arcane-os 0.10.1 → 0.11.0

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # Changelog
 
+## 0.11.0
+
+- Add explicit browser chat CPU selection with `loadDefaults:{gpuLayers:0}`
+  or `load({gpuLayers:0})`. Use the packaged Wllama CPU path without WebGPU
+  requirements or adapter initialization. Omission retains full GPU offload;
+  model sources, cache, complete responses, streaming, and cancellation remain
+  on their existing owners.
+- On mobile browsers and tablets, keep shared Speech focused on voice output:
+  hide and disable transcription activation and recording, omit transcription
+  status/progress, and retain mute/unmute. Chat retains End. Desktop speech and
+  the dedicated transcription component retain their existing behavior.
+
 ## 0.10.1
 
 - Open developer error modals only for newly captured live incidents. Restoring

package/README.md

@@ -19,7 +19,7 @@ version-locked SDK runtime, while an integrated Arcane checkout uses its live
 `arcane/` runtime. Both profiles preserve the same app URLs, theme, packaging,
 event, cancellation, and browser run contracts.
 
-This checkout defines the `0.10.1` SDK contract. Applications pin one exact npm
+This checkout defines the `0.11.0` SDK contract. Applications pin one exact npm
 version and lockfile; registry state is deliberately not baked into application
 artifacts.
 
@@ -41,7 +41,7 @@ Create one browser application, install its pinned SDK, and start its source
 server:
 
 ```bash
-npx arcane-os@0.10.1 new hello-speech --path ./hello-speech --target browser
+npx arcane-os@0.11.0 new hello-speech --path ./hello-speech --target browser
 cd hello-speech
 npm install
 npm run dev
@@ -392,7 +392,7 @@ uses the same controller for automatic memory extraction.
 Create a new repository-shaped Arcane application with the exact stable SDK:
 
 ```bash
-npx arcane-os@0.10.1 new my-app --path ./my-app --target portable --git
+npx arcane-os@0.11.0 new my-app --path ./my-app --target portable --git
 cd my-app
 npm install
 npm run dev
@@ -402,7 +402,7 @@ To enroll an existing repository, install the exact SDK and initialize only
 missing Arcane files:
 
 ```bash
-npm install --save-dev --save-exact arcane-os@0.10.1
+npm install --save-dev --save-exact arcane-os@0.11.0
 npm exec -- arcane init my-app --target portable
 ```
 
@@ -418,7 +418,7 @@ npm exec -- arcane-os targets
 No global SDK install or standalone Arcane CLI is required. The application
 repository's exact npm dependency and lockfile own the CLI and toolchain version.
 
-Use `npx arcane-os@0.10.1` for the initial bootstrap because it names this npm
+Use `npx arcane-os@0.11.0` for the initial bootstrap because it names this npm
 package explicitly; bare `npx arcane` outside an installed project could resolve
 a different package. Both installed commands invoke the same headless toolchain.
 Project-local npm scripts use the SDK pinned by that app's `package-lock.json`,
@@ -438,7 +438,7 @@ node ./bin/arcane.mjs new local-app --path ../local-app --target portable --git
 
 # From the generated app repository
 cd ../local-app
-npm install --save-dev --save-exact ../arcane-os-sdk/arcane-os-0.10.1.tgz
+npm install --save-dev --save-exact ../arcane-os-sdk/arcane-os-0.11.0.tgz
 npm ci
 ```
 
@@ -447,7 +447,7 @@ same location. The lockfile retains the selected package dependency while
 Arcane uses the installed package name and version. Local directory `file:` dependencies are not
 accepted because npm may install them as links; use a packed `.tgz`. A GitHub
 runner also needs that tarball at the locked path. After publication, replace
-the local declaration with the exact `arcane-os@0.10.1` registry package and
+the local declaration with the exact `arcane-os@0.11.0` registry package and
 commit the regenerated lock.
 
 Generated repositories use `npm ci --ignore-scripts` in CI. Run dependency
@@ -586,7 +586,7 @@ package installation, or assertions.
 
 ## Current target support
 
-Version `0.10.1` exposes one browser target and five explicitly paired
+Version `0.11.0` exposes one browser target and five explicitly paired
 native development targets: a non-runnable portable directory, a
 Windows x64 unsigned-local-test EXE bundle, Linux x64 and Linux ARM64
 unsigned-local-test DEBs, and an Android development-signed APK. The

package/browser-runtime/ai/browser-wasm-llm-provider.mjs

@@ -2605,7 +2605,7 @@ function capabilityLoadPlan(runtimeCapabilities, defaults, options = {}) {
     contextTokens,
     batchTokens,
     microBatchTokens,
-    gpuLayers: 99_999,
+    gpuLayers: configured.gpuLayers === 0 ? 0 : 99_999,
     ...(reasoning!==undefined?{ reasoning }:{}),
     ...(chatTemplate!==undefined?{ chatTemplate }:{}),
     ...(jinja!==undefined?{ jinja }:{}),
@@ -2625,14 +2625,20 @@ function sameLoadPlan(left, right) {
     && JSON.stringify(left?.templateDefaults) === JSON.stringify(right?.templateDefaults);
 }
 
-function stableModelFailure(error) {
+function stableModelFailure(error, loadPlan) {
   const code = is.string(error?.code) ? error.code : "";
   const message = is.string(error?.message) ? error.message : "";
   if (code === "ARCANE_AI_MODEL_SHARD_TOO_LARGE") {
     return completeValue({ code });
   }
   if (/(?:out of memory|allocation failed|failed to allocate|memory exhausted)/iu.test(message)) {
-    return completeValue({ code: "ARCANE_AI_MODEL_GPU_MEMORY_INSUFFICIENT" });
+    return completeValue(
+        {
+            code: loadPlan.gpuLayers === 0
+                ? 'ARCANE_AI_MODEL_CPU_MEMORY_INSUFFICIENT'
+                : 'ARCANE_AI_MODEL_GPU_MEMORY_INSUFFICIENT'
+        }
+    );
   }
   if (code === "ARCANE_AI_WEBGPU_EVIDENCE_INVALID") {
     return completeValue({ code: "ARCANE_AI_MODEL_FULL_OFFLOAD_UNPROVEN" });
@@ -2667,10 +2673,10 @@ function capabilityPolicy(
   if (runtimeCapabilities.secureContext !== true) {
     add("ARCANE_AI_SECURE_CONTEXT_REQUIRED", "incompatible");
   }
-  if (runtimeCapabilities.webgpuApiPresent !== true) {
+  if (loadPlan.gpuLayers > 0 && runtimeCapabilities.webgpuApiPresent !== true) {
     add("ARCANE_AI_WEBGPU_API_UNAVAILABLE", "incompatible");
   }
-  if (failure) {
+  if (failure && failure.gpuLayers === loadPlan.gpuLayers) {
     add(failure.code, "incompatible");
   }
   if (storage?.compatibility === "incompatible") {
@@ -2679,7 +2685,13 @@ function capabilityPolicy(
     add(storage?.code ?? "ARCANE_AI_STORAGE_NOT_MEASURED", "unknown");
   }
   const webgpu = runtimeEvidence?.webgpu;
-  if (state === "ready" && webgpu?.observed === true) {
+  if (loadPlan.gpuLayers === 0) {
+    const cpuLoaded = state === 'ready' && runtimeEvidence?.executionDevice === 'cpu';
+    add(
+      cpuLoaded ? 'ARCANE_AI_CPU_MODEL_LOADED' : 'ARCANE_AI_CPU_MODEL_NOT_LOADED',
+      cpuLoaded ? 'compatible' : 'unknown',
+    );
+  } else if (state === "ready" && webgpu?.observed === true) {
     add(
       "ARCANE_AI_WEBGPU_EXECUTION_OBSERVED",
       "compatible",
@@ -2799,6 +2811,8 @@ export function createBrowserWasmLlmProvider({
       webgpu: runtimeCapabilities.webgpu,
       webgpuApiPresent: runtimeCapabilities.webgpuApiPresent,
       webgpuOperational: runtimeCapabilities.webgpuOperational,
+      webgpuRequired: activeLoadPlan.gpuLayers > 0,
+      executionDevice: activeLoadPlan.gpuLayers === 0 ? 'cpu' : 'webgpu',
       webgpuEvidenceProtocol: runtimeCapabilities.webgpuEvidenceProtocol,
       webgpuAdapterSelectionEvent: WEBGPU_ADAPTER_SELECTED_EVENT,
       crossOriginIsolated: runtimeCapabilities.crossOriginIsolated,
@@ -2920,7 +2934,7 @@ export function createBrowserWasmLlmProvider({
       if (!sameLoadPlan(activeLoadPlan, requestedLoadPlan)) {
         throw fail(
           "ARCANE_AI_LOAD_PLAN_RELOAD_REQUIRED",
-          "Unload the browser-WASM model before changing its context or batch load plan.",
+          "Unload the browser-WASM model before changing its execution device, context, or batch load plan.",
         );
       }
       activeSecurity = effectiveSecurity;
@@ -2936,7 +2950,7 @@ export function createBrowserWasmLlmProvider({
       if (!sameLoadPlan(activeLoadPlan, requestedLoadPlan)) {
         throw fail(
           "ARCANE_AI_LOAD_PLAN_RELOAD_REQUIRED",
-          "The in-flight browser-WASM load uses a different context or batch plan.",
+          "The in-flight browser-WASM load uses a different execution device, context, or batch plan.",
         );
       }
       activeSecurity = effectiveSecurity;
@@ -3082,8 +3096,13 @@ export function createBrowserWasmLlmProvider({
           operation: "load",
           signal: normalizationSignal(surfaced, signal),
         });
-        const modelFailure = stableModelFailure(normalized);
-        if (modelFailure) modelFailures.set(activeSource.id, modelFailure);
+        const modelFailure = stableModelFailure(normalized, activeLoadPlan);
+        if (modelFailure) {
+          modelFailures.set(
+              activeSource.id,
+              {...modelFailure, gpuLayers: activeLoadPlan.gpuLayers}
+          );
+        }
         if (generation === lifecycleGeneration && state === "loading") {
           state = "error";
           errorState = completeValue({ code: normalized.code, message: normalized.message });
@@ -3196,6 +3215,10 @@ export function createBrowserWasmLlmProvider({
         state = "unloaded";
         errorState = null;
         activeSecurity = null;
+        activeLoadPlan = capabilityLoadPlan(
+          measuredRuntimeCapabilities(runtime.capabilities()),
+          runtimeLoadDefaults,
+        );
         return status();
       } catch (error) {
         const normalized = normalizeArcaneAIError(error, {
@@ -3394,7 +3417,7 @@ export function adaptV1LlmProvider(provider) {
         [capabilities?.webAssembly === true, "WebAssembly"],
         [capabilities?.opfs === true, "OPFS"],
         [capabilities?.secureContext === true, "a secure context"],
-        [capabilities?.webgpuApiPresent === true, "the WebGPU API"],
+        [capabilities?.webgpuRequired === false || capabilities?.webgpuApiPresent === true, "the WebGPU API"],
       ];
       const missing = requirements.find(([available]) => !available)?.[1] ?? null;
       if (missing) {

package/browser-runtime/ai/browser-wllama-runtime.mjs

@@ -20,7 +20,9 @@ export const BROWSER_WASM_RUNTIME_AUTHORITY = completeValue({
   protocol: "arcane-ai-browser-wasm/2",
   provider: "wllama",
   executionPolicy: {
-    webgpuRequired: true,
+    webgpuRequired: false,
+    defaultGpuLayers: FULL_GPU_LAYERS,
+    cpuGpuLayers: 0,
     cpuFallback: false,
     cancellation: "abortSignal-plus-llama-cancel-acknowledgement",
     cleanup: "worker-termination-only",
@@ -71,6 +73,7 @@ function runtimeCapabilitySnapshot(evidence) {
     webgpu: webgpuOperational,
     webgpuApiPresent: Boolean(navigatorObject?.gpu),
     webgpuOperational,
+    executionDevice: evidence?.executionDevice ?? null,
     webgpuEvidenceProtocol: RUNTIME_EVIDENCE_PROTOCOL,
     crossOriginIsolated: globalThis.crossOriginIsolated === true,
     secureContext: globalThis.isSecureContext === true,
@@ -329,6 +332,7 @@ function initialEvidence() {
   return completeValue({
     protocol: RUNTIME_EVIDENCE_PROTOCOL,
     state: "unloaded",
+    executionDevice: null,
     webgpu: {
       observed: false,
       apiPresent: Boolean(globalThis.navigator?.gpu),
@@ -339,7 +343,8 @@ function initialEvidence() {
 }
 
 /**
- * Creates one packaged, WebGPU-required Wllama session. This factory has no
+ * Creates one packaged Wllama session. GPU loading is the default; gpuLayers: 0
+ * selects the upstream CPU path. This factory has no
  * network or browser side effects until load() is called. Runtime URLs are
  * fixed relative to this module for npm and materialized /arcane/sdk trees.
  */
@@ -473,7 +478,12 @@ export function createPackagedWllamaRuntime({ logger = arcaneLogging } = {}) {
     if (!is.object(globalThis.WebAssembly)) {
       throw new Error("WebAssembly is unavailable in this browser.");
     }
-    if (!globalThis.navigator?.gpu) {
+    const gpuLayers = options.gpuLayers === 0
+      ? 0
+      : normalizePositiveInteger(options.gpuLayers, FULL_GPU_LAYERS);
+    const webgpuRequired = gpuLayers > 0;
+    const executionDevice = webgpuRequired ? 'webgpu' : 'cpu';
+    if (webgpuRequired && !globalThis.navigator?.gpu) {
       throw runtimeFailure(
         "ARCANE_AI_WEBGPU_REQUIRED",
         "A WebGPU API is required, but navigator presence alone will not establish operational execution.",
@@ -486,9 +496,6 @@ export function createPackagedWllamaRuntime({ logger = arcaneLogging } = {}) {
     const next = newEngine();
     const threads = normalizePositiveInteger(options.threads, 1);
     const contextTokens = normalizePositiveInteger(options.contextTokens, 4_096);
-    const gpuLayers = options.gpuLayers === undefined
-      ? FULL_GPU_LAYERS
-      : normalizePositiveInteger(options.gpuLayers, FULL_GPU_LAYERS);
     const loadOptions = {
       n_threads: threads,
       n_ctx: contextTokens,
@@ -527,7 +534,8 @@ export function createPackagedWllamaRuntime({ logger = arcaneLogging } = {}) {
 
     publishEvidence({
       state: "loading",
-      webgpu: { observed: false, apiPresent: true },
+      executionDevice,
+      webgpu: { observed: false, apiPresent: Boolean(globalThis.navigator?.gpu) },
       cancellation: null,
       cleanup: null,
     });
@@ -568,43 +576,45 @@ export function createPackagedWllamaRuntime({ logger = arcaneLogging } = {}) {
         );
       }
       let adapter = null;
-      const telemetryOperation = trackOperation(
-          Promise.resolve().then(
-              function observeLoadedGpu() {
-                  return next.arcaneTelemetry();
+      if (webgpuRequired) {
+          const telemetryOperation = trackOperation(
+              Promise.resolve().then(
+                  function observeLoadedGpu() {
+                      return next.arcaneTelemetry();
+                  }
+              )
+          );
+          function cancelGpuObservation() {
+              telemetryOperation.cancel(loadController.signal.reason);
+          }
+          loadController.signal.addEventListener(
+              'abort',
+              cancelGpuObservation,
+              {once: true}
+          );
+          if (loadController.signal.aborted) cancelGpuObservation();
+          try {
+              // Observe the adapter this loaded Worker already selected; do not probe
+              // a second adapter in the page and attribute it to the model.
+              const telemetry = await telemetryOperation.result;
+              if (telemetry?.worker?.invalid === true) {
+                  arcaneLogging.warn("The loaded model's GPU adapter observation is inconsistent.", telemetry.worker);
+              } else {
+                  adapter = telemetry?.worker?.adapter ?? null;
               }
-          )
-      );
-      function cancelGpuObservation() {
-          telemetryOperation.cancel(loadController.signal.reason);
-      }
-      loadController.signal.addEventListener(
-          'abort',
-          cancelGpuObservation,
-          {once: true}
-      );
-      if (loadController.signal.aborted) cancelGpuObservation();
-      try {
-          // Observe the adapter this loaded Worker already selected; do not probe
-          // a second adapter in the page and attribute it to the model.
-          const telemetry = await telemetryOperation.result;
-          if (telemetry?.worker?.invalid === true) {
-              arcaneLogging.warn("The loaded model's GPU adapter observation is inconsistent.", telemetry.worker);
-          } else {
-              adapter = telemetry?.worker?.adapter ?? null;
+          } catch (error) {
+              if (loadController.signal.aborted || error instanceof WllamaRuntimeError) throw error;
+              arcaneLogging.warn("The loaded model's GPU adapter details are unavailable.", error);
+          } finally {
+              loadController.signal.removeEventListener('abort', cancelGpuObservation);
           }
-      } catch (error) {
-          if (loadController.signal.aborted || error instanceof WllamaRuntimeError) throw error;
-          arcaneLogging.warn("The loaded model's GPU adapter details are unavailable.", error);
-      } finally {
-          loadController.signal.removeEventListener('abort', cancelGpuObservation);
       }
       if (progressFailure) throw progressFailure;
       if (loadController.signal.aborted) throw cancellationError(loadController.signal.reason);
       if (pending?.engine !== next) throw new Error('Wllama load was cancelled.');
       const webgpu = {
-          observed: true,
-          apiPresent: true,
+          observed: webgpuRequired,
+          apiPresent: Boolean(globalThis.navigator?.gpu),
           adapter
       };
       pending = null;
@@ -638,6 +648,7 @@ export function createPackagedWllamaRuntime({ logger = arcaneLogging } = {}) {
       contextTokens,
       threads,
       gpuLayers,
+      executionDevice,
       evidence: evidenceState,
       metadata: engine.getModelMetadata?.() ?? null,
     });

package/docs/reference/ai/browser-wasm.md

@@ -80,6 +80,19 @@ GPU request, model operation, listener, or viewport-dependent reload. Its
 dependency-free entrypoint does not import the inference runtime. Profile
 selection does not alter prompts, documents, history, or output limits.
 
+## CPU selection
+
+Set `loadDefaults:{gpuLayers:0}` when creating the provider, or pass
+`gpuLayers:0` to `ai.load()` or `provider.load()`, to run the selected model on
+CPU through the packaged Wllama WebAssembly runtime. CPU selection skips
+WebGPU capability checks, adapter initialization, and GPU execution reporting.
+It uses the same complete model sources, cache, request, streaming, and
+cancellation contracts. Existing context, batch, micro-batch, and thread
+settings remain independently configurable.
+
+Omitting `gpuLayers` preserves full GPU offload (`99999`). CPU is an explicit
+selection; a failed GPU load does not silently switch to CPU.
+
 ## Lifecycle at a glance
 
 `createArcaneAI()` owns one LLM controller. Its default `loadPolicy` is
@@ -206,15 +219,16 @@ Applications remain responsible for model selection and license compliance.
 cannot download. Source downloads use CORS, omit credentials and referrer,
 disable HTTP caching, and honor `AbortSignal`.
 
-After a model loads, its existing Worker reports the adapter it selected.
+After a GPU model loads, its existing Worker reports the adapter it selected.
 The runtime retains that adapter in `evidence().webgpu.adapter`, including
 optional browser-reported `type` and `isFallbackAdapter` fields. This reads
 the loaded Worker's metadata once; it does not request another adapter.
 Missing fields remain unknown. Cancellation also cancels the metadata wait;
 an unavailable observation is logged without rejecting an otherwise loaded model.
 
-On Windows desktop Chromium browsers, an explicitly reported integrated GPU,
-CPU, or fallback adapter triggers one GPU-only browser alert per page session.
+For GPU loads on Windows desktop Chromium browsers, an explicitly reported
+integrated GPU, CPU, or fallback adapter triggers one GPU-only browser alert
+per page session.
 The alert names the selected adapter and **Force High Performance GPU**
 (`#force-high-performance-gpu`), conditional on the computer also having a
 discrete GPU. It contains the address to paste manually and asks the user to
@@ -278,9 +292,10 @@ serialized; provider status exposes `busy` and `queued`. Supported request
 generation fields include temperature, top-K, top-P, min-P, repeat penalty,
 and seed. Load settings separately include
 `contextTokens`, `batchTokens`, `microBatchTokens`, `threads`, and GPU-layer
-count. The shipped runtime always sets `gpuLayers: 99999`: WebGPU and proved
-full offload are mandatory, and there is no CPU fallback. Secure context and
-WebGPU/full-offload availability remain browser platform requirements;
+count. The shipped runtime defaults to `gpuLayers:99999` for full GPU offload;
+`gpuLayers:0` explicitly selects CPU. WebGPU/full-offload availability applies
+only to the GPU route, which has no automatic CPU fallback. Secure context
+remains a browser platform requirement;
 cross-origin isolation and coarse hardware fields remain observations.
 
 Tool definitions, tool choice, parallel-tool-call preference, and JSON or JSON
@@ -528,11 +543,11 @@ intent. They do not activate checking in the ordinary development contract.
 
 ### Availability and normalization
 
-**Browser secure context with WebAssembly, OPFS/DBOPFS, WebGPU, and full-offload
-support.** Inference is local after a successful Wllama load.
-The runtime forces `gpuLayers: 99999`; callers cannot select CPU or partial
-offload. Status discloses the optional `secure` intent, capability state,
-storage/model compatibility, and lifecycle state; it does not claim that
+**Browser secure context with WebAssembly and OPFS/DBOPFS.** The default GPU
+route additionally requires WebGPU and full-offload support; `gpuLayers:0`
+selects CPU without those GPU requirements. Inference is local after a
+successful Wllama load. Status discloses the optional `secure` intent,
+capability state, storage/model compatibility, and lifecycle state; it does not claim that
 hardening ran.
 
 ### Example

package/docs/reference/availability-and-normalization.md

@@ -70,7 +70,7 @@ version; WebKitGTK availability must not be generalized to macOS.
 | Publish application events or review a complete event history | `arcane-os/event-manager` | **Node** and **Browser**; optional DOM capture needs a browser DOM or compatible host | Live listeners receive original arguments. Ordinary `secure:false` recording preserves complete URLs, public details, and captured stack text in deeply frozen `arcane-event-stack/1` snapshots while credential-named fields remain redacted. The stack format is local diagnostic data, not a host transport. |
 | Build browser UI and app-local behavior | `/arcane/modules/*.js`, shared entities, and components | **Browser**; many modules also run inside every native renderer | Pure modules own their result contracts. Modules that call `Arcane` inherit the bridge boundary described below. |
 | Select and observe independent LLM/STT/TTS roles | `/arcane/modules/AIProviderRuntime.js` and `AIRuntimeState.js` | **Cross-host** controller/state; registered providers retain their own host requirements | Required/projected provider members, route/configuration records, and status fields; per-role lifecycle, cancellation, stream cleanup, sticky state, and startup barriers are normalized. `localOnly` creates no fallback. |
-| Run a caller-selected local LLM entirely in a browser renderer | `arcane-os/ai/browser-wasm` through `createArcaneAI()` | **Browser** only; secure context, WebAssembly, OPFS/DBOPFS, WebGPU, and requested full offload are required; no CPU fallback | The public AI API module normalizes multi-model lifecycle, status, complete all-choice streaming, cancellation, exact ordered structural tool-call visibility, and session persistence. Model sources are canonical ordered file descriptors; licenses and model choice remain application policy. |
+| Run a caller-selected local LLM entirely in a browser renderer | `arcane-os/ai/browser-wasm` through `createArcaneAI()` | **Browser** only; secure context, WebAssembly, and OPFS/DBOPFS; full WebGPU offload by default or explicit CPU with `gpuLayers:0` | The public AI API module normalizes multi-model lifecycle, status, complete all-choice streaming, cancellation, exact ordered structural tool-call visibility, and session persistence. Model sources are canonical ordered file descriptors; licenses and model choice remain application policy. |
 | Run caller-selected Whisper or Kokoro in a browser renderer | `arcane-os/ai/browser-speech` registered with `AIProviderRuntime` | **Browser** only; DBOPFS, Web Locks, Workers, Fetch/object URLs, and a caller-supplied self-contained runtime/model closure are required | STT/TTS use independent provider/2 lifecycle and status. Kokoro adds bounded Worker/session concurrency and explicit `auto`, `webgpu`, or `wasm` execution. Complete model/runtime selection, offline behavior, cancellation, Worker teardown, and request/result shapes are normalized. No runtime/model content or cloud fallback is supplied. |
 | Prepare ordered speech playback or reuse the speech-input formatting filter | `arcane-os/speech-playback` and `arcane-os/speech-text` | **Node** with injected media adapters, or **Browser / Native WebView** media; the text filter itself is **Cross-host** | Stored and caller-owned text stays exact. Only the outbound synthesis copy automatically loses repeated same formatting marks. A capacity-advertising provider receives complete segments immediately while retaining indexed playback; native/custom synthesis stays serialized. |
 | Preserve complete chat history and memory | `/arcane/modules/PersistentAIChatSession.js` | **Browser / native WebView** with ChatEntity/DBOPFS and a configured chat function | Existing DBOPFS names and memory semantics are preserved. Live-context commit is atomic; durable persistence is explicit and coherent across user/assistant turns and atomic all-ID tool-result batches. |
@@ -133,9 +133,12 @@ surface does not require an Arcane Core method grant because it does not call a
 Core host. Browser Fetch, CORS, storage policy, secure-context behavior, and
 resource limits still apply.
 
-The current browser runtime requires WebGPU and has no CPU fallback. A successful
-load requests full GPU offload (`gpuLayers: 99999`). `navigator.gpu` presence by
-itself is not readiness. The provider emits the instrumented
+The browser runtime defaults to full GPU offload (`gpuLayers:99999`). Explicit
+`loadDefaults:{gpuLayers:0}` or `load({gpuLayers:0})` selects CPU and skips
+WebGPU requirements and adapter work; GPU failure never selects CPU implicitly.
+Both routes wait for Wllama to report the complete model loaded.
+`navigator.gpu` presence by itself is not readiness. On the GPU route, the
+provider emits the instrumented
 `arcane.ai.browser-wasm.webgpu.adapter.selected` capability event after adapter
 selection.
 

package/docs/reference/inventory/package-api.json

@@ -2932,7 +2932,7 @@
       "primaryImport": "arcane-os/ai/browser-wasm",
       "group": "Browser-WASM local AI",
       "summary": "Creates the packaged Wllama browser provider with complete serialized inference, streaming, abort, unload, and probe operations.",
-      "availability": "Browser context with WebAssembly, OPFS/DBOPFS, and WebGPU; no CPU fallback",
+      "availability": "Browser context with WebAssembly and OPFS/DBOPFS; WebGPU by default or explicit CPU with gpuLayers:0",
       "protocol": "arcane-ai-adapter/1 with OpenAI-like chat completion and streaming envelopes",
       "normalization": "Mutable complete lifecycle/status records, declaration/history validation, atomic nonblank all-ID tool-result sequencing, complete all-choice nonstructural ordinary iterator projections, all-choice terminal validation, terminal-only structural-call acceptance, and exact structured tool-call data with required user-facing message and retained extension fields; the SDK never invokes application tools"
     },

package/docs/reference/protocols.md

@@ -355,10 +355,12 @@ does not silently delete the app-owned cache. A complete whole member supersedes
 its current resumable fragments. Cleanup failure is warned without hiding the
 usable model.
 
-SDK `0.5.17` requires WebGPU. Load requests full offload and waits for the runtime
-to report a loaded model. `navigator.gpu` presence alone is
-not readiness. There is no CPU fallback, partial-offload success mode, or
-silent switch to native/Core/cloud inference.
+Browser-WASM loads default to full WebGPU offload. Explicit
+`loadDefaults:{gpuLayers:0}` or `load({gpuLayers:0})` selects CPU through the
+same packaged Wllama runtime without GPU initialization or GPU-only
+requirements. Both routes wait for the runtime to report the complete model
+loaded. `navigator.gpu` presence alone is not readiness. There is no automatic
+CPU fallback or silent switch to native/Core/cloud inference.
 
 ### Browser speech lifecycle
 
@@ -556,6 +558,11 @@ model or selects a fallback. `speech.html` and `voice-transcription.html` consum
 one shared `createSTTActivationController()` contract for selected, unloaded,
 loading, unloading, error, and ready presentation plus cancelable user intent.
 Both keep capture unavailable until sticky STT state is exactly ready.
+On mobile browsers, `speech.html` additionally omits transcription activation,
+capture, and transcription status/progress while retaining its TTS control and
+voice status. Its public activation callback returns `false` without dispatch.
+This uses the SDK's browser-device classification, including tablets, and does
+not change the dedicated `voice-transcription.html` workflow or provider APIs.
 
 Each shared speech component owns an `AbortController` for its STT request and
 passes its signal through `AI.fetchSTT()`. `voice-transcription.html` also adds

package/docs/reference/runtime-components.md

@@ -1163,8 +1163,17 @@ Shared dependencies: [`AI.js`](runtime-modules.md#aijs),
 [`ComponentContracts.js`](runtime-modules.md#componentcontractsjs), and
 [`DBLS.js`](runtime-modules.md#dblsjs).
 
-The Hold to talk control remains disabled unless sticky STT state is `ready`,
-the role is not busy, and microphone capture is available. For an explicitly
+On mobile browsers, including tablets, the shared browser-device classifier
+selects a Speech interface for voice output. Transcription activation,
+recording, sending recorded audio, and their status/progress are unavailable;
+`requestSTTActivation()` returns `false` without requesting a provider action.
+The keyboard remains the voice-input path. The existing TTS mute/unmute control
+and voice status remain available, and Chat retains its End control. Desktop
+speech and the dedicated `voice-transcription.html` component retain their
+existing behavior.
+
+On desktop, the Hold to talk control remains disabled unless sticky STT state
+is `ready`, the role is not busy, and microphone capture is available. For an explicitly
 selected STT route, a separate keyboard-operable control presents Start
 transcription while `unloaded`, Cancel loading while `loading`, a disabled
 Canceling state while `unloading`, and Try again with the sticky error while

package/docs/reference/sdk-api.md

@@ -122,7 +122,7 @@ browser map are cataloged separately in [Runtime modules](runtime-modules.md).
 | `createBrowserModelSource()` | function | `arcane-os/ai/browser-wasm` | Browser-WASM local AI | Browser Fetch with a readable response body |
 | `createBrowserSpeechArtifactGraph()` | function | `arcane-os/ai/browser-speech` | Browser speech providers | Browser metadata; construction starts no fetch, cache, Worker, provider, or event operation |
 | `createBrowserSpeechAuthority()` | function | `arcane-os/ai/browser-speech` | Browser speech providers | Browser descriptor construction; use requires the selected storage and provider Web APIs |
-| `createBrowserWasmLlmProvider()` | function | `arcane-os/ai/browser-wasm` | Browser-WASM local AI | Browser context with WebAssembly, OPFS/DBOPFS, and WebGPU; no CPU fallback |
+| `createBrowserWasmLlmProvider()` | function | `arcane-os/ai/browser-wasm` | Browser-WASM local AI | Browser context with WebAssembly and OPFS/DBOPFS; WebGPU by default or explicit CPU with gpuLayers:0 |
 | `createBrowserWhisperProvider()` | function | `arcane-os/ai/browser-speech` | Browser speech providers | Browser with Workers, object URLs, caller-selected Whisper runtime/model artifacts, and an SDK DBOPFS speech store |
 | `createCanonicalUstarHeader()` | function | `arcane-os` | Packaging and release bundles | Node |
 | `createDbopfsModelStore()` | function | `arcane-os/ai/browser-wasm` | Browser-WASM local AI | Browser with a ready DBOPFS instance and OPFS |
@@ -6112,8 +6112,9 @@ exposes `protocol`, `id`, default `model`, `catalog`, `capabilities`, `status`,
 Direct `load()` selects a catalog model and returns `{model,status}`; the
 public AI API module's `ai.load()` returns the flat controller status. Load settings include
 offline mode, `AbortSignal`, progress, threads, and context/batch/micro-batch
-tokens. The runtime always forces `gpuLayers:99999`; callers cannot request CPU
-or partial offload.
+tokens. `loadDefaults:{gpuLayers:0}` or `load({gpuLayers:0})` explicitly selects
+CPU. Omission retains full GPU offload (`gpuLayers:99999`); GPU failures never
+silently switch to CPU.
 
 Chat supports OpenAI-like message/generation fields, tools, tool choice,
 parallel tool-call preference, and JSON/JSON-Schema structured output.
@@ -6133,11 +6134,13 @@ turn.
 
 ### Availability and normalization
 
-**Browser context with WebAssembly, OPFS/DBOPFS, and WebGPU.** A load succeeds
-after Wllama confirms the complete model is loaded. There is no CPU fallback.
+**Browser context with WebAssembly and OPFS/DBOPFS.** The default GPU route
+also requires WebGPU; explicit `gpuLayers:0` CPU loading skips GPU requirements
+and adapter work. A load succeeds after Wllama confirms the complete model is
+loaded. There is no automatic CPU fallback.
 Cross-origin isolation and coarse hardware fields remain observations rather
 than hard gates. Ordinary status reports complete catalog compatibility and
-lifecycle state. Adapter selection is instrumented as
+lifecycle state. GPU adapter selection is instrumented as
 `arcane.ai.browser-wasm.webgpu.adapter.selected`.
 Cancellation normalizes to `ARCANE_AI_REQUEST_ABORTED`; load or availability failures
 surface stable `ARCANE_AI_*` codes such as `ARCANE_AI_WEBGPU_REQUIRED`,
@@ -6278,8 +6281,8 @@ Re-adapting the same provider returns the same adapter object.
 ### Availability and normalization
 
 **Browser; the wrapped provider retains its own secure-context, WebAssembly,
-OPFS/DBOPFS, WebGPU, model, and lifecycle requirements.** The adapter normalizes
-provider/2 selection and lifecycle without adding Node, native, cloud, network,
+OPFS/DBOPFS, selected CPU or WebGPU route, model, and lifecycle requirements.**
+The adapter normalizes provider/2 selection and lifecycle without adding Node, native, cloud, network,
 or CPU fallback behavior.
 
 ### Example

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "arcane-os",
-  "version": "0.10.1",
+  "version": "0.11.0",
   "description": "Arcane OS JavaScript SDK, project-local CLI, browser runtime, and repository-portable application packager.",
   "type": "module",
   "main": "./src/index.mjs",
@@ -68,7 +68,7 @@
     "test:release": "node ./bin/arcane-test.mjs test/npm-release.test.mjs",
     "test:unit": "node ./bin/arcane-test.mjs test/app-descriptor.test.mjs test/app-schema.test.mjs test/contracts.test.mjs test/doctor.test.mjs test/mail-credentials.test.mjs test/mail-outbox.test.mjs test/mail-public-api.test.mjs test/mail-send.test.mjs test/mail-transport.test.mjs test/targets.test.mjs test/workspace-operation-lock.test.mjs",
     "test:functional": "node ./bin/arcane-test.mjs test/browser-speech-providers.test.mjs test/browser-wasm-gpu-notice.test.mjs test/browser-wasm-download-resume.test.mjs test/cli.test.mjs test/dbopfs-document-library.test.mjs test/dev-server.test.mjs test/dev-pwa.test.mjs test/dom-event-instrumentation.test.mjs test/event-manager.test.mjs test/events.test.mjs test/import-map.test.mjs test/mail-cli.test.mjs test/mail-runtime.test.mjs test/mail-server.test.mjs test/packaging.test.mjs test/pwa-packaging.test.mjs test/pwa-client.test.mjs test/pwa-worker.test.mjs test/persistent-ai-chat-session.test.mjs test/reference-completeness.test.mjs test/runtime-api-behavior.test.mjs test/runtime.test.mjs test/scaffold.test.mjs test/speech-playback.test.mjs test/site.test.mjs test/update-check.test.mjs",
-    "test:integration": "node ./bin/arcane-test.mjs test/integrated-shared.test.mjs test/integrated-workspace.test.mjs test/mail-browser.test.mjs test/native-plan.test.mjs test/native-provider-loader.test.mjs test/npm-release.test.mjs test/release-bundle.test.mjs test/release-capability-smoke.test.mjs test/shared-payload-batch.test.mjs test/tarball.test.mjs test/wllama-webgpu-runtime.test.mjs",
+    "test:integration": "node ./bin/arcane-test.mjs test/integrated-shared.test.mjs test/integrated-workspace.test.mjs test/mail-browser.test.mjs test/native-plan.test.mjs test/native-provider-loader.test.mjs test/npm-release.test.mjs test/release-bundle.test.mjs test/release-capability-smoke.test.mjs test/shared-payload-batch.test.mjs test/tarball.test.mjs test/browser-wasm-cpu.test.mjs test/wllama-webgpu-runtime.test.mjs",
     "test:regression": "node ./bin/arcane-test.mjs test/channel-workflows.test.mjs test/html-import-registration.test.mjs test/logging-regression.test.mjs test/markdown-speech.test.mjs test/prepared-speech.test.mjs test/native-provider-generation.test.mjs test/speech-queue-regression.test.mjs test/testing.test.mjs test/test-sets.test.mjs",
     "check": "node tools/check-source.mjs && npm test",
     "check:release": "node tools/check-source.mjs --package-only && npm run test:release",

package/runtime/arcane/components/speech.html

@@ -137,10 +137,24 @@
             grid-column: 3;
         }
     }
+
+    :host([data-device-class="mobile"]) {
+        grid-template-columns: minmax(0, 1fr) auto;
+    }
+
+    :host([data-device-class="mobile"]) .speech_feedback {
+        grid-row: 1;
+        grid-column: 1;
+    }
+
+    :host([data-device-class="mobile"]) #muteButton {
+        grid-row: 1;
+        grid-column: 2;
+    }
 </style>
 
 <p class="hidden">Will auto send after X seconds of silence set in code.</p>
-<button id="recordButton" type="button" aria-label="Transcription unavailable" disabled>Hold to talk</button>
+<button id="recordButton" type="button" aria-label="Transcription unavailable" hidden disabled>Hold to talk</button>
 <button id="sttActivationButton" type="button" hidden disabled>Start transcription</button>
 <button id="stopButton" type="button" class="hidden" disabled>Send</button>
 <div class="speech_feedback">
@@ -155,6 +169,7 @@
 <button id="muteButton" type="button" class="mute" aria-pressed="true" aria-label="Voice unavailable" disabled>Voice unavailable</button>
 
 <script type="module">
+    const browserDeviceImport = import('arcane-os/browser-device');
     const {default:Is}=await import('strong-type');
     const is=new Is(false);
 
@@ -180,8 +195,12 @@
         createSTTActivationController,
         formatAIRuntimeProgress
     } = await import('../modules/ComponentContracts.js');
+    const {getBrowserDeviceClass} = await browserDeviceImport;
 
     const host = this;
+    const deviceClass = getBrowserDeviceClass();
+    const transcriptionEnabled = deviceClass !== 'mobile';
+    host.dataset.deviceClass = deviceClass;
     const events = createArcaneEventSource(
         host,
         {
@@ -326,28 +345,36 @@
     host.configure = configure;
     host.destroy = destroy;
     host.reportTTSError = reportTTSError;
-    host.requestSTTActivation = host.requestSTTActivation
-        || function requestSTTActivation(intent) {
-            return requestAIRuntimeIntent(intent);
+    if (transcriptionEnabled) {
+        host.requestSTTActivation = host.requestSTTActivation
+            || function requestSTTActivation(intent) {
+                return requestAIRuntimeIntent(intent);
+            };
+    } else {
+        host.requestSTTActivation = function unavailableMobileSTTActivation() {
+            return false;
         };
+    }
     host.setMuted = setMuted;
     host.availability = {
-        microphone: Boolean(
+        microphone: transcriptionEnabled && Boolean(
             navigator.mediaDevices?.getUserMedia
             && globalThis.MediaRecorder
         ),
         stt: false,
         tts: false
     };
-    const sttActivationController = createSTTActivationController(
-        {
-            host,
-            button: sttActivationButton,
-            progress: sttActivationProgress,
-            onChange: renderSTTActivationState,
-            eventSource: events
-        }
-    );
+    const sttActivationController = transcriptionEnabled
+        ? createSTTActivationController(
+            {
+                host,
+                button: sttActivationButton,
+                progress: sttActivationProgress,
+                onChange: renderSTTActivationState,
+                eventSource: events
+            }
+        )
+        : null;
 
     const lifecycleListenerOptions = completeValue(
         { signal: runtimeStateAbortController.signal }
@@ -516,13 +543,13 @@
     }
 
     function applyRoleTransitions(previousSTTRole, previousTTSRole) {
-        sttActivationController.synchronize(sttRole);
+        sttActivationController?.synchronize(sttRole);
         const sttBecameUnready = previousSTTRole.state === 'ready'
             && sttRole.state !== 'ready';
         const ttsBecameUnready = previousTTSRole.state === 'ready'
             && ttsRole.state !== 'ready';
 
-        host.availability.stt = sttRole.state === 'ready';
+        host.availability.stt = transcriptionEnabled && sttRole.state === 'ready';
         host.availability.tts = ttsRole.state === 'ready';
 
         if (sttBecameUnready) {
@@ -571,6 +598,17 @@
     }
 
     function renderControls() {
+        if (!transcriptionEnabled) {
+            recordButton.hidden = true;
+            recordButton.disabled = true;
+            sttActivationButton.hidden = true;
+            sttActivationButton.disabled = true;
+            stopButton.hidden = true;
+            stopButton.disabled = true;
+            sttActivationProgress.hidden = true;
+            renderTTSControl();
+            return;
+        }
         const transcriptionReady = sttRole.state === 'ready';
         const microphoneReady = host.availability.microphone
             && is.function(globalThis.MediaRecorder)
@@ -693,17 +731,19 @@
         if (localStatus?.message) {
             messages.push(localStatus.message);
             tone = localStatus.tone;
-            if (sttActivationController.pending
-                || ['loading', 'unloading'].includes(sttRole.state)) {
+            if (transcriptionEnabled
+                && (sttActivationController.pending
+                    || ['loading', 'unloading'].includes(sttRole.state))) {
                 messages.push(
                     describeSTTRole()
                 );
             }
-        } else {
+        } else if (transcriptionEnabled) {
             messages.push(describeSTTRole());
         }
         messages.push(describeTTSRole());
-        if (sttRole.state === 'error' || ttsRole.state === 'error') {
+        if ((transcriptionEnabled && sttRole.state === 'error')
+            || ttsRole.state === 'error') {
             tone = 'error';
         }
 
@@ -712,13 +752,14 @@
         speechStatus.setAttribute(
             'aria-busy',
             String(
-                sttRole.state === 'loading'
-                || sttRole.state === 'unloading'
+                (transcriptionEnabled
+                    && (sttRole.state === 'loading'
+                        || sttRole.state === 'unloading'
+                        || sttRole.busy
+                        || sttActivationController.pending
+                        || transcriptionActive))
                 || ttsRole.state === 'loading'
-                || sttRole.busy
                 || ttsRole.busy
-                || sttActivationController.pending
-                || transcriptionActive
             )
         );
     }
@@ -776,7 +817,9 @@
     }
 
     async function observeMicrophonePermission() {
-        if (!navigator.permissions?.query || !navigator.mediaDevices?.getUserMedia) {
+        if (!transcriptionEnabled
+            || !navigator.permissions?.query
+            || !navigator.mediaDevices?.getUserMedia) {
             return;
         }
 
@@ -1226,7 +1269,8 @@
 
     async function record() {
         if (
-            destroyed
+            !transcriptionEnabled
+            || destroyed
             || sttRole.state !== 'ready'
             || sttRole.busy
             || !host.availability.microphone
@@ -1725,7 +1769,7 @@
 
         destroyed = true;
         runtimeStateAbortController.abort();
-        sttActivationController.destroy();
+        sttActivationController?.destroy();
         window.removeEventListener('pagehide', handlePageHide);
         window.removeEventListener('pageshow', handlePageShow);
         microphonePermissionStatus?.removeEventListener?.(