit {
+        case .localMLX:    return "gemma-macmini"
+        case .remoteDS:    return "mac-mini-default"   // 명시 remoteDS = DS 기본 합성
+        case .onDevice:    return "mac-mini-default"   // onDevice 는 코퍼스 합성 불가 → DS 기본
+        case .specialized: return "mac-mini-default"   // specialized 코퍼스 backend 없음 → DS 기본
+        }
+        // TODO(qwen-macbook): 현재 어떤 AIProviderID 도 'qwen-macbook'(M5 Max Qwen VLM) 로 매핑 안 됨.
+        //   해당 provider case 가 생기면 위 exhaustive switch 가 컴파일 실패 → backend 결정 강제(S2-1b 게이트 b).
+        // TODO(claude-cloud): cloud backend = 'claude-cloud' 는 DS 가 503(scaffold, S2-4b). 매핑하는 case 없음.
+        //   503 은 client 가 backendError(.remoteDS, status:503, …) 로 표면화 — 절대 로컬 침묵 폴백 X(과금 버킷 분리).
+    }
 }
diff --git a/Sources/AI/Providers/SpecializedProvider.swift b/Sources/AI/Providers/SpecializedProvider.swift
new file mode 100644
index 0000000..0d11025
--- /dev/null
+++ b/Sources/AI/Providers/SpecializedProvider.swift
@@ -0,0 +1,27 @@
+// SpecializedProvider.swift — S2 scaffold (GPU 특화 통로: rerank / embed / vision / OCR).
+//
+// 이번 PR = **scaffold-only**: isAvailable=false, complete→notImplemented(.specialized). HTTP client/API key/cost = 0.
+// 목적: 라우터 .vision 체인 [.specialized, .onDevice] 에서 specialized 가 **dict 에 부재 → 침묵 continue** 하던 것을,
+//       등록된 명시-불가 provider 로 바꿔 폴백을 **가시화**(라우터가 'specialized unavailable → onDevice' log).
+//
+// 실 vision 통합 = backend 확정 + fixture 캡처 후 **별 PR**(결정 #1):
+//   후보 = MacBook M5 Max Qwen VLM  http://100.118.112.84:8810 (OpenAI 호환, wake-on-call) — 패브릭 내 유일 생성형 VLM.
+//   GPU Ollama 는 embedding(bge-m3) 전용이라 생성형 vision 아님.
+import Foundation
+
+public struct SpecializedProvider: AIProvider {
+    public let id: AIProviderID = .specialized
+
+    public init() {}
+
+    /// scaffold — 미구현. 라우터가 가시 폴백하도록 정직하게 false.
+    public var isAvailable: Bool {
+        get async { false }
+    }
+
+    public func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
+        // 침묵 nil-skip 금지 — 명시 notImplemented 로 vision-chain 폴백을 가시화.
+        // TODO(vision): .specialized → M5 Max Qwen VLM(:8810, wake-on-call) 실 call-shape + fixture 후 별 PR.
+        throw AIProviderError.notImplemented(id)
+    }
+}
diff --git a/Tests/AITests/AIRouterSmokeTests.swift b/Tests/AITests/AIRouterSmokeTests.swift
new file mode 100644
index 0000000..a65f6da
--- /dev/null
+++ b/Tests/AITests/AIRouterSmokeTests.swift
@@ -0,0 +1,84 @@
+import XCTest
+@testable import AIFabric
+
+/// S2-Fc — AI-ROUTING.md §3 스모크 매트릭스를 mock provider(라이브 net 0)로 검증.
+/// 정성 게이트(올바른 라우팅 시맨틱 + 정확한 에러 enum + 폴백 가시성). latency 는 기록만.
+final class AIRouterSmokeTests: XCTestCase {
+
+    private func askFixture() throws -> AskResponse {
+        try Fixture.decode(AskResponse.self, from: "ask.json")
+    }
+
+    /// 모든 태스크가 1순위 성공하는 정상 패브릭.
+    private func healthyRouter(log: @escaping @Sendable (String) -> Void = { _ in }) throws -> AIRouter {
+        AIRouter(providers: [
+            .onDevice: EchoProvider(id: .onDevice),
+            .localMLX: EchoProvider(id: .localMLX),
+            .remoteDS: RemoteDSProvider(client: MockDSAskClient(response: try askFixture())),
+            .specialized: SpecializedProvider(),   // 불가 scaffold
+        ], log: log)
+    }
+
+    func testQuickSummarizeToOnDevice() async throws {
+        let resp = try await healthyRouter().route(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .onDevice)
+        XCTAssertNil(resp.routingNote)
+    }
+
+    func testCorpusAskToRemoteDSWithCitations() async throws {
+        let resp = try await healthyRouter().route(AICompletionRequest(task: .corpusAsk, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .remoteDS)
+        XCTAssertEqual(resp.citations.count, 1)
+    }
+
+    func testClassifyToLocalMLX() async throws {
+        let resp = try await healthyRouter().route(AICompletionRequest(task: .classify, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .localMLX)
+        XCTAssertNil(resp.routingNote)
+    }
+
+    func testVisionSpecializedUnavailableFallsToOnDeviceVisibly() async throws {
+        let sink = LogSink()
+        let resp = try await healthyRouter(log: { sink.append($0) })
+            .route(AICompletionRequest(task: .vision, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .onDevice)
+        XCTAssertEqual(resp.routingNote, "fallback from specialized → onDevice")
+        XCTAssertTrue(sink.lines.contains { $0.contains("specialized") && $0.contains("unavailable") })
+    }
+
+    func testExplicitOnDeviceUnavailableErrorsNoFallback() async throws {
+        let router = AIRouter(providers: [
+            .onDevice: EchoProvider(id: .onDevice, available: false),
+            .localMLX: EchoProvider(id: .localMLX, available: true),
+        ])
+        do {
+            _ = try await router.route(AICompletionRequest(task: .quickSummarize, prompt: "p", explicitProvider: .onDevice))
+            XCTFail("explicit onDevice 불가 → 에러(자동 fallback X)")
+        } catch let AIRoutingError.explicitProviderUnavailable(id) {
+            XCTAssertEqual(id, .onDevice)
+        }
+    }
+
+    func testRuleFallbackOnDeviceDownToLocalMLXWithNote() async throws {
+        let router = AIRouter(providers: [
+            .onDevice: EchoProvider(id: .onDevice, available: false),
+            .localMLX: EchoProvider(id: .localMLX, available: true),
+        ])
+        let resp = try await router.route(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .localMLX)
+        XCTAssertEqual(resp.routingNote, "fallback from onDevice → localMLX")
+    }
+
+    func testAllUnavailableYieldsNoProviderAvailable() async throws {
+        let router = AIRouter(providers: [
+            .onDevice: EchoProvider(id: .onDevice, available: false),
+            .localMLX: EchoProvider(id: .localMLX, available: false),
+        ])
+        do {
+            _ = try await router.route(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+            XCTFail("전부 불가 → noProviderAvailable")
+        } catch let AIRoutingError.noProviderAvailable(task) {
+            XCTAssertEqual(task, .quickSummarize)
+        }
+    }
+}
diff --git a/Tests/AITests/CompositionTests.swift b/Tests/AITests/CompositionTests.swift
new file mode 100644
index 0000000..cb2bf2b
--- /dev/null
+++ b/Tests/AITests/CompositionTests.swift
@@ -0,0 +1,66 @@
+import XCTest
+@testable import AIFabric
+
+/// S2-Fa(config 단일소스) + S2-Fb(컴포지션 루트) + S2-Fe(타임아웃/취소 전파).
+final class CompositionTests: XCTestCase {
+
+    // MARK: S2-Fa — 엔드포인트 단일소스 config
+
+    func testConfigDefaults() {
+        let c = AIProviderConfiguration.resolved(environment: [:])
+        XCTAssertEqual(c.localMLXBaseURL.absoluteString, "http://100.76.254.116:8890")
+        XCTAssertEqual(c.localMLXModel, "gemma-macmini")
+        XCTAssertEqual(c.dsBaseURL.absoluteString, "https://document.hyungi.net/api")
+        XCTAssertEqual(c.probeTimeout, 2)
+    }
+
+    func testConfigEnvOverride() {
+        let c = AIProviderConfiguration.resolved(environment: [
+            "AIFABRIC_LOCALMLX_URL": "http://127.0.0.1:9999",
+            "AIFABRIC_LOCALMLX_MODEL": "test-model",
+            "AIFABRIC_DS_URL": "http://100.110.63.63:8000/api",
+        ])
+        XCTAssertEqual(c.localMLXBaseURL.absoluteString, "http://127.0.0.1:9999")
+        XCTAssertEqual(c.localMLXModel, "test-model")
+        XCTAssertEqual(c.dsBaseURL.absoluteString, "http://100.110.63.63:8000/api")
+    }
+
+    // MARK: S2-Fb — 컴포지션 루트 (4 provider 전부 등록)
+
+    func testMakeDefaultRouterRegistersAllFour() async throws {
+        let client = MockDSAskClient(response: try Fixture.decode(AskResponse.self, from: "ask.json"))
+        let router = makeDefaultRouter(client: client, session: MockURLProtocol.session(), log: { _ in })
+        XCTAssertEqual(Set(router.providers.keys), Set(AIProviderID.allCases))
+        // corpusAsk 는 와이어링된 RemoteDS 로 흘러 citations 매핑.
+        let resp = try await router.route(AICompletionRequest(task: .corpusAsk, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .remoteDS)
+        XCTAssertEqual(resp.citations.count, 1)
+    }
+
+    func testMakeDefaultRouterVisionFallbackVisible() async throws {
+        let client = MockDSAskClient(response: try Fixture.decode(AskResponse.self, from: "ask.json"))
+        let sink = LogSink()
+        let router = makeDefaultRouter(client: client, session: MockURLProtocol.session(), log: { sink.append($0) })
+        // specialized scaffold(불가) → onDevice. (onDevice 가용 여부는 머신 의존 — 최소한 specialized 불가 log 는 떠야)
+        _ = try? await router.route(AICompletionRequest(task: .vision, prompt: "p"))
+        XCTAssertTrue(sink.lines.contains { $0.contains("specialized") && $0.contains("unavailable") },
+                      "specialized 불가가 침묵 아닌 log 로 가시화")
+    }
+
+    // MARK: S2-Fe — 취소 전파 (URLSession 경로 아날로그)
+
+    func testCancellationPropagatesThroughRouter() async throws {
+        let router = AIRouter(providers: [.localMLX: SleepingProvider(id: .localMLX)])
+        let task = Task {
+            try await router.route(AICompletionRequest(task: .classify, prompt: "p"))
+        }
+        try? await Task.sleep(nanoseconds: 100_000_000)
+        task.cancel()
+        do {
+            _ = try await task.value
+            XCTFail("취소된 생성은 CancellationError 전파")
+        } catch is CancellationError {
+            // 기대: URLSession async/Task.sleep 둘 다 취소 자동 honor. OnDevice respond() 도 협조적(S2-3a).
+        }
+    }
+}
diff --git a/Tests/AITests/FixtureSupport.swift b/Tests/AITests/FixtureSupport.swift
new file mode 100644
index 0000000..6bae563
--- /dev/null
+++ b/Tests/AITests/FixtureSupport.swift
@@ -0,0 +1,26 @@
+// FixtureSupport.swift — canonical fixture 로더 (contract/fixtures/ 를 #filePath 기준으로 직접 읽음).
+//
+// 픽스처는 repo 루트의 `contract/fixtures/` 에 단일 소유(S1 ask.json 등 + S2 가 추가하는
+// foundationmodels-respond / llm-router-chat). 테스트 타깃 안에 복제하면 드리프트가 생기므로
+// 복제 대신 #filePath 에서 repo 루트를 계산해 canonical 파일을 직접 로드한다.
+import Foundation
+
+enum Fixture {
+    /// repo 루트(.../ds-app-s2) — 이 파일은 <root>/Tests/AITests/FixtureSupport.swift.
+    static let repoRoot: URL = URL(fileURLWithPath: #filePath)
+        .deletingLastPathComponent()   // Tests/AITests
+        .deletingLastPathComponent()   // Tests
+        .deletingLastPathComponent()   // <root>
+
+    static func url(_ name: String) -> URL {
+        repoRoot.appendingPathComponent("contract/fixtures").appendingPathComponent(name)
+    }
+
+    static func data(_ name: String) throws -> Data {
+        try Data(contentsOf: url(name))
+    }
+
+    static func decode<T: Decodable>(_ type: T.Type, from name: String, using decoder: JSONDecoder = JSONDecoder()) throws -> T {
+        try decoder.decode(type, from: data(name))
+    }
+}
diff --git a/Tests/AITests/HarnessSmokeTests.swift b/Tests/AITests/HarnessSmokeTests.swift
new file mode 100644
index 0000000..d592970
--- /dev/null
+++ b/Tests/AITests/HarnessSmokeTests.swift
@@ -0,0 +1,18 @@
+import XCTest
+@testable import AIFabric
+
+/// Phase 0 하니스 스모크 — Sources/AI 단독 빌드/테스트가 도는지, 동결 타입이 보이는지만 확인.
+final class HarnessSmokeTests: XCTestCase {
+
+    func testFrozenTypesVisible() {
+        // 동결 인터페이스 심볼이 AIFabric product 에서 보인다.
+        XCTAssertEqual(AIProviderID.allCases.count, 4)
+        XCTAssertEqual(AIRoutingPolicy.default.chain(for: .corpusAsk), [.remoteDS])
+    }
+
+    func testAskFixtureReadable() throws {
+        // canonical contract/fixtures/ask.json 이 로드된다 (FixtureSupport 경로 확인).
+        let data = try Fixture.data("ask.json")
+        XCTAssertGreaterThan(data.count, 0)
+    }
+}
diff --git a/Tests/AITests/LocalMLXProviderTests.swift b/Tests/AITests/LocalMLXProviderTests.swift
new file mode 100644
index 0000000..0cf5a19
--- /dev/null
+++ b/Tests/AITests/LocalMLXProviderTests.swift
@@ -0,0 +1,149 @@
+import XCTest
+@testable import AIFabric
+
+final class LocalMLXProviderTests: XCTestCase {
+
+    private let baseURL = URL(string: "http://100.76.254.116:8890")!
+
+    override func tearDown() {
+        MockURLProtocol.reset()
+        super.tearDown()
+    }
+
+    private func provider() -> LocalMLXProvider {
+        LocalMLXProvider(baseURL: baseURL, model: "gemma-macmini", session: MockURLProtocol.session())
+    }
+
+    // MARK: isAvailable probe (wake 미트리거)
+
+    func testProbeAvailable() async throws {
+        MockURLProtocol.handler = { req in
+            MockURLProtocol.ok(req.url!, json: Data(#"{"data":[{"id":"gemma-macmini"}]}"#.utf8))
+        }
+        let available = await provider().isAvailable
+        XCTAssertTrue(available)
+        // probe 는 GET /v1/models
+        XCTAssertEqual(MockURLProtocol.recorder.lastURL?.path, "/v1/models")
+        XCTAssertEqual(MockURLProtocol.recorder.lastMethod, "GET")
+    }
+
+    func testProbeUnavailableOnError() async throws {
+        MockURLProtocol.handler = { _ in throw URLError(.cannotConnectToHost) }
+        let available = await provider().isAvailable
+        XCTAssertFalse(available)   // 연결오류 → false(throw 아님)
+    }
+
+    func testProbeUnavailableOn500() async throws {
+        MockURLProtocol.handler = { req in MockURLProtocol.status(req.url!, 500) }
+        let available = await provider().isAvailable
+        XCTAssertFalse(available)
+    }
+
+    // MARK: complete — 응답 매핑 + 요청 call-shape 락
+
+    func testCompleteMapsResponseFixture() async throws {
+        let body = try Fixture.data("llm-router-chat.response.json")
+        MockURLProtocol.handler = { req in MockURLProtocol.ok(req.url!, json: body) }
+
+        let resp = try await provider().complete(
+            AICompletionRequest(task: .quickSummarize, prompt: "충격시험 면제 기준을 한 문장으로 요약해줘.",
+                                systemPrompt: "You are a concise technical assistant.", maxTokens: 512)
+        )
+        XCTAssertEqual(resp.providerUsed, .localMLX)
+        XCTAssertEqual(resp.finishReason, .completed)
+        XCTAssertTrue(resp.citations.isEmpty)
+        XCTAssertNotNil(resp.latencyMs)
+        XCTAssertTrue(resp.text.contains("UCS-66"))
+    }
+
+    func testCompleteRequestCallShape() async throws {
+        let body = try Fixture.data("llm-router-chat.response.json")
+        MockURLProtocol.handler = { req in MockURLProtocol.ok(req.url!, json: body) }
+
+        _ = try await provider().complete(
+            AICompletionRequest(task: .quickSummarize, prompt: "PROMPT_X",
+                                systemPrompt: "SYS_Y", maxTokens: 512)
+        )
+        // POST /v1/chat/completions
+        XCTAssertEqual(MockURLProtocol.recorder.lastURL?.path, "/v1/chat/completions")
+        XCTAssertEqual(MockURLProtocol.recorder.lastMethod, "POST")
+
+        // messages system/user 분리 call-shape (load-bearing)
+        let sent = try XCTUnwrap(MockURLProtocol.recorder.lastBody)
+        let decoded = try JSONDecoder().decode(SentRequest.self, from: sent)
+        XCTAssertEqual(decoded.model, "gemma-macmini")
+        XCTAssertEqual(decoded.maxTokens, 512)
+        XCTAssertEqual(decoded.stream, false)
+        XCTAssertEqual(decoded.messages.count, 2)
+        XCTAssertEqual(decoded.messages[0].role, "system")
+        XCTAssertEqual(decoded.messages[0].content, "SYS_Y")
+        XCTAssertEqual(decoded.messages[1].role, "user")
+        XCTAssertEqual(decoded.messages[1].content, "PROMPT_X")
+    }
+
+    func testNilSystemPromptSendsEmptySystemMessage() async throws {
+        let body = try Fixture.data("llm-router-chat.response.json")
+        MockURLProtocol.handler = { req in MockURLProtocol.ok(req.url!, json: body) }
+        _ = try await provider().complete(AICompletionRequest(task: .quickSummarize, prompt: "P"))
+        let sent = try XCTUnwrap(MockURLProtocol.recorder.lastBody)
+        let decoded = try JSONDecoder().decode(SentRequest.self, from: sent)
+        XCTAssertEqual(decoded.messages[0].role, "system")
+        XCTAssertEqual(decoded.messages[0].content, "")   // plan S2-2c: systemPrompt ?? ""
+    }
+
+    func testNon200BackendError() async throws {
+        MockURLProtocol.handler = { req in MockURLProtocol.status(req.url!, 503, body: "model loading") }
+        do {
+            _ = try await provider().complete(AICompletionRequest(task: .quickSummarize, prompt: "P"))
+            XCTFail("non-200 must throw backendError, not silent empty text")
+        } catch let AIProviderError.backendError(id, status, reason) {
+            XCTAssertEqual(id, .localMLX)
+            XCTAssertEqual(status, 503)
+            XCTAssertEqual(reason, "model loading")
+        }
+    }
+
+    func testRequestFixtureMatchesEncoder() throws {
+        // 동결 request fixture 의 call-shape 가 encodeRequest 산출과 일치(템플릿 placeholder 회귀 방지).
+        let fixtureData = try Fixture.data("llm-router-chat.request.json")
+        let fixture = try JSONDecoder().decode(SentRequest.self, from: fixtureData)
+        XCTAssertEqual(fixture.messages.count, 2)
+        XCTAssertEqual(fixture.messages[0].role, "system")
+        XCTAssertEqual(fixture.messages[1].role, "user")
+        XCTAssertEqual(fixture.stream, false)
+    }
+
+    // MARK: rule-fallback 라우팅 (S2-2d) — onDevice 불가 → localMLX
+
+    func testFallbackFromOnDeviceToLocalMLX() async throws {
+        let body = try Fixture.data("llm-router-chat.response.json")
+        MockURLProtocol.handler = { req in MockURLProtocol.ok(req.url!, json: body) }
+        let router = AIRouter(providers: [
+            .onDevice: MockAIProvider(id: .onDevice, available: false),   // 불가
+            .localMLX: provider(),
+        ])
+        let resp = try await router.route(AICompletionRequest(task: .quickSummarize, prompt: "P"))
+        XCTAssertEqual(resp.providerUsed, .localMLX)
+        XCTAssertEqual(resp.routingNote, "fallback from onDevice → localMLX")
+    }
+
+    func testNoFallbackNoteOnFirstChoiceSuccess() async throws {
+        let body = try Fixture.data("llm-router-chat.response.json")
+        MockURLProtocol.handler = { req in MockURLProtocol.ok(req.url!, json: body) }
+        // classify 체인 = [.localMLX, .remoteDS, .onDevice] → 1순위 localMLX 성공 → note nil
+        let router = AIRouter(providers: [.localMLX: provider()])
+        let resp = try await router.route(AICompletionRequest(task: .classify, prompt: "P"))
+        XCTAssertEqual(resp.providerUsed, .localMLX)
+        XCTAssertNil(resp.routingNote)
+    }
+
+    /// 나간 요청 디코딩용(테스트 전용 미러).
+    struct SentRequest: Decodable {
+        struct Message: Decodable { let role: String; let content: String }
+        let model: String
+        let messages: [Message]
+        let maxTokens: Int?
+        let stream: Bool
+        enum CodingKeys: String, CodingKey { case model, messages, stream; case maxTokens = "max_tokens" }
+    }
+}
diff --git a/Tests/AITests/MockURLProtocol.swift b/Tests/AITests/MockURLProtocol.swift
new file mode 100644
index 0000000..7c3d08b
--- /dev/null
+++ b/Tests/AITests/MockURLProtocol.swift
@@ -0,0 +1,86 @@
+import Foundation
+
+/// 테스트용 URLProtocol — URLSession 을 가로채 canned 응답/에러를 돌려주고, 나간 요청을 기록.
+/// 라이브 네트워크 0 으로 LocalMLX 의 probe/complete call-shape 를 검증.
+final class MockURLProtocol: URLProtocol {
+    /// (request) -> (response, body). throw 하면 URLSession 에러 경로.
+    nonisolated(unsafe) static var handler: (@Sendable (URLRequest) throws -> (HTTPURLResponse, Data))?
+    /// 마지막으로 가로챈 요청(body 포함) 기록.
+    nonisolated(unsafe) static var recorder = RequestRecorder()
+
+    static func reset() {
+        handler = nil
+        recorder = RequestRecorder()
+    }
+
+    override class func canInit(with request: URLRequest) -> Bool { true }
+    override class func canonicalRequest(for request: URLRequest) -> URLRequest { request }
+
+    override func startLoading() {
+        Self.recorder.record(request)
+        guard let handler = Self.handler else {
+            client?.urlProtocol(self, didFailWithError: URLError(.unsupportedURL))
+            return
+        }
+        do {
+            let (response, data) = try handler(request)
+            client?.urlProtocol(self, didReceive: response, cacheStoragePolicy: .notAllowed)
+            client?.urlProtocol(self, didLoad: data)
+            client?.urlProtocolDidFinishLoading(self)
+        } catch {
+            client?.urlProtocol(self, didFailWithError: error)
+        }
+    }
+
+    override func stopLoading() {}
+
+    // MARK: helpers
+
+    static func session() -> URLSession {
+        let config = URLSessionConfiguration.ephemeral
+        config.protocolClasses = [MockURLProtocol.self]
+        return URLSession(configuration: config)
+    }
+
+    static func ok(_ url: URL, json: Data) -> (HTTPURLResponse, Data) {
+        (HTTPURLResponse(url: url, statusCode: 200, httpVersion: nil, headerFields: nil)!, json)
+    }
+
+    static func status(_ url: URL, _ code: Int, body: String = "") -> (HTTPURLResponse, Data) {
+        (HTTPURLResponse(url: url, statusCode: code, httpVersion: nil, headerFields: nil)!, Data(body.utf8))
+    }
+}
+
+/// 나간 요청 기록(body 는 httpBody 또는 httpBodyStream 에서 추출 — URLProtocol 은 보통 stream 으로 전달).
+final class RequestRecorder: @unchecked Sendable {
+    private(set) var lastURL: URL?
+    private(set) var lastMethod: String?
+    private(set) var lastBody: Data?
+    private(set) var callCount = 0
+
+    func record(_ request: URLRequest) {
+        callCount += 1
+        lastURL = request.url
+        lastMethod = request.httpMethod
+        lastBody = request.bodyData
+    }
+}
+
+extension URLRequest {
+    /// URLProtocol 단계에서 body 추출 — httpBody 가 nil 이면 httpBodyStream 에서 읽음.
+    var bodyData: Data? {
+        if let httpBody { return httpBody }
+        guard let stream = httpBodyStream else { return nil }
+        stream.open()
+        defer { stream.close() }
+        var data = Data()
+        let bufSize = 8192
+        var buffer = [UInt8](repeating: 0, count: bufSize)
+        while stream.hasBytesAvailable {
+            let read = stream.read(&buffer, maxLength: bufSize)
+            if read <= 0 { break }
+            data.append(buffer, count: read)
+        }
+        return data
+    }
+}
diff --git a/Tests/AITests/OnDeviceProviderTests.swift b/Tests/AITests/OnDeviceProviderTests.swift
new file mode 100644
index 0000000..88676ce
--- /dev/null
+++ b/Tests/AITests/OnDeviceProviderTests.swift
@@ -0,0 +1,201 @@
+import XCTest
+@testable import AIFabric
+#if canImport(FoundationModels)
+import FoundationModels
+#endif
+
+/// HW 비의존 mock backend — availability + generate 결과를 주입.
+struct MockOnDeviceBackend: OnDeviceModelBackend {
+    let avail: OnDeviceAvailability
+    let outcome: Result<String, OnDeviceGenerationError>
+
+    init(avail: OnDeviceAvailability = .available,
+         outcome: Result<String, OnDeviceGenerationError> = .success("on-device ok")) {
+        self.avail = avail
+        self.outcome = outcome
+    }
+    var availability: OnDeviceAvailability { avail }
+    func generate(prompt: String, systemPrompt: String?, maxTokens: Int?) async throws -> String {
+        switch outcome {
+        case .success(let s): return s
+        case .failure(let e): throw e
+        }
+    }
+}
+
+/// complete() 호출 횟수를 세는 provider — 무폴백 단언용.
+actor CountingProvider: AIProvider {
+    nonisolated let id: AIProviderID
+    let available: Bool
+    private(set) var completeCalls = 0
+
+    init(id: AIProviderID, available: Bool) {
+        self.id = id
+        self.available = available
+    }
+    var isAvailable: Bool { get async { available } }
+    func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
+        completeCalls += 1
+        return AICompletionResponse(text: "should-not-be-called", providerUsed: id)
+    }
+}
+
+final class LogSink: @unchecked Sendable {
+    private let lock = NSLock()
+    private var storage: [String] = []
+    func append(_ s: String) { lock.lock(); storage.append(s); lock.unlock() }
+    var lines: [String] { lock.lock(); defer { lock.unlock() }; return storage }
+}
+
+final class OnDeviceProviderTests: XCTestCase {
+
+    // MARK: 가용성 + happy path (주입 backend)
+
+    func testAvailableReturnsText() async throws {
+        let p = OnDeviceProvider(backend: MockOnDeviceBackend(avail: .available, outcome: .success("요약 결과")))
+        let available = await p.isAvailable
+        XCTAssertTrue(available)
+        let resp = try await p.complete(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .onDevice)
+        XCTAssertEqual(resp.finishReason, .completed)
+        XCTAssertEqual(resp.text, "요약 결과")
+        XCTAssertNotNil(resp.latencyMs)
+    }
+
+    func testUnavailableReportsFalse() async throws {
+        let p = OnDeviceProvider(backend: MockOnDeviceBackend(avail: .unavailable(reason: "appleIntelligenceNotEnabled")))
+        let available = await p.isAvailable
+        XCTAssertFalse(available)
+    }
+
+    // MARK: GenerationError 매핑 (S2-3c)
+
+    func testGuardrailAndRefusalReturnRefused() async throws {
+        for err in [OnDeviceGenerationError.guardrailViolation, .refusal] {
+            let p = OnDeviceProvider(backend: MockOnDeviceBackend(outcome: .failure(err)))
+            let resp = try await p.complete(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+            XCTAssertEqual(resp.finishReason, .refused, "\(err) → .refused (답변의 일종, 폴백 X)")
+            XCTAssertEqual(resp.providerUsed, .onDevice)
+        }
+    }
+
+    func testRateLimitedThrowsUnavailableAndLoudLogs() async throws {
+        let sink = LogSink()
+        let p = OnDeviceProvider(backend: MockOnDeviceBackend(outcome: .failure(.rateLimited)),
+                                 log: { sink.append($0) })
+        do {
+            _ = try await p.complete(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+            XCTFail("rateLimited → throw unavailable")
+        } catch let AIProviderError.unavailable(id) {
+            XCTAssertEqual(id, .onDevice)
+        }
+        XCTAssertTrue(sink.lines.contains { $0.contains("rateLimited") }, "stateless 위반은 loud log")
+    }
+
+    func testConcurrentRequestsThrowsUnavailableAndLoudLogs() async throws {
+        let sink = LogSink()
+        let p = OnDeviceProvider(backend: MockOnDeviceBackend(outcome: .failure(.concurrentRequests)),
+                                 log: { sink.append($0) })
+        do {
+            _ = try await p.complete(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+            XCTFail("concurrentRequests → throw unavailable")
+        } catch let AIProviderError.unavailable(id) {
+            XCTAssertEqual(id, .onDevice)
+        }
+        XCTAssertTrue(sink.lines.contains { $0.contains("concurrentRequests") })
+    }
+
+    func testContextOverflowThrowsUnavailable() async throws {
+        let p = OnDeviceProvider(backend: MockOnDeviceBackend(outcome: .failure(.exceededContextWindowSize)))
+        do {
+            _ = try await p.complete(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+            XCTFail("exceededContextWindowSize → throw unavailable (폴백 유도)")
+        } catch let AIProviderError.unavailable(id) {
+            XCTAssertEqual(id, .onDevice)
+        }
+    }
+
+    // MARK: 라우팅 (S2-3d)
+
+    func testRouterFallsBackOnDeviceOverflowToLocalMLX() async throws {
+        let router = AIRouter(providers: [
+            .onDevice: OnDeviceProvider(backend: MockOnDeviceBackend(outcome: .failure(.exceededContextWindowSize))),
+            .localMLX: MockAIProvider(id: .localMLX, available: true),
+        ])
+        let resp = try await router.route(AICompletionRequest(task: .quickSummarize, prompt: "p"))
+        XCTAssertEqual(resp.providerUsed, .localMLX)
+        XCTAssertEqual(resp.routingNote, "fallback from onDevice → localMLX")
+    }
+
+    func testExplicitOnDeviceUnavailableNoFallback() async throws {
+        let counting = CountingProvider(id: .localMLX, available: true)
+        let router = AIRouter(providers: [
+            .onDevice: OnDeviceProvider(backend: MockOnDeviceBackend(avail: .unavailable(reason: "deviceNotEligible"))),
+            .localMLX: counting,
+        ])
+        do {
+            _ = try await router.route(AICompletionRequest(task: .quickSummarize, prompt: "p", explicitProvider: .onDevice))
+            XCTFail("explicit onDevice unavailable → explicitProviderUnavailable, 자동 폴백 금지")
+        } catch let AIRoutingError.explicitProviderUnavailable(id) {
+            XCTAssertEqual(id, .onDevice)
+        }
+        let calls = await counting.completeCalls
+        XCTAssertEqual(calls, 0, "명시 불가 시 타 provider complete() 호출 0")
+    }
+
+    // MARK: SDK GenerationError → 중립 매핑 lock (구성 가능한 케이스)
+
+    #if canImport(FoundationModels)
+    func testTranslateGenerationErrorCases() {
+        let ctx = LanguageModelSession.GenerationError.Context(debugDescription: "test")
+        XCTAssertEqual(FoundationModelsBackend.translate(.exceededContextWindowSize(ctx)), .exceededContextWindowSize)
+        XCTAssertEqual(FoundationModelsBackend.translate(.guardrailViolation(ctx)), .guardrailViolation)
+        XCTAssertEqual(FoundationModelsBackend.translate(.rateLimited(ctx)), .rateLimited)
+        XCTAssertEqual(FoundationModelsBackend.translate(.concurrentRequests(ctx)), .concurrentRequests)
+        XCTAssertEqual(FoundationModelsBackend.translate(.unsupportedLanguageOrLocale(ctx)), .unsupportedLanguageOrLocale)
+        XCTAssertEqual(FoundationModelsBackend.translate(.assetsUnavailable(ctx)), .assetsUnavailable)
+    }
+    #endif
+
+    // MARK: 라이브 통합 (M5 Max — 비-AI Mac 은 skip)
+
+    func testLiveOnDeviceIntegration() async throws {
+        let p = OnDeviceProvider()   // 라이브 FoundationModels backend
+        guard await p.isAvailable else {
+            throw XCTSkip("FoundationModels not available on this machine — live test skipped")
+        }
+        let resp = try await p.complete(
+            AICompletionRequest(task: .quickSummarize,
+                                prompt: "엘보 내경 가공의 핵심 관리 포인트를 한 문장으로 요약해줘.",
+                                maxTokens: 120)
+        )
+        XCTAssertEqual(resp.providerUsed, .onDevice)
+        XCTAssertEqual(resp.finishReason, .completed)
+        XCTAssertFalse(resp.text.isEmpty, "라이브 응답은 비어있지 않아야")
+    }
+
+    func testLiveCancellationCooperative() async throws {
+        let p = OnDeviceProvider()
+        guard await p.isAvailable else {
+            throw XCTSkip("FoundationModels not available — cancellation live test skipped")
+        }
+        let started = Date()
+        let task = Task { () -> AIFinishReason in
+            let r = try await p.complete(
+                AICompletionRequest(task: .quickSummarize,
+                                    prompt: "대한민국 압력용기 산업과 ASME 표준 채택 역사를 아주 길고 자세하게 여러 단락으로 서술해줘.",
+                                    maxTokens: 4000)
+            )
+            return r.finishReason
+        }
+        try? await Task.sleep(nanoseconds: 500_000_000)
+        task.cancel()
+        do {
+            _ = try await task.value
+            // 협조적이지 않으면 완주 — 정보로만(테스트 실패시키지 않음, 환경 의존).
+        } catch is CancellationError {
+            let elapsed = Date().timeIntervalSince(started)
+            XCTAssertLessThan(elapsed, 8.0, "협조적 취소면 빠르게 중단(S2-3a: ~33ms 후)")
+        }
+    }
+}
diff --git a/Tests/AITests/RemoteDSProviderTests.swift b/Tests/AITests/RemoteDSProviderTests.swift
new file mode 100644
index 0000000..992a51c
--- /dev/null
+++ b/Tests/AITests/RemoteDSProviderTests.swift
@@ -0,0 +1,151 @@
+import XCTest
+@testable import AIFabric
+
+/// 테스트용 DS client — ask.json 등 fixture 를 반환하거나 에러를 던지고, 받은 call-shape 를 기록.
+actor MockDSAskClient: DSAskClient {
+    let response: AskResponse?
+    let error: Error?
+    private(set) var lastBackend: String?
+    private(set) var lastQuery: String?
+    private(set) var callCount = 0
+
+    init(response: AskResponse? = nil, error: Error? = nil) {
+        self.response = response
+        self.error = error
+    }
+
+    func ask(query: String, backend: String) async throws -> AskResponse {
+        callCount += 1
+        lastBackend = backend
+        lastQuery = query
+        if let error { throw error }
+        return response!
+    }
+}
+
+final class RemoteDSProviderTests: XCTestCase {
+
+    private func askFixture() throws -> AskResponse {
+        try Fixture.decode(AskResponse.self, from: "ask.json")
+    }
+
+    // MARK: ask.json 디코딩 + 매핑 (call-shape 회귀)
+
+    func testAskJsonDecodeAndMap() throws {
+        let r = try askFixture()
+        XCTAssertEqual(r.synthesisStatus, "completed")
+        XCTAssertEqual(r.confidence, "high")
+        XCTAssertEqual(r.backendUsed, "gemma-macmini")
+        XCTAssertEqual(r.citations.count, 1)
+        XCTAssertEqual(r.citations[0].docId, 4912)
+        XCTAssertEqual(r.citations[0].n, 1)
+        XCTAssertEqual(r.citations[0].sectionTitle, "2. UCS-66 면제 곡선")
+
+        let mapped = RemoteDSProvider.map(r)
+        XCTAssertEqual(mapped.providerUsed, .remoteDS)
+        XCTAssertEqual(mapped.finishReason, .completed)
+        XCTAssertEqual(mapped.citations.count, 1)
+        XCTAssertEqual(mapped.citations[0].docId, 4912)
+        XCTAssertEqual(mapped.confidence, .high)
+        XCTAssertEqual(mapped.routingNote, "gemma-macmini")
+        XCTAssertEqual(mapped.latencyMs, 2841.5)
+        XCTAssertEqual(mapped.text, r.aiAnswer)
+    }
+
+    func testCompleteMapsFixture() async throws {
+        let mock = MockDSAskClient(response: try askFixture())
+        let provider = RemoteDSProvider(client: mock)
+        let resp = try await provider.complete(
+            AICompletionRequest(task: .corpusAsk, prompt: "충격시험은 언제 면제되나")
+        )
+        XCTAssertEqual(resp.providerUsed, .remoteDS)
+        XCTAssertEqual(resp.citations.count, 1)
+        XCTAssertEqual(resp.finishReason, .completed)
+        XCTAssertEqual(resp.routingNote, "gemma-macmini")
+    }
+
+    // MARK: backend exhaustive switch call-shape (쿼리 파라미터 락)
+
+    func testBackendCallShape_nilExplicit() async throws {
+        let mock = MockDSAskClient(response: try askFixture())
+        let provider = RemoteDSProvider(client: mock)
+        _ = try await provider.complete(AICompletionRequest(task: .corpusAsk, prompt: "q"))
+        let backend = await mock.lastBackend
+        XCTAssertEqual(backend, "mac-mini-default")   // 미지정 → DS 기본
+    }
+
+    func testBackendCallShape_localMLXExplicit() async throws {
+        let mock = MockDSAskClient(response: try askFixture())
+        let provider = RemoteDSProvider(client: mock)
+        _ = try await provider.complete(
+            AICompletionRequest(task: .corpusAsk, prompt: "q", explicitProvider: .localMLX)
+        )
+        let backend = await mock.lastBackend
+        XCTAssertEqual(backend, "gemma-macmini")
+    }
+
+    func testBackendMapPure() {
+        XCTAssertEqual(RemoteDSProvider.dsBackend(for: nil), "mac-mini-default")
+        XCTAssertEqual(RemoteDSProvider.dsBackend(for: .localMLX), "gemma-macmini")
+        XCTAssertEqual(RemoteDSProvider.dsBackend(for: .remoteDS), "mac-mini-default")
+        XCTAssertEqual(RemoteDSProvider.dsBackend(for: .onDevice), "mac-mini-default")
+        XCTAssertEqual(RemoteDSProvider.dsBackend(for: .specialized), "mac-mini-default")
+    }
+
+    func testNonCorpusTaskNotImplemented() async throws {
+        let mock = MockDSAskClient(response: try askFixture())
+        let provider = RemoteDSProvider(client: mock)
+        do {
+            _ = try await provider.complete(AICompletionRequest(task: .quickSummarize, prompt: "q"))
+            XCTFail("non-corpus task should not be served by RemoteDS")
+        } catch let AIProviderError.notImplemented(id) {
+            XCTAssertEqual(id, .remoteDS)
+        }
+    }
+
+    // MARK: 라우팅 — corpusAsk 무폴백 (오프라인에서도 정답)
+
+    func testCorpusAskRoutesToRemoteDSOnly() async throws {
+        let router = AIRouter(providers: [
+            .remoteDS: RemoteDSProvider(client: MockDSAskClient(response: try askFixture())),
+            .onDevice: MockAIProvider(id: .onDevice, available: true),   // available 이어도 corpusAsk 폴백 금지
+        ])
+        let resp = try await router.route(AICompletionRequest(task: .corpusAsk, prompt: "q"))
+        XCTAssertEqual(resp.providerUsed, .remoteDS)
+        XCTAssertEqual(resp.citations.count, 1)
+    }
+
+    func testCorpusAskRemoteDSDown_NoLocalFallback() async throws {
+        // remoteDS 불가 → 엉뚱한 온디바이스 코퍼스 답변 대신 명확한 실패(무폴백).
+        struct Net: Error {}
+        let router = AIRouter(providers: [
+            .remoteDS: RemoteDSProvider(client: MockDSAskClient(error: Net())),
+            .onDevice: MockAIProvider(id: .onDevice, available: true),
+        ])
+        do {
+            _ = try await router.route(AICompletionRequest(task: .corpusAsk, prompt: "q"))
+            XCTFail("corpusAsk must not fall back to onDevice")
+        } catch is Net {
+            // 기대: remoteDS 의 실패가 전파(코퍼스 체인 = [.remoteDS] only)
+        }
+    }
+
+    // MARK: S2-4b — cloud 'claude-cloud' = 명시 503 (로컬 대체 응답 금지)
+
+    func testCloud503Surfaces_NoSilentFallback() async throws {
+        let err = AIProviderError.backendError(.remoteDS, status: 503, reason: "cloud backend pending activation")
+        let router = AIRouter(providers: [
+            .remoteDS: RemoteDSProvider(client: MockDSAskClient(error: err)),
+            .onDevice: MockAIProvider(id: .onDevice, available: true),
+        ])
+        do {
+            _ = try await router.route(
+                AICompletionRequest(task: .corpusAsk, prompt: "q", explicitProvider: .remoteDS)
+            )
+            XCTFail("503 must surface, not fall back")
+        } catch let AIProviderError.backendError(id, status, _) {
+            XCTAssertEqual(id, .remoteDS)
+            XCTAssertEqual(status, 503)
+        }
+    }
+}
diff --git a/Tests/AITests/SpecializedProviderTests.swift b/Tests/AITests/SpecializedProviderTests.swift
new file mode 100644
index 0000000..7809f53
--- /dev/null
+++ b/Tests/AITests/SpecializedProviderTests.swift
@@ -0,0 +1,39 @@
+import XCTest
+@testable import AIFabric
+
+final class SpecializedProviderTests: XCTestCase {
+
+    func testScaffoldUnavailableAndNotImplemented() async throws {
+        let p = SpecializedProvider()
+        let available = await p.isAvailable
+        XCTAssertFalse(available)
+        do {
+            _ = try await p.complete(AICompletionRequest(task: .vision, prompt: "p"))
+            XCTFail("scaffold must throw notImplemented")
+        } catch let AIProviderError.notImplemented(id) {
+            XCTAssertEqual(id, .specialized)
+        }
+    }
+
+    /// .vision 체인 [.specialized, .onDevice] — specialized 등록·불가 → onDevice 로 **가시** 폴백(라우터 log).
+    /// (onDevice 는 providerUsed=id 를 정직히 반환하는 CountingProvider 사용 — MockAIProvider 는 vision 케이스에서
+    ///  providerUsed 를 .specialized 로 하드코딩하므로 부적합.)
+    func testVisionFallsBackToOnDeviceVisibly() async throws {
+        let sink = LogSink()
+        let onDevice = CountingProvider(id: .onDevice, available: true)
+        let router = AIRouter(
+            providers: [
+                .specialized: SpecializedProvider(),
+                .onDevice: onDevice,
+            ],
+            log: { sink.append($0) }
+        )
+        let resp = try await router.route(AICompletionRequest(task: .vision, prompt: "도면 보기"))
+        XCTAssertEqual(resp.providerUsed, .onDevice)
+        XCTAssertEqual(resp.routingNote, "fallback from specialized → onDevice")
+        let calls = await onDevice.completeCalls
+        XCTAssertEqual(calls, 1)
+        XCTAssertTrue(sink.lines.contains { $0.contains("specialized") && $0.contains("unavailable") },
+                      "specialized 불가가 침묵 아닌 log 로 가시화")
+    }
+}
diff --git a/Tests/AITests/TestProviders.swift b/Tests/AITests/TestProviders.swift
new file mode 100644
index 0000000..ae512bc
--- /dev/null
+++ b/Tests/AITests/TestProviders.swift
@@ -0,0 +1,28 @@
+import Foundation
+@testable import AIFabric
+
+/// providerUsed=id 를 정직히 반환하는 테스트 provider(MockAIProvider 는 일부 태스크에서 providerUsed 를 하드코딩).
+struct EchoProvider: AIProvider {
+    let id: AIProviderID
+    let available: Bool
+    init(id: AIProviderID, available: Bool = true) {
+        self.id = id
+        self.available = available
+    }
+    var isAvailable: Bool { get async { available } }
+    func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
+        AICompletionResponse(text: "echo:\(id.rawValue)", providerUsed: id)
+    }
+}
+
+/// 취소 전파 테스트용 — sleep 중 Task 취소 시 CancellationError(S2-Fe URLSession 경로 아날로그).
+struct SleepingProvider: AIProvider {
+    let id: AIProviderID
+    init(id: AIProviderID = .localMLX) { self.id = id }
+    var isAvailable: Bool { get async { true } }
+    func complete(_ request: AICompletionRequest) async throws -> AICompletionResponse {
+        try await Task.sleep(nanoseconds: 5_000_000_000)   // 취소되면 CancellationError throw
+        try Task.checkCancellation()
+        return AICompletionResponse(text: "done", providerUsed: id)
+    }
+}
diff --git a/contract/fixtures/foundationmodels-respond.json b/contract/fixtures/foundationmodels-respond.json
new file mode 100644
index 0000000..294cd2d
--- /dev/null
+++ b/contract/fixtures/foundationmodels-respond.json
@@ -0,0 +1,96 @@
+{
+  "_meta": {
+    "fixture": "foundationmodels-respond",
+    "purpose": "S2-3a — Apple FoundationModels live capture (OnDeviceProvider 결선 + 테스트 동결 기준)",
+    "captured_on": "M5 Max MacBook Pro (128GB, Apple Intelligence)",
+    "captured_date": "2026-06-04",
+    "sdk": "macOS 26.5 SDK / FoundationModels.framework",
+    "note": "SDK 가 marshaling 하므로 raw request_body 는 없음. 이 파일은 응답 모양 + 에러 타입 + 취소 동작의 동결 기준."
+  },
+
+  "availability": {
+    "observed": "available",
+    "is_available_convenience": true,
+    "read_is_synchronous": true,
+    "api": "SystemLanguageModel.default.availability",
+    "enum": {
+      "available": "case available",
+      "unavailable_reasons": ["deviceNotEligible", "appleIntelligenceNotEnabled", "modelNotReady"]
+    },
+    "supports_korean": true,
+    "supported_language_count": 23
+  },
+
+  "happy_path": {
+    "api": "session.respond(to: String, options: GenerationOptions) async throws -> Response<String>",
+    "content_accessor": "response.content   (Response<String>.content : String)",
+    "observed_content": "압력용기의 충격시험(Charpy) 면제 판정은, 용기의 압력 등급이 10MPa 이하인 경우, 충격 시험을 면제할 수 있으며, 이는 용기의 안전성을 보장하기 위한 중요한 기준입니다.",
+    "observed_latency_ms": 1291.3,
+    "transcript_entries_count": 1,
+    "is_responding_after": false,
+    "quality_note": "내용은 부정확(온디바이스 ~3B/2-bit QAT). corpusAsk 부적합·quickSummarize/classify 적합 라우팅 정합 — 사실성은 RemoteDS 코퍼스가 담당."
+  },
+
+  "session_init": {
+    "api": "LanguageModelSession(model: .default, tools: [], instructions: String?)",
+    "instructions_timing": "init (per-call 아님)",
+    "instructions_nil_handling": "systemPrompt == nil 이면 instructions 인자 생략 (빈 문자열 금지)",
+    "prewarm": "session.prewarm() — 동기 반환(관찰 ~1.3ms), 백그라운드 워밍",
+    "stateless_per_request": "호출마다 새 세션 생성 → instructions(init-time) + rateLimited/concurrentRequests(세션 상태) 둘 다 우회"
+  },
+
+  "generation_options": {
+    "api": "GenerationOptions(sampling: SamplingMode? = nil, temperature: Double? = nil, maximumResponseTokens: Int? = nil)",
+    "mapping": "AICompletionRequest.maxTokens -> maximumResponseTokens",
+    "temperature_note": "AICompletionRequest 에 temperature 필드 없음(동결) → 미설정(모델 기본). LocalMLX 와 동일 정책(둘 다 미설정)."
+  },
+
+  "generation_error": {
+    "_source": "Xcode jump-to-def / swiftinterface (LanguageModelSession.GenerationError) — authoritative, version-accurate",
+    "type": "LanguageModelSession.GenerationError : Error, LocalizedError",
+    "associated_value": "각 case 는 GenerationError.Context (refusal 은 (Refusal, Context))",
+    "cases": [
+      "exceededContextWindowSize(Context)",
+      "assetsUnavailable(Context)",
+      "guardrailViolation(Context)",
+      "unsupportedGuide(Context)",
+      "unsupportedLanguageOrLocale(Context)",
+      "decodingFailure(Context)",
+      "rateLimited(Context)",
+      "concurrentRequests(Context)",
+      "refusal(Refusal, Context)"
+    ],
+    "plan_corrections": [
+      "plan 가정 'refusal 케이스명 없음' = 틀림 → refusal 은 별도 case 로 존재(guardrailViolation 과 구분).",
+      "plan 에 없던 concurrentRequests case 존재 — rateLimited 와 함께 stateless 세션에서 뜨면 세션 공유 버그 신호.",
+      "assetsUnavailable 정확명 확정(모델 자산 미가용)."
+    ],
+    "reproduced_live": {
+      "exceededContextWindowSize": {
+        "trigger": "의도적 컨텍스트 오버플로(긴 프롬프트)",
+        "errorDescription": "Exceeded model context window size"
+      }
+    },
+    "finish_reason_mapping": {
+      "guardrailViolation": ".refused",
+      "refusal": ".refused",
+      "exceededContextWindowSize": ".unavailable",
+      "rateLimited": ".unavailable + loud log (stateless 인데 발생 = 세션 재사용 버그 신호)",
+      "concurrentRequests": ".unavailable + loud log (동일 — stateless 위반 신호)",
+      "unsupportedLanguageOrLocale": ".unavailable (+ supportedLocale 사전체크로 회피)",
+      "unsupportedGuide": ".unavailable",
+      "decodingFailure": ".unavailable",
+      "assetsUnavailable": ".unavailable",
+      "@unknown default": ".unavailable + loud log"
+    }
+  },
+
+  "cancellation": {
+    "_finding": "S2-Fe 전제 확정 — COOPERATIVE",
+    "cancel_requested_at_ms": 500,
+    "threw": "CancellationError",
+    "elapsed_ms": 533.6,
+    "interpretation": "respond() 는 mid-flight Task 취소를 협조적으로 honor(요청 33ms 후 CancellationError throw).",
+    "implication": "OnDevice complete() 에 surrounding Task.checkCancellation() 은 belt-and-suspenders(실제 중단은 respond() 내부). streamResponse 토큰단위 취소 폴백 불필요(선전환 금지)."
+  }
+}
diff --git a/contract/fixtures/llm-router-chat.request.json b/contract/fixtures/llm-router-chat.request.json
new file mode 100644
index 0000000..1fb6fbe
--- /dev/null
+++ b/contract/fixtures/llm-router-chat.request.json
@@ -0,0 +1,19 @@
+{
+  "_meta": {
+    "fixture": "llm-router-chat.request",
+    "status": "PROVISIONAL_SYNTHETIC",
+    "synthetic": true,
+    "captured_date": null,
+    "reason": "맥미니(hyungi-macmini 100.76.254.116) Tailscale offline(last seen 8h+, 2026-06-04) → 라이브 캡처 불가. OpenAI /v1/chat/completions 표준 스펙 기반 합성. 맥미니 복귀 시 S2-2a 라이브 재캡처로 교체(별 fixture-update PR, S2-Ff drift-check 경유).",
+    "endpoint": "POST http://100.76.254.116:8890/v1/chat/completions",
+    "call_shape_note": "messages = [system, user] 분리 고정(load-bearing). system.content = AICompletionRequest.systemPrompt ?? \"\" (plan S2-2c). max_tokens = AICompletionRequest.maxTokens.",
+    "model_note": "model 문자열은 llm-router 가 기대하는 별칭 — 라이브 캡처로 확정 필요(provisional: 'gemma-macmini')."
+  },
+  "model": "gemma-macmini",
+  "messages": [
+    { "role": "system", "content": "You are a concise technical assistant." },
+    { "role": "user", "content": "충격시험 면제 기준을 한 문장으로 요약해줘." }
+  ],
+  "max_tokens": 512,
+  "stream": false
+}
diff --git a/contract/fixtures/llm-router-chat.response.json b/contract/fixtures/llm-router-chat.response.json
new file mode 100644
index 0000000..5a948bb
--- /dev/null
+++ b/contract/fixtures/llm-router-chat.response.json
@@ -0,0 +1,25 @@
+{
+  "_meta": {
+    "fixture": "llm-router-chat.response",
+    "status": "PROVISIONAL_SYNTHETIC",
+    "synthetic": true,
+    "captured_date": null,
+    "reason": "맥미니 offline → 라이브 캡처 불가. OpenAI chat.completion 표준 응답 모양 기반 합성. 라이브 재캡처로 교체(S2-Ff).",
+    "shape": "OpenAI chat.completion (choices[0].message.content → text, finish_reason → AIFinishReason)"
+  },
+  "id": "chatcmpl-provisional-0001",
+  "object": "chat.completion",
+  "created": 0,
+  "model": "gemma-macmini",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "충격시험 면제는 재료군(Curve A~D)과 거버닝 두께에 따른 UCS-66 면제 곡선으로 MDMT에서 판정합니다."
+      },
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": { "prompt_tokens": 24, "completion_tokens": 41, "total_tokens": 65 }
+}