Fail model load early when free RAM is below the catalog estimate

Sources/VocaMac/Models/AppState.swift

@@ -324,6 +324,10 @@ final class AppState: ObservableObject {
     /// Set to `true` in tests to avoid side effects.
     let skipSystemIntegration: Bool
 
+    /// Pre-load memory gate. Production defaults to SystemInfo; tests stub this
+    /// so CI free+inactive pages cannot flake medium/large mock loads.
+    var modelFitsInMemory: (ModelSize) -> Bool = { SystemInfo.canFitModelInMemory($0) }
+
     // MARK: - Initialization
 
     init(
@@ -1252,6 +1256,26 @@ final class AppState: ObservableObject {
         // we don't know yet — we'll detect it after loading completes.
         let targetSize = size
 
+        // Refuse known-too-large loads before WhisperKit/CoreML can hang the
+        // UI spinner under memory pressure (vocamac#250).
+        if let targetSize,
+           !modelFitsInMemory(targetSize) {
+            let needed = String(format: "%.1f", targetSize.ramRequiredGB)
+            let failureMessage =
+                "Not enough free memory to load \(targetSize.displayName) "
+                + "(~\(needed) GB needed). Free RAM or choose a smaller model."
+            showTemporaryError(failureMessage)
+            VocaLogger.error(.appState, failureMessage)
+            await restorePreviousModelIfNeeded(
+                afterFailedLoadFor: targetSize,
+                previousSize: previousModelSize,
+                previousName: previousLoadedModelName,
+                hadLoadedModel: hadLoadedModel,
+                originalFailureMessage: failureMessage
+            )
+            return
+        }
+
         // Mark the model as loading in the UI
         if let targetSize = targetSize, let idx = availableModels.firstIndex(where: { $0.size == targetSize }) {
             availableModels[idx].isLoading = true

Sources/VocaMac/Services/SystemInfo.swift

@@ -3,6 +3,7 @@
 //
 // Detects system hardware capabilities and recommends optimal whisper model size.
 
+import Darwin
 import Foundation
 
 // MARK: - SystemCapabilities
@@ -147,4 +148,42 @@ enum SystemInfo {
         // Use at most half the cores, minimum 2, maximum 8
         return max(2, min(cores / 2, 8))
     }
+
+    /// Approximate reclaimable free memory in bytes. Zero means the probe failed.
+    static var availableMemoryBytes: UInt64 {
+        var stats = vm_statistics64()
+        var count = mach_msg_type_number_t(
+            MemoryLayout<vm_statistics64_data_t>.stride / MemoryLayout<integer_t>.stride
+        )
+        let host = mach_host_self()
+        let kr = withUnsafeMutablePointer(to: &stats) {
+            $0.withMemoryRebound(to: integer_t.self, capacity: Int(count)) {
+                host_statistics64(host, HOST_VM_INFO64, $0, &count)
+            }
+        }
+        guard kr == KERN_SUCCESS else { return 0 }
+        var pageSize: vm_size_t = 0
+        guard host_page_size(host, &pageSize) == KERN_SUCCESS, pageSize > 0 else { return 0 }
+        let pages = UInt64(stats.free_count) + UInt64(stats.inactive_count)
+        return pages * UInt64(pageSize)
+    }
+
+    /// Whether loading `size` is likely to fit without thrashing.
+    ///
+    /// Uses the catalog RAM estimate against installed memory and against
+    /// reclaimable free memory from host_statistics64. A zero available
+    /// reading is treated as unknown so we do not block loads on a failed probe.
+    static func canFitModelInMemory(
+        _ size: ModelSize,
+        physicalMemoryGB: Int = physicalMemoryGB,
+        availableBytes: UInt64 = availableMemoryBytes
+    ) -> Bool {
+        let requiredGB = size.ramRequiredGB
+        guard Double(physicalMemoryGB) + 0.001 >= requiredGB else {
+            return false
+        }
+        guard availableBytes > 0 else { return true }
+        let requiredBytes = UInt64((requiredGB * 1024 * 1024 * 1024).rounded(.up))
+        return availableBytes >= requiredBytes
+    }
 }

Tests/VocaMacTests/AppStateTests.swift

@@ -347,6 +347,43 @@ final class AppStateModelLoadingTests: XCTestCase {
         )
     }
 
+    @MainActor
+    func testLowMemoryGateRefusesMediumBeforeWhisperAndRestoresPrevious() async {
+        UserDefaults.standard.set(ModelSize.small.rawValue, forKey: "vocamac.selectedModelSize")
+
+        let modelManager = MockModelManager()
+        modelManager.downloadedModels = [.small, .medium]
+
+        let whisperService = MockWhisperService()
+        whisperService.loadedModelName = "openai_whisper-small"
+        whisperService.isModelLoaded = true
+        whisperService.loadResponses = [
+            .success("openai_whisper-small"),
+        ]
+
+        let (appState, mocks) = AppState.makeTestState(
+            modelManager: modelManager,
+            whisperService: whisperService
+        )
+        appState.modelFitsInMemory = { $0 != .medium }
+
+        await appState.loadModel(.medium)
+
+        XCTAssertTrue(
+            appState.errorMessage?.contains("Not enough") == true
+                || appState.errorMessage?.localizedCaseInsensitiveContains("free memory") == true
+        )
+        XCTAssertFalse(
+            mocks.whisperService.loadRequests.map { $0.name }.contains("openai_whisper-medium")
+        )
+        XCTAssertEqual(
+            mocks.whisperService.loadRequests.map { $0.name },
+            ["openai_whisper-small"]
+        )
+        XCTAssertEqual(appState.currentModel?.size, .small)
+        XCTAssertEqual(appState.selectedModelSize, ModelSize.small.rawValue)
+    }
+
     @MainActor
     func testDeleteModelRemovesDownloadedModel() async {
         let modelManager = MockModelManager()

Tests/VocaMacTests/Mocks/MockServices.swift

@@ -572,6 +572,8 @@ extension AppState {
             permissionManager: permissionManager,
             skipSystemIntegration: true
         )
+        // Bypass host free-RAM probe so mock loads are not refused on CI.
+        appState.modelFitsInMemory = { _ in true }
         return (appState, mocks)
     }
 }

Tests/VocaMacTests/ModelTests.swift

@@ -53,6 +53,47 @@ final class SystemInfoTests: XCTestCase {
         XCTAssertTrue(summary.contains("Metal:"))
         XCTAssertTrue(summary.contains("Recommended Model:"))
     }
+
+    func testCanFitModelRejectsWhenPhysicalMemoryIsTooLow() {
+        XCTAssertFalse(
+            SystemInfo.canFitModelInMemory(
+                .largeV3Latest,
+                physicalMemoryGB: 4,
+                availableBytes: UInt64(64) * 1024 * 1024 * 1024
+            )
+        )
+    }
+
+    func testCanFitModelRejectsWhenAvailableMemoryIsTooLow() {
+        XCTAssertFalse(
+            SystemInfo.canFitModelInMemory(
+                .base,
+                physicalMemoryGB: 16,
+                availableBytes: 64 * 1024 * 1024
+            )
+        )
+    }
+
+    func testCanFitModelAllowsWhenAvailableIsUnknown() {
+        XCTAssertTrue(
+            SystemInfo.canFitModelInMemory(
+                .tiny,
+                physicalMemoryGB: 8,
+                availableBytes: 0
+            )
+        )
+    }
+
+    func testCanFitModelAllowsWhenPhysicalAndAvailableAreEnough() {
+        let required = UInt64((ModelSize.base.ramRequiredGB * 1024 * 1024 * 1024).rounded(.up))
+        XCTAssertTrue(
+            SystemInfo.canFitModelInMemory(
+                .base,
+                physicalMemoryGB: 16,
+                availableBytes: required
+            )
+        )
+    }
 }
 
 // MARK: - ModelSize Tests