arcane-os 0.1.2 → 0.2.0

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

@@ -2,6 +2,16 @@ import { Wllama } from "./wllama/index.mjs";
 
 const MODULE_URL = new URL("./wllama/index.mjs", import.meta.url).href;
 const WASM_URL = new URL("./wllama/wllama.wasm", import.meta.url).href;
+const WEBGPU_EVIDENCE_PROTOCOL = "arcane-wllama-webgpu-evidence/1";
+const RUNTIME_EVIDENCE_PROTOCOL = "arcane-wllama-runtime-evidence/1";
+const FULL_GPU_LAYERS = 99_999;
+const WEBGPU_ADAPTER_PATTERN = /^ggml_webgpu: adapter_info: vendor_id: (\d+) \| vendor: (.*?) \| architecture: (.*?) \| device_id: (\d+) \| name: (.*?) \| device_desc: (.*)$/u;
+const GPU_OFFLOAD_PATTERN = /^[^:]+: offloaded (\d+)\/(\d+) layers to GPU$/u;
+const PEG_NATIVE_OUTPUT_PREFIX = "common_chat_peg_parse: unparsed peg-native output: ";
+const PEG_NATIVE_FINAL_PREFIX = "<|channel|>final <|constrain|>content<|message|>";
+const PEG_NATIVE_FAILURE = "The model produced output that does not match the expected peg-native format";
+const MAX_RECOVERED_COMPLETION_CHARACTERS = 1_048_576;
+const MAX_RECOVERED_COMPLETION_LINES = 16_384;
 
 function deepFreeze(value) {
   if (!value || typeof value !== "object" || Object.isFrozen(value)) return value;
@@ -10,8 +20,19 @@ function deepFreeze(value) {
 }
 
 export const BROWSER_WASM_RUNTIME_AUTHORITY = deepFreeze({
-  protocol: "arcane-ai-browser-wasm/1",
+  protocol: "arcane-ai-browser-wasm/2",
   provider: "wllama",
+  executionPolicy: {
+    webgpuRequired: true,
+    cpuFallback: false,
+    operationalEvidence: RUNTIME_EVIDENCE_PROTOCOL,
+    navigatorPresenceIsOperationalEvidence: false,
+    cancellation: "abortSignal-plus-llama-cancel-acknowledgement",
+    cleanup: "worker-termination-only",
+    nativeUnloadClaimed: false,
+    physicalVramReclamationClaimed: false,
+    telemetryThreatModel: "authenticated-module-closure-and-frozen-prototypes-not-hostile-platform-global-attestation",
+  },
   package: {
     name: "@wllama/wllama",
     version: "3.6.0",
@@ -40,9 +61,17 @@ export const BROWSER_WASM_RUNTIME_AUTHORITY = deepFreeze({
     module: {
       path: "ai/wllama/index.mjs",
       url: MODULE_URL,
-      bytes: 373_519,
-      sha256: "4637e42d636010493a9b274fbbe70bfd8120365da726b1d9e589d85ca84a00d6",
+      bytes: 392_852,
+      sha256: "b119a7cdffabc8541dce283381d18ada4027c0560728aac1fe45bdd30cdac8e2",
       mediaType: "text/javascript",
+      projection: {
+        protocol: WEBGPU_EVIDENCE_PROTOCOL,
+        tool: "tools/project-wllama-webgpu-runtime.mjs",
+        sourcePath: "node_modules/@wllama/wllama/esm/index.js",
+        sourceBytes: 373_519,
+        sourceSha256: "4637e42d636010493a9b274fbbe70bfd8120365da726b1d9e589d85ca84a00d6",
+        wasmModified: false,
+      },
     },
     wasm: {
       path: "ai/wllama/wllama.wasm",
@@ -59,12 +88,23 @@ export const BROWSER_WASM_RUNTIME_AUTHORITY = deepFreeze({
   },
 });
 
-function runtimeCapabilitySnapshot() {
+function runtimeFailure(code, message, cause) {
+  const error = new Error(message, cause === undefined ? undefined : { cause });
+  error.name = "ArcaneWllamaRuntimeError";
+  error.code = code;
+  return error;
+}
+
+function runtimeCapabilitySnapshot(evidence) {
   const navigatorObject = globalThis.navigator;
+  const webgpuOperational = evidence?.state === "ready" && evidence?.webgpu?.observed === true;
   return Object.freeze({
     webAssembly: typeof globalThis.WebAssembly === "object",
     opfs: typeof navigatorObject?.storage?.getDirectory === "function",
-    webgpu: Boolean(navigatorObject?.gpu),
+    webgpu: webgpuOperational,
+    webgpuApiPresent: Boolean(navigatorObject?.gpu),
+    webgpuOperational,
+    webgpuEvidenceProtocol: RUNTIME_EVIDENCE_PROTOCOL,
     crossOriginIsolated: globalThis.crossOriginIsolated === true,
     secureContext: globalThis.isSecureContext === true,
     hardwareConcurrency: Number.isSafeInteger(navigatorObject?.hardwareConcurrency)
@@ -82,16 +122,308 @@ function normalizePositiveInteger(value, fallback, { maximum = Number.MAX_SAFE_I
   return number;
 }
 
+function createEvidenceLogger(logger) {
+  let adapter = null;
+  let offload = null;
+  let invalid = false;
+  let completionCapture = null;
+
+  function same(left, right) {
+    return JSON.stringify(left) === JSON.stringify(right);
+  }
+
+  function observeCompletionLine(level, value) {
+    if (!completionCapture) return;
+    const line = String(value).replace(/\r$/u, "");
+    if (!completionCapture.started) {
+      if (level !== "warn" || !line.startsWith(PEG_NATIVE_OUTPUT_PREFIX)) return;
+      completionCapture.started = true;
+      completionCapture.lines.push(line.slice(PEG_NATIVE_OUTPUT_PREFIX.length));
+    } else if (level === "error" && line.trim() === PEG_NATIVE_FAILURE) {
+      completionCapture.complete = true;
+      return;
+    } else if (!completionCapture.complete) {
+      if (level !== "log") completionCapture.invalid = true;
+      completionCapture.lines.push(line);
+    }
+    completionCapture.characters += line.length + 1;
+    if (
+      completionCapture.characters > MAX_RECOVERED_COMPLETION_CHARACTERS
+      || completionCapture.lines.length > MAX_RECOVERED_COMPLETION_LINES
+    ) {
+      completionCapture.invalid = true;
+      completionCapture.lines.length = 0;
+    }
+  }
+
+  function observeLine(level, value) {
+    observeCompletionLine(level, value);
+    const line = String(value).trim();
+    if (!line) return;
+    const adapterMatch = line.match(WEBGPU_ADAPTER_PATTERN);
+    if (adapterMatch) {
+      const next = Object.freeze({
+        vendorId: Number(adapterMatch[1]),
+        vendor: adapterMatch[2],
+        architecture: adapterMatch[3],
+        deviceId: Number(adapterMatch[4]),
+        name: adapterMatch[5],
+        description: adapterMatch[6],
+      });
+      if (adapter && !same(adapter, next)) invalid = true;
+      else adapter = next;
+    }
+    const offloadMatch = line.match(GPU_OFFLOAD_PATTERN);
+    if (offloadMatch) {
+      const next = Object.freeze({
+        layers: Number(offloadMatch[1]),
+        totalLayers: Number(offloadMatch[2]),
+      });
+      if (offload && !same(offload, next)) invalid = true;
+      else offload = next;
+    }
+  }
+
+  function observe(level, args) {
+    for (const value of args) {
+      if (typeof value !== "string") continue;
+      for (const line of value.split(/\r?\n/u)) observeLine(level, line);
+    }
+  }
+
+  const wrapped = {};
+  for (const level of ["debug", "log", "warn", "error"]) {
+    wrapped[level] = (...args) => {
+      observe(level, args);
+      if (typeof logger?.[level] === "function") logger[level](...args);
+    };
+  }
+
+  return Object.freeze({
+    logger: Object.freeze(wrapped),
+    beginCompletionCapture() {
+      if (completionCapture) {
+        throw runtimeFailure(
+          "ARCANE_AI_RUNTIME_BUSY",
+          "A Wllama completion capture is already active.",
+        );
+      }
+      const capture = {
+        started: false,
+        complete: false,
+        invalid: false,
+        characters: 0,
+        lines: [],
+      };
+      completionCapture = capture;
+
+      function release() {
+        if (completionCapture === capture) completionCapture = null;
+      }
+
+      return Object.freeze({
+        recover(error, { aborted = false } = {}) {
+          release();
+          const stack = String(error?.stack ?? "");
+          if (
+            aborted
+            || error?.name !== "Error"
+            || error?.message !== "Invalid magic number"
+            || !stack.includes("glueDeserialize")
+            || !stack.includes("ProxyToWorker")
+            || !capture.started
+            || !capture.complete
+            || capture.invalid
+          ) return null;
+
+          const raw = capture.lines.join("\n");
+          if (!raw.startsWith(PEG_NATIVE_FINAL_PREFIX)) return null;
+          const content = raw.slice(PEG_NATIVE_FINAL_PREFIX.length);
+          if (!content.trim() || content.includes("<|")) return null;
+          return content;
+        },
+        release,
+      });
+    },
+    snapshot() {
+      return deepFreeze({ adapter, offload, invalid });
+    },
+  });
+}
+
+function createStructuredStreamCapture() {
+  let content = "";
+  let invalid = false;
+  let sawContent = false;
+  let chunks = 0;
+
+  function observe(value) {
+    chunks += 1;
+    if (!value || typeof value !== "object" || !Array.isArray(value.choices)) {
+      invalid = true;
+      return;
+    }
+    if (value.choices.length !== 1) {
+      invalid = true;
+      return;
+    }
+    const choice = value.choices[0];
+    const delta = choice?.delta;
+    if (
+      choice?.index !== 0
+      || !delta
+      || typeof delta !== "object"
+      || Array.isArray(delta)
+      || (delta.role !== undefined && delta.role !== "assistant")
+      || (choice.finish_reason !== undefined && choice.finish_reason !== null)
+      || delta.tool_calls !== undefined
+      || delta.function_call !== undefined
+      || delta.reasoning_content !== undefined
+    ) {
+      invalid = true;
+      return;
+    }
+    if (delta.content === undefined || delta.content === null) return;
+    if (typeof delta.content !== "string") {
+      invalid = true;
+      return;
+    }
+    content += delta.content;
+    sawContent = true;
+    if (content.length > MAX_RECOVERED_COMPLETION_CHARACTERS) invalid = true;
+  }
+
+  return Object.freeze({
+    observe,
+    matches(value) {
+      return chunks > 0
+        && !invalid
+        && sawContent
+        && content.length > 0
+        && content === value;
+    },
+  });
+}
+
+function validCounter(value) {
+  return Number.isSafeInteger(value) && value >= 0;
+}
+
+function verifyProjectedTelemetry(value) {
+  const worker = value?.worker;
+  if (
+    value?.protocol !== WEBGPU_EVIDENCE_PROTOCOL
+    || worker?.protocol !== WEBGPU_EVIDENCE_PROTOCOL
+    || !validCounter(worker.bufferCount)
+    || !validCounter(worker.bufferBytes)
+    || !validCounter(worker.queueSubmissions)
+    || !validCounter(worker.commandBuffers)
+    || !validCounter(worker.queueFenceRequests)
+    || !validCounter(worker.queueFenceCompletions)
+    || worker.queueFenceCompletions > worker.queueFenceRequests
+    || worker.invalid === true
+  ) {
+    throw runtimeFailure(
+      "ARCANE_AI_WEBGPU_EVIDENCE_INVALID",
+      "The projected Wllama WebGPU evidence was missing or invalid.",
+    );
+  }
+  return value;
+}
+
+function adapterEvidenceConflicts(workerAdapter, loggedAdapter) {
+  if (!workerAdapter || !loggedAdapter) return false;
+  function loggedText(value) {
+    return typeof value === "string" ? value.slice(0, 256) : "";
+  }
+  return workerAdapter.vendor !== loggedText(loggedAdapter.vendor)
+    || workerAdapter.architecture !== loggedText(loggedAdapter.architecture)
+    || workerAdapter.name !== loggedText(loggedAdapter.name)
+    || workerAdapter.description !== loggedText(loggedAdapter.description);
+}
+
+function admittedLoadEvidence(logs, projected) {
+  const worker = projected.worker;
+  const adapter = worker.adapter;
+  const offload = logs.offload;
+  const failures = [];
+  if (logs.invalid) failures.push("conflicting-log-evidence");
+  if (!adapter) failures.push("adapter-selection");
+  else if (adapterEvidenceConflicts(adapter, logs.adapter)) failures.push("adapter-log-conflict");
+  if (!offload) failures.push("offload-log");
+  else if (
+    !Number.isSafeInteger(offload.layers)
+    || !Number.isSafeInteger(offload.totalLayers)
+    || offload.totalLayers < 1
+  ) failures.push("offload-shape");
+  else if (offload.layers !== offload.totalLayers) {
+    failures.push(`full-offload(${offload.layers}/${offload.totalLayers})`);
+  }
+  if (worker.bufferCount < 1) failures.push(`buffer-count(${worker.bufferCount})`);
+  if (worker.bufferBytes < 1) failures.push(`buffer-bytes(${worker.bufferBytes})`);
+  if (worker.queueSubmissions < 1) failures.push(`queue-submissions(${worker.queueSubmissions})`);
+  if (worker.commandBuffers < 1) failures.push(`command-buffers(${worker.commandBuffers})`);
+  if (worker.queueFenceRequests < 1) failures.push(`fence-requests(${worker.queueFenceRequests})`);
+  if (worker.queueFenceCompletions < worker.queueFenceRequests) {
+    failures.push(`fence-completions(${worker.queueFenceCompletions}/${worker.queueFenceRequests})`);
+  }
+  if (failures.length > 0) {
+    throw runtimeFailure(
+      "ARCANE_AI_WEBGPU_REQUIRED",
+      `Wllama WebGPU admission failed: ${failures.join(", ")}.`,
+    );
+  }
+  return deepFreeze({
+    observed: true,
+    adapter,
+    offload: { ...offload, allReportedModelLayers: true },
+    buffers: { count: worker.bufferCount, descriptorBytes: worker.bufferBytes },
+    queue: {
+      submissions: worker.queueSubmissions,
+      commandBuffers: worker.commandBuffers,
+      fenceRequests: worker.queueFenceRequests,
+      fenceCompletions: worker.queueFenceCompletions,
+    },
+    cpuUnusedClaimed: false,
+    gpuOnlyClaimed: false,
+  });
+}
+
+function initialEvidence() {
+  return deepFreeze({
+    protocol: RUNTIME_EVIDENCE_PROTOCOL,
+    state: "unloaded",
+    webgpu: {
+      observed: false,
+      apiPresent: Boolean(globalThis.navigator?.gpu),
+    },
+    cancellation: null,
+    cleanup: null,
+  });
+}
+
 /**
- * Creates one packaged Wllama session. This factory has no network or browser
- * side effects until load() is called. Runtime URLs are fixed relative to this
- * module so the same bytes work from npm and a materialized /arcane/sdk tree.
+ * Creates one packaged, WebGPU-required Wllama session. 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.
  */
 export function createPackagedWllamaRuntime({ logger = console } = {}) {
   let engine = null;
   let pending = null;
+  let inferenceActive = false;
+  let evidenceState = initialEvidence();
   const trackedOperations = new Set();
   const sessionExitPromises = new WeakMap();
+  const sessionObservers = new WeakMap();
+
+  function publishEvidence(update) {
+    evidenceState = deepFreeze({
+      ...evidenceState,
+      ...update,
+      protocol: RUNTIME_EVIDENCE_PROTOCOL,
+    });
+    return evidenceState;
+  }
 
   function cancellationError(reason, fallback = "The Wllama operation was cancelled.") {
     return reason instanceof Error ? reason : new Error(reason ? String(reason) : fallback);
@@ -104,11 +436,11 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
     const cancellation = new Promise((_, reject) => {
       rejectCancellation = reject;
     });
-    let cancelled = false;
+    let locallySuppressed = false;
     const record = {
       cancel(reason) {
-        if (cancelled) return false;
-        cancelled = true;
+        if (locallySuppressed) return false;
+        locallySuppressed = true;
         rejectCancellation(cancellationError(reason));
         return true;
       },
@@ -118,102 +450,48 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
       trackedOperations.delete(record);
     });
     result.catch(() => undefined);
-    return Object.freeze({ raw, result, cancel: record.cancel });
+    return Object.freeze({
+      raw,
+      result,
+      cancel: record.cancel,
+      locallySuppressed: () => locallySuppressed,
+    });
   }
 
   function cancelTrackedOperations(reason) {
     for (const operation of [...trackedOperations]) operation.cancel(reason);
   }
 
-  function requireConfigurableDataProperty(object, key, label) {
-    const descriptor = Object.getOwnPropertyDescriptor(object, key);
-    if (!descriptor?.configurable || !("value" in descriptor) || descriptor.get || descriptor.set) {
-      throw new Error(`Pinned Wllama ${label} is not a configurable data property.`);
-    }
-    return descriptor;
-  }
-
-  function guardLoadingSession(session) {
-    const sessionDescriptor = requireConfigurableDataProperty(session, "proxy", "proxy");
-    let proxy = sessionDescriptor.value;
-    let cancelled = null;
-    const workerGuards = new Map();
-
-    function guardProxy(nextProxy) {
-      if (!nextProxy || workerGuards.has(nextProxy)) return;
-      const descriptor = requireConfigurableDataProperty(nextProxy, "worker", "proxy worker");
-      let worker = descriptor.value;
-      Object.defineProperty(nextProxy, "worker", {
-        enumerable: descriptor.enumerable,
-        configurable: true,
-        get: () => worker,
-        set(value) {
-          if (value && cancelled) {
-            value.terminate?.();
-            throw cancelled;
-          }
-          worker = value;
-        },
-      });
-      workerGuards.set(nextProxy, () => {
-        Object.defineProperty(nextProxy, "worker", { ...descriptor, value: worker });
-      });
-      if (worker && cancelled) {
-        worker.terminate?.();
-        throw cancelled;
-      }
+  function recordCleanup(snapshot) {
+    let cleanup = snapshot?.cleanup ?? null;
+    if (!["worker-terminated", "no-worker-observed-at-exit"].includes(cleanup?.kind)) {
+      throw runtimeFailure(
+        "ARCANE_AI_WORKER_TERMINATION_UNCONFIRMED",
+        "Wllama cleanup did not confirm Worker termination.",
+      );
     }
-
-    if (proxy) guardProxy(proxy);
-    Object.defineProperty(session, "proxy", {
-      enumerable: sessionDescriptor.enumerable,
-      configurable: true,
-      get: () => proxy,
-      set(value) {
-        if (value && cancelled) {
-          value.worker?.terminate?.();
-          throw cancelled;
-        }
-        proxy = value;
-        if (value) guardProxy(value);
-      },
-    });
-
-    return Object.freeze({
-      cancel(reason) {
-        cancelled ||= cancellationError(reason, "The Wllama model load was cancelled.");
-        const current = proxy;
-        current?.worker?.terminate?.();
-        try {
-          current?.abort?.(cancelled.message, "");
-        } catch {
-          // The tracked operation gate remains the stable cancellation result.
-        }
-        return cancelled;
-      },
-      restore() {
-        for (const restoreWorker of workerGuards.values()) restoreWorker();
-        workerGuards.clear();
-        Object.defineProperty(session, "proxy", { ...sessionDescriptor, value: proxy });
+    publishEvidence({
+      state: "unloaded",
+      webgpu: {
+        ...evidenceState.webgpu,
+        observed: false,
+        lastObservedOperational: evidenceState.webgpu?.observed === true
+          || evidenceState.webgpu?.lastObservedOperational === true,
       },
+      cleanup,
     });
+    return cleanup;
   }
 
-  function exitSession(session, reason) {
-    if (!session) return Promise.resolve(false);
+  function exitSession(session) {
+    if (!session) return Promise.resolve(null);
     let exitPromise = sessionExitPromises.get(session);
     if (!exitPromise) {
-      const attempt = Promise.resolve().then(() => {
-        try {
-          session.proxy?.abort?.(cancellationError(reason).message, "");
-        } catch {
-          // Public session.exit() still owns Worker termination.
-        }
-        return session.exit();
-      }).then(() => true);
+      const attempt = Promise.resolve().then(() => session.arcaneTerminate()).then((snapshot) => {
+        recordCleanup(snapshot);
+        return snapshot;
+      });
       exitPromise = attempt.catch((error) => {
-        // A failed cleanup attempt is not proof that this session is closed.
-        // Evict only this attempt so a later exit() can retry the same handle.
         if (sessionExitPromises.get(session) === exitPromise) {
           sessionExitPromises.delete(session);
         }
@@ -225,20 +503,31 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
   }
 
   function newEngine() {
+    const observer = createEvidenceLogger(logger);
     const next = new Wllama({ default: WASM_URL }, {
-      logger,
+      logger: observer.logger,
       allowOffline: true,
     });
     next.setCompat(null);
+    for (const method of ["arcaneLoadModel", "arcaneTelemetry", "arcaneTerminate"]) {
+      if (typeof next[method] !== "function") {
+        throw new Error(`The packaged Wllama projection is missing public ${method}().`);
+      }
+    }
     const resources = next.getWorkerResources();
     if (resources.compat !== false || resources.wasmPath !== WASM_URL || resources.jsPath) {
       throw new Error("The packaged Wllama resource projection was not exact.");
     }
+    sessionObservers.set(next, observer);
     return next;
   }
 
   function capabilities() {
-    return runtimeCapabilitySnapshot();
+    return runtimeCapabilitySnapshot(evidenceState);
+  }
+
+  function evidence() {
+    return evidenceState;
   }
 
   async function load(files, options = {}) {
@@ -248,21 +537,29 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
     if (typeof globalThis.WebAssembly !== "object") {
       throw new Error("WebAssembly is unavailable in this browser.");
     }
+    if (!globalThis.navigator?.gpu) {
+      throw runtimeFailure(
+        "ARCANE_AI_WEBGPU_REQUIRED",
+        "A WebGPU API is required, but navigator presence alone will not establish operational execution.",
+      );
+    }
     if (engine || pending) {
       throw new Error("The packaged Wllama runtime is already loaded or loading.");
     }
 
-    // Wllama defaults to a CDN compatibility runtime. Arcane never admits it.
     const next = newEngine();
-
     const threads = normalizePositiveInteger(options.threads, 1, { maximum: 64 });
     const contextTokens = normalizePositiveInteger(options.contextTokens, 4_096, {
       maximum: 1_048_576,
     });
+    if (options.gpuLayers !== undefined && options.gpuLayers !== FULL_GPU_LAYERS) {
+      throw new RangeError(`WebGPU-required Wllama must request exactly ${FULL_GPU_LAYERS} GPU layers.`);
+    }
+    const gpuLayers = FULL_GPU_LAYERS;
     const loadOptions = {
       n_threads: threads,
       n_ctx: contextTokens,
-      n_gpu_layers: Number.isSafeInteger(options.gpuLayers) ? options.gpuLayers : 0,
+      n_gpu_layers: gpuLayers,
     };
     if (options.batchTokens !== undefined) {
       loadOptions.n_batch = normalizePositiveInteger(options.batchTokens, 512, {
@@ -275,34 +572,56 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
       });
     }
 
-    const loadGuard = guardLoadingSession(next);
-    const operation = trackOperation(Promise.resolve().then(() => next.loadModel(files, loadOptions)));
-    operation.raw.finally(loadGuard.restore).catch(() => undefined);
+    publishEvidence({
+      state: "loading",
+      webgpu: { observed: false, apiPresent: true },
+      cancellation: null,
+      cleanup: null,
+    });
+    const loadController = new AbortController();
+    const loadOperation = Promise.resolve().then(() => (
+      next.arcaneLoadModel(files, loadOptions, loadController.signal)
+    ));
+    loadOperation.catch(() => undefined);
     const cancel = (reason) => {
-      const error = loadGuard.cancel(reason);
-      operation.cancel(error);
+      const error = cancellationError(reason, "The Wllama model load was cancelled.");
+      loadController.abort(error);
     };
-    pending = Object.freeze({
-      engine: next,
-      cancel,
-    });
+    pending = Object.freeze({ engine: next, cancel });
     const signal = options.signal ?? null;
     const onAbort = () => cancel(signal.reason);
     if (signal?.aborted) onAbort();
     else signal?.addEventListener?.("abort", onAbort, { once: true });
     try {
-      await operation.result;
+      await loadOperation;
       if (pending?.engine !== next) throw new Error("Wllama load was cancelled.");
+      if (typeof next.isModelLoaded !== "function" || next.isModelLoaded() !== true) {
+        throw runtimeFailure(
+          "ARCANE_AI_LOAD_FAILED",
+          "Wllama did not confirm a successfully loaded model.",
+        );
+      }
+      const projected = verifyProjectedTelemetry(await next.arcaneTelemetry());
+      const webgpu = admittedLoadEvidence(sessionObservers.get(next).snapshot(), projected);
       pending = null;
       engine = next;
+      publishEvidence({ state: "ready", webgpu, cancellation: null, cleanup: null });
     } catch (error) {
+      let cleanupFailure = null;
       try {
-        await exitSession(next, error);
+        await exitSession(next);
         if (pending?.engine === next) pending = null;
-      } catch {
-        // Preserve the exact pending session for a later runtime.exit() retry.
-        // The original load failure remains the stable public rejection.
+      } catch (cleanupError) {
+        cleanupFailure = cleanupError?.code === "ARCANE_AI_WORKER_TERMINATION_UNCONFIRMED"
+          ? cleanupError
+          : runtimeFailure(
+            "ARCANE_AI_WORKER_TERMINATION_UNCONFIRMED",
+            "Wllama model load failed and Worker termination could not be proved.",
+            cleanupError,
+          );
+        // Preserve the pending handle if Worker termination could not be proved.
       }
+      if (cleanupFailure) throw cleanupFailure;
       throw error;
     } finally {
       signal?.removeEventListener?.("abort", onAbort);
@@ -312,32 +631,261 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
       loaded: true,
       contextTokens,
       threads,
+      gpuLayers,
+      evidence: evidenceState,
       metadata: engine.getModelMetadata?.() ?? null,
     });
   }
 
   function assertLoaded() {
-    if (!engine?.isModelLoaded?.()) throw new Error("The packaged Wllama model is not loaded.");
+    if (!engine?.isModelLoaded?.() || evidenceState.state !== "ready") {
+      throw new Error("The packaged Wllama model is not loaded with admitted WebGPU evidence.");
+    }
     return engine;
   }
 
-  async function chat(options) {
-    const session = assertLoaded();
-    return trackOperation(
-      Promise.resolve().then(() => session.createChatCompletion({ ...options, stream: false })),
-    ).result;
+  async function terminateUnacknowledgedCancellation(session, reason) {
+    let snapshot;
+    try {
+      snapshot = await exitSession(session);
+    } catch (error) {
+      throw runtimeFailure(
+        "ARCANE_AI_WORKER_TERMINATION_UNCONFIRMED",
+        "Wllama cancellation was not acknowledged and Worker termination could not be proved.",
+        error,
+      );
+    }
+    if (engine === session) engine = null;
+    publishEvidence({
+      cancellation: deepFreeze({
+        deliverySuppressed: true,
+        upstream: {
+          kind: "worker-terminated",
+          cancellationAcknowledged: false,
+          cleanup: snapshot.cleanup,
+        },
+        nativeUnloadClaimed: false,
+        physicalVramReclamationClaimed: false,
+      }),
+    });
+  }
+
+  async function recordInference(
+    session,
+    before,
+    { aborted, requireCancellationAcknowledgement = false },
+  ) {
+    let after;
+    try {
+      after = verifyProjectedTelemetry(await session.arcaneTelemetry());
+    } catch (error) {
+      if (!aborted && !requireCancellationAcknowledgement) throw error;
+      await terminateUnacknowledgedCancellation(session, error);
+      if (requireCancellationAcknowledgement) {
+        throw runtimeFailure(
+          "ARCANE_AI_COMPLETION_RECOVERY_UNCONFIRMED",
+          "Wllama completion recovery could not prove request settlement.",
+          error,
+        );
+      }
+      return;
+    }
+    const previousSequence = before?.cancellation?.sequence ?? 0;
+    const cancellation = after.cancellation;
+    const cancellationAcknowledged = cancellation?.sequence > previousSequence
+      && cancellation.responseName === "cncl_res"
+      && cancellation.acknowledged === true
+      && cancellation.failed === false;
+    if (aborted) {
+      if (cancellationAcknowledged) {
+        publishEvidence({
+          cancellation: deepFreeze({
+            deliverySuppressed: true,
+            upstream: {
+              kind: "llama-request-cancel-acknowledged",
+              sequence: cancellation.sequence,
+              requestId: cancellation.requestId,
+              responseName: cancellation.responseName,
+              acknowledged: true,
+              failed: false,
+            },
+            immediateGpuKernelPreemptionClaimed: false,
+          }),
+        });
+        return;
+      }
+      await terminateUnacknowledgedCancellation(
+        session,
+        "Wllama cancellation was not acknowledged.",
+      );
+      return;
+    }
+    if (requireCancellationAcknowledgement && !cancellationAcknowledged) {
+      await terminateUnacknowledgedCancellation(
+        session,
+        "Wllama completion recovery could not prove cancellation acknowledgement.",
+      );
+      throw runtimeFailure(
+        "ARCANE_AI_COMPLETION_RECOVERY_UNCONFIRMED",
+        "Wllama completion recovery could not prove request settlement.",
+      );
+    }
+
+    const submissions = after.worker.queueSubmissions - before.worker.queueSubmissions;
+    const commandBuffers = after.worker.commandBuffers - before.worker.commandBuffers;
+    const fenceRequests = after.worker.queueFenceRequests - before.worker.queueFenceRequests;
+    const fenceCompletions = after.worker.queueFenceCompletions - before.worker.queueFenceCompletions;
+    if (
+      submissions < 1
+      || commandBuffers < 1
+      || fenceRequests < 1
+      || fenceCompletions < fenceRequests
+    ) {
+      await exitSession(session);
+      if (engine === session) engine = null;
+      throw runtimeFailure(
+        "ARCANE_AI_WEBGPU_REQUIRED",
+        "Inference completed without positive, settled WebGPU queue evidence.",
+      );
+    }
+    publishEvidence({
+      webgpu: deepFreeze({
+        ...evidenceState.webgpu,
+        queue: {
+          submissions: after.worker.queueSubmissions,
+          commandBuffers: after.worker.commandBuffers,
+          fenceRequests: after.worker.queueFenceRequests,
+          fenceCompletions: after.worker.queueFenceCompletions,
+        },
+        lastInference: { submissions, commandBuffers, fenceRequests, fenceCompletions },
+      }),
+      cancellation: requireCancellationAcknowledgement
+        ? deepFreeze({
+          deliverySuppressed: false,
+          recovery: "peg-native-final-output",
+          upstream: {
+            kind: "llama-request-cancel-acknowledged",
+            sequence: cancellation.sequence,
+            requestId: cancellation.requestId,
+            responseName: cancellation.responseName,
+            acknowledged: true,
+            failed: false,
+          },
+          immediateGpuKernelPreemptionClaimed: false,
+        })
+        : null,
+    });
+  }
+
+  async function invalidateFatalSession(session, error) {
+    try {
+      await exitSession(session);
+    } catch (cleanupError) {
+      throw cleanupError?.code === "ARCANE_AI_WORKER_TERMINATION_UNCONFIRMED"
+        ? cleanupError
+        : runtimeFailure(
+          "ARCANE_AI_WORKER_TERMINATION_UNCONFIRMED",
+          "Wllama failed and Worker termination could not be proved.",
+          cleanupError,
+        );
+    }
+    if (engine === session) engine = null;
+    publishEvidence({
+      state: "error",
+      webgpu: {
+        ...evidenceState.webgpu,
+        observed: false,
+        lastObservedOperational: evidenceState.webgpu?.lastObservedOperational === true,
+      },
+      failure: deepFreeze({
+        code: typeof error?.code === "string" ? error.code : "ARCANE_AI_RUNTIME_FAILED",
+      }),
+    });
+  }
+
+  async function inference(options, onData = null) {
+    if (!options || typeof options !== "object" || Array.isArray(options)) {
+      throw new TypeError("Wllama inference options must be an object.");
+    }
+    if (Object.hasOwn(options, "signal")) {
+      throw new TypeError("Pinned Wllama 3.6.0 accepts abortSignal, not signal.");
+    }
+    if (inferenceActive) {
+      throw runtimeFailure(
+        "ARCANE_AI_RUNTIME_BUSY",
+        "The packaged Wllama runtime admits one inference at a time.",
+      );
+    }
+    inferenceActive = true;
+    try {
+      const session = assertLoaded();
+      const observer = sessionObservers.get(session);
+      const streamCapture = createStructuredStreamCapture();
+      let capture = null;
+      let operation = null;
+      let before = null;
+      try {
+        before = verifyProjectedTelemetry(await session.arcaneTelemetry());
+        const abortSignal = options.abortSignal ?? null;
+        capture = observer.beginCompletionCapture();
+        const deliver = onData
+          ? (chunk) => {
+            streamCapture.observe(chunk);
+            onData(chunk);
+          }
+          : null;
+        operation = trackOperation(Promise.resolve().then(() => session.createChatCompletion({
+          ...options,
+          stream: Boolean(onData),
+          ...(deliver ? { onData: deliver } : {}),
+          abortSignal,
+        })));
+        const result = await operation.result;
+        capture.release();
+        await recordInference(session, before, { aborted: false });
+        return result;
+      } catch (error) {
+        const abortSignal = options.abortSignal ?? null;
+        const locallySuppressed = operation?.locallySuppressed() === true;
+        const aborted = operation !== null
+          && !locallySuppressed
+          && (abortSignal?.aborted || error?.name === "AbortError");
+        const recoveredContent = locallySuppressed
+          ? null
+          : capture?.recover(error, { aborted }) ?? null;
+        if (onData && recoveredContent !== null && streamCapture.matches(recoveredContent)) {
+          // Wllama has already awaited cancelRequest() before rejecting here.
+          // The streamed text is exact, but the native stop reason is unknown.
+          await recordInference(session, before, {
+            aborted: false,
+            requireCancellationAcknowledgement: true,
+          });
+          return null;
+        }
+        if (aborted) {
+          await operation.raw.catch(() => undefined);
+          await recordInference(session, before, { aborted: true });
+        } else if (operation === null || !locallySuppressed) {
+          await invalidateFatalSession(session, error);
+        }
+        throw error;
+      } finally {
+        capture?.release();
+      }
+    } finally {
+      inferenceActive = false;
+    }
+  }
+
+  function chat(options) {
+    return inference({ ...options, stream: false });
   }
 
-  async function stream(options, onData) {
+  function stream(options, onData) {
     if (typeof onData !== "function") {
       throw new TypeError("Wllama stream() requires an onData callback.");
     }
-    // Wllama owns the request slot until this promise settles. Its documented
-    // abortSignal path cancels the llama.cpp request in a finally block.
-    const session = assertLoaded();
-    return trackOperation(
-      Promise.resolve().then(() => session.createChatCompletion({ ...options, stream: true, onData })),
-    ).result;
+    return inference({ ...options, stream: true }, onData);
   }
 
   async function exit() {
@@ -348,10 +896,7 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
     cancelTrackedOperations(reason);
     const sessions = new Set([current, loading?.engine].filter(Boolean));
     if (!sessions.size) return false;
-    await Promise.all([...sessions].map((session) => exitSession(session, reason)));
-    // Retain ownership through cleanup failure. Clear only the exact handles
-    // whose exit attempts completed successfully; replacement sessions, if
-    // any, remain owned by the runtime.
+    await Promise.all([...sessions].map((session) => exitSession(session)));
     if (engine === current) engine = null;
     if (pending === loading) pending = null;
     return true;
@@ -363,14 +908,13 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
       throw new TypeError("Wllama probe args must be an array of strings.");
     }
     const temporary = newEngine();
-    // testBackendOps owns and closes its temporary worker. It requires browser
-    // multithreading, but that is a probe capability—not model admission.
     const result = await temporary.testBackendOps(args);
     return Object.freeze({
       ...result,
       args: Object.freeze([...args]),
       origin: globalThis.location?.origin ?? null,
       capabilities: capabilities(),
+      evidence: evidenceState,
       runtime: BROWSER_WASM_RUNTIME_AUTHORITY,
     });
   }
@@ -379,12 +923,13 @@ export function createPackagedWllamaRuntime({ logger = console } = {}) {
     authority: BROWSER_WASM_RUNTIME_AUTHORITY,
     runtimeAssets: BROWSER_WASM_RUNTIME_AUTHORITY.runtimeAssets,
     capabilities,
+    evidence,
     load,
     chat,
     stream,
     probe,
     exit,
-    isLoaded: () => Boolean(engine?.isModelLoaded?.()),
+    isLoaded: () => Boolean(engine?.isModelLoaded?.()) && evidenceState.state === "ready",
     isLoading: () => Boolean(pending),
   });
 }