space-data-module-sdk 0.8.10 → 0.8.12

package/src/host/isomorphicLoader.js

@@ -21,12 +21,14 @@
 import {
   createBrowserModuleHarness,
   toLoadableWasmBytes,
+  zeroWasmBytes,
 } from "./browserModuleHarness.js";
 import {
   resolveModuleSignaturePolicy,
   verifyModuleArtifact,
 } from "../bundle/signing.js";
 import { attachHostDispatch, inspectModule } from "./isomorphicLoaderCore.js";
+import { assertArtifactRuntimeTarget } from "./runtimeTargetGate.js";
 
 export { inspectModule, attachHostDispatch };
 
@@ -203,21 +205,65 @@ export async function loadModule(options = {}) {
     );
   }
 
-  if (signaturePolicy) {
+  const runtimeKind = options.runtimeKind ?? "wasmedge";
+
+  // ONE READ of the artifact, serving every question this function asks of the
+  // bytes: the signature, the runtime-target declaration, and the profile
+  // inspection below. Each of those used to open the file for itself. The
+  // plaintext buffer this function materialized is scrubbed the moment the
+  // last of them is answered, matching the browser harness's
+  // ownedArtifactBytes contract.
+  const runtimeHostRequested =
+    String(options.hostProfile ?? "").trim().toLowerCase() === "runtime-host" ||
+    Array.isArray(options.modules) ||
+    (typeof options.defaultModuleId === "string" &&
+      options.defaultModuleId.trim().length > 0);
+  // The profile inspection is the ONLY consumer here that needs a compiled
+  // module. The runtime-host path does not take it, and it must not start
+  // paying a V8 compile of the whole artifact on every load just because the
+  // runtime-target gate arrived — the gate reads a custom section, which is a
+  // byte-level question.
+  const needsCompiledModule =
+    runtimeKind === "wasmedge" &&
+    !runtimeHostRequested &&
+    !options.wasmEdgeRunnerBinary;
+  const needsBytes = Boolean(signaturePolicy) || runtimeKind === "wasmedge";
+  let artifactBytes = null;
+  let wasmModule = null;
+  let embeddedManifest = null;
+  if (needsBytes) {
     const { readFile } = await import("node:fs/promises");
-    await verifyModuleArtifact(await readFile(source), signaturePolicy);
+    artifactBytes = new Uint8Array(await readFile(source));
+    if (signaturePolicy) {
+      await verifyModuleArtifact(artifactBytes, signaturePolicy);
+    }
+    if (runtimeKind === "wasmedge") {
+      const { locateEmbeddedPlgManifest } = await import(
+        "../compliance/index.js"
+      );
+      embeddedManifest =
+        locateEmbeddedPlgManifest(artifactBytes)?.decoded ?? null;
+    }
+    if (needsCompiledModule) {
+      wasmModule = await WebAssembly.compile(toLoadableWasmBytes(artifactBytes));
+    }
+    zeroWasmBytes(artifactBytes);
+    artifactBytes = null;
   }
 
-  const runtimeKind = options.runtimeKind ?? "wasmedge";
   if (runtimeKind === "wasmedge") {
-    const runtimeHostRequested =
-      String(options.hostProfile ?? "").trim().toLowerCase() === "runtime-host" ||
-      Array.isArray(options.modules) ||
-      (typeof options.defaultModuleId === "string" &&
-        options.defaultModuleId.trim().length > 0);
-    if (!runtimeHostRequested && !options.wasmEdgeRunnerBinary) {
-      const { readFile } = await import("node:fs/promises");
-      const inspection = await inspectModule(await readFile(source));
+    // MIRROR OF THE BROWSER GATE. Enforcing the declaration on one leg only is
+    // itself a divergence between the legs: a `["browser"]`-only artifact
+    // handed to WasmEdge would run until it reached something WasmEdge does
+    // not serve, while the reverse case is refused at the door. Both legs read
+    // the same declaration and refuse the same way.
+    assertArtifactRuntimeTarget({
+      embeddedManifest,
+      manifest: options.manifest,
+      leg: "wasmedge",
+    });
+    if (needsCompiledModule) {
+      const inspection = await inspectModule(wasmModule);
       if (
         (inspection.profile === "standalone" || inspection.profile === "module-host-abi") &&
         inspection.exports.includes("_start")

package/src/host/runtimeTargetGate.js

@@ -0,0 +1,256 @@
+/**
+ * THE RUNTIME-TARGET GATE — one assert, called by every loader that stands for
+ * a runtime leg.
+ *
+ * Composed flows derive their `runtimeTargets` from their parts (see
+ * `collectFlowRuntimeTargets` in `src/flow/flowCompiler.js`), so an artifact
+ * that legitimately runs on only ONE leg now exists. That artifact must never
+ * be loaded QUIETLY on a leg it does not declare: the failure would surface as
+ * a hostcall trap several frames deep in a capability that leg cannot serve,
+ * which is indistinguishable from the cross-runtime divergence the isomorphism
+ * invariant exists to catch. Refuse at the door, by name, quoting the
+ * declaration that made it ineligible.
+ *
+ * The gate lives here rather than inside one harness because there are two
+ * doors into a composed artifact — `createBrowserModuleHarness` (a single
+ * module) and `createFlowRuntimeHost` (the composed flow runtime) — and a gate
+ * on only one of them is not a gate.
+ *
+ * Browser-safe: no node builtins, no bundle/signing surface, and the embedded
+ * declaration is read with `WebAssembly.Module.customSections`, which needs no
+ * copy of the source bytes (they are scrubbed by the time this runs).
+ */
+
+import { SDS_MANIFEST_SECTION_NAME } from "../bundle/constants.js";
+import { decodePlgManifest, isPlgManifestBuffer } from "../manifest/plgCodec.js";
+import { LegIncompatibleCapabilityIds } from "../capabilities.js";
+
+/**
+ * `wasi` is the STRICT PORTABILITY BASELINE, not a fourth runtime: compliance
+ * confines a `wasi`-declaring artifact to the pure WASI capability subset and
+ * the `command` invoke surface. So a declaration of `wasi` admits either leg —
+ * WITH ONE EXCEPTION, which is why capabilities are part of this question.
+ * `pipe` is in both the standalone-WASI subset and the browser-incompatible
+ * set, so `runtimeTargets:["wasi"] + capabilities:["pipe"]` passes compliance
+ * (it names no browser target for the rule to fire on) and would otherwise be
+ * admitted to the browser leg against the SDK's own policy. The baseline
+ * admits a leg only when the artifact carries nothing that leg cannot serve.
+ *
+ * DO NOT "UNIFY" THIS WITH THE LITERAL PATH. A declaration that names the leg
+ * outright is NOT capability-checked here, and that asymmetry is deliberate:
+ *
+ *   declaration -> trust the author;  inference -> prove it.
+ *
+ * An explicit `browser` target is the author's statement and the embedder's
+ * business — five of the browser-incompatible capabilities (`network`, `ipfs`,
+ * `protocol_dial`, `protocol_handle`, `wallet_sign`) ARE served by
+ * `BrowserHost` when the embedder injects an adapter, so refusing them at load
+ * time would break legitimate embedders. Publish-time compliance already
+ * refuses the incoherent declaration. The `wasi` path, by contrast, is an
+ * inference the SDK makes on the author's behalf, and an inference must be
+ * conservative.
+ */
+const LEG_SATISFYING_TARGETS = Object.freeze({
+  browser: Object.freeze(["browser", "wasi"]),
+  wasmedge: Object.freeze(["wasmedge", "wasi"]),
+});
+
+const WASI_PORTABILITY_TARGET = "wasi";
+
+function legIncompatibleCapabilities(capabilities, leg) {
+  const incompatible = LegIncompatibleCapabilityIds[leg];
+  if (!incompatible || !Array.isArray(capabilities)) return [];
+  return capabilities
+    .map((capability) =>
+      typeof capability === "string"
+        ? capability
+        : (capability?.capabilityId ?? capability?.name ?? ""),
+    )
+    .filter((capability) => incompatible.includes(capability))
+    .sort();
+}
+
+/**
+ * Does a declared target set admit this leg?
+ *
+ * ONE RULE, used by the loaders AND by the flow compiler's derivation. Two
+ * rules for one field is how a manifest comes to mean different things to the
+ * compiler and to the door: the compiler stripped the browser leg from a
+ * `["wasi","wasmedge"]` part while the gate happily admitted the same
+ * declaration to the browser.
+ *
+ * An empty/absent declaration admits everything, matching compliance, which
+ * skips the runtime-target rule on an absent field.
+ *
+ * @param {string[]|undefined} targets declared runtimeTargets
+ * @param {string} leg the runtime target being asked about
+ * @param {Array<string|{capabilityId?: string, name?: string}>} [capabilities]
+ *   the artifact's declared capabilities; consulted only to keep the `wasi`
+ *   baseline from admitting a leg that cannot serve them
+ */
+export function runtimeTargetSatisfies(targets, leg, capabilities) {
+  if (!Array.isArray(targets) || targets.length === 0) return true;
+  const accepted = LEG_SATISFYING_TARGETS[leg] ?? [leg];
+  if (targets.includes(leg)) return true;
+  if (!targets.some((target) => accepted.includes(target))) return false;
+  // Admitted only via the portability baseline — prove it, do not assume it.
+  return (
+    targets.includes(WASI_PORTABILITY_TARGET) &&
+    legIncompatibleCapabilities(capabilities, leg).length === 0
+  );
+}
+
+export class RuntimeTargetError extends Error {
+  constructor(message, { declaredTargets, source, leg }) {
+    super(message);
+    this.name = "RuntimeTargetError";
+    this.code = "runtime-target-out-of-scope";
+    this.declaredTargets = declaredTargets;
+    this.leg = leg;
+    // "embedded" (the artifact's own signed record) or "caller" (the manifest
+    // the loader was handed). Which one refused is evidence, not trivia.
+    this.declarationSource = source;
+  }
+}
+
+export function normalizedRuntimeTargets(manifest) {
+  return Array.isArray(manifest?.runtimeTargets)
+    ? manifest.runtimeTargets
+        .map((target) => String(target ?? "").trim().toLowerCase())
+        .filter(Boolean)
+    : [];
+}
+
+/**
+ * The artifact's OWN declaration, read straight out of the embedded `$PLG`
+ * custom section, so a caller that hands a loader nothing but bytes still gets
+ * the refusal.
+ *
+ * @param {WebAssembly.Module} wasmModule
+ * @returns {string[]} declared targets, or [] when the artifact declares none
+ */
+export function embeddedPlgManifest(wasmModule) {
+  let sections;
+  try {
+    sections = WebAssembly.Module.customSections(
+      wasmModule,
+      SDS_MANIFEST_SECTION_NAME,
+    );
+  } catch {
+    return null;
+  }
+  for (const section of sections ?? []) {
+    try {
+      const bytes = new Uint8Array(section);
+      if (!isPlgManifestBuffer(bytes)) continue;
+      // FIRST DECODABLE MANIFEST WINS — the same rule the node-side locator
+      // (`locateEmbeddedPlgManifest`) applies. Skipping a manifest merely
+      // because its runtimeTargets are empty made this reader disagree with
+      // that one on the very same bytes, so an artifact could be scoped by one
+      // declaration and enforced against another.
+      return decodePlgManifest(bytes);
+    } catch {
+      // An UNDECODABLE section is not a declaration. Keep looking.
+    }
+  }
+  return null;
+}
+
+export function embeddedRuntimeTargets(wasmModule) {
+  return normalizedRuntimeTargets(embeddedPlgManifest(wasmModule));
+}
+
+/**
+ * Resolve which declaration refuses this leg, if either does.
+ *
+ * EITHER source can refuse, and the EMBEDDED one wins on conflict. The
+ * embedded `$PLG` is what the artifact's signature covers; `manifest` is
+ * caller input. Trusting the caller first would make
+ * `{manifest: {runtimeTargets: ["browser"]}}` a bypass for a signed
+ * WasmEdge-only artifact — and the isomorphic flow host passes exactly such a
+ * caller-side manifest for every child it mounts.
+ *
+ * An artifact that declares NOTHING is unconstrained, matching compliance,
+ * which skips the runtime-target rule on an absent field.
+ *
+ * @returns {{targets: string[], source: "embedded"|"caller"}|null}
+ */
+export function resolveRuntimeTargetRefusal({
+  wasmModule,
+  manifest,
+  leg,
+  embeddedManifest: providedEmbeddedManifest,
+}) {
+  // A caller that already located the artifact's own manifest from BYTES (the
+  // node leg does, via `locateEmbeddedPlgManifest`) passes it here rather than
+  // making this gate compile the module just to read a custom section. The two
+  // locators apply the same first-decodable-wins rule; the byte-side one can
+  // additionally reach a manifest carried in a bundle entry or a data segment,
+  // which only ever makes that leg MORE likely to refuse — the safe direction.
+  const embeddedManifest =
+    providedEmbeddedManifest ??
+    (wasmModule ? embeddedPlgManifest(wasmModule) : null);
+  const embedded = normalizedRuntimeTargets(embeddedManifest);
+  if (!runtimeTargetSatisfies(embedded, leg, embeddedManifest?.capabilities)) {
+    return { targets: embedded, source: "embedded" };
+  }
+  const caller = normalizedRuntimeTargets(manifest);
+  if (!runtimeTargetSatisfies(caller, leg, manifest?.capabilities)) {
+    return { targets: caller, source: "caller" };
+  }
+  return null;
+}
+
+/**
+ * @param {object} options
+ * @param {WebAssembly.Module} [options.wasmModule] compiled artifact
+ * @param {object} [options.manifest] caller-supplied manifest
+ * @param {string} options.leg the runtime target this loader stands for
+ * @param {string} [options.what] noun used in the message ("module" by default)
+ */
+export function assertArtifactRuntimeTarget({
+  wasmModule,
+  manifest,
+  embeddedManifest,
+  leg,
+  what = "module",
+}) {
+  const refusal = resolveRuntimeTargetRefusal({
+    wasmModule,
+    manifest,
+    embeddedManifest,
+    leg,
+  });
+  if (!refusal) return;
+  const { targets, source } = refusal;
+  throw new RuntimeTargetError(
+    `Refusing to load a ${what} that does not target "${leg}": its ` +
+      `${source === "embedded" ? "embedded manifest" : "supplied manifest"} declares ` +
+      `runtimeTargets [${targets.join(", ")}], which does not admit this leg. Loading it ` +
+      `here would trap on the first capability the "${leg}" host cannot serve. Run it on ` +
+      "a host that provides one of its declared targets, or split the out-of-scope " +
+      "surface into its own module.",
+    { declaredTargets: targets, source, leg },
+  );
+}
+
+/**
+ * Is this JavaScript running in a browser (window or a Worker)? Used only to
+ * DEFAULT the leg for loaders that are genuinely runtime-agnostic; every
+ * caller may state its leg explicitly and that always wins.
+ */
+export function detectBrowserLeg() {
+  try {
+    if (typeof window !== "undefined" && typeof window.document !== "undefined") {
+      return true;
+    }
+    const workerScope = globalThis.WorkerGlobalScope;
+    return (
+      typeof workerScope === "function" &&
+      typeof self !== "undefined" &&
+      self instanceof workerScope
+    );
+  } catch {
+    return false;
+  }
+}

package/src/host/workerModuleHarness.js

@@ -27,6 +27,7 @@ import {
 } from "./sabHostcallChannel.js";
 import { WASI_THREAD_HOSTCALL_MESSAGE } from "./wasiThreadWorkerRuntime.js";
 import { importNodeBuiltin } from "./nodeBuiltinSpecifier.js";
+import { assertBrowserRuntimeTarget } from "./browserModuleHarness.js";
 
 const WORKER_URL = new URL("./workerModuleHarnessWorker.js", import.meta.url);
 
@@ -130,6 +131,12 @@ async function toWasmModule(source, label) {
  */
 export async function createWorkerModuleHarness(options = {}) {
   const wasmModule = await toWasmModule(options.wasmSource, "wasmSource");
+  // Fail on THIS thread, before a worker, a SAB channel and a BroadcastChannel
+  // are stood up for an artifact the browser leg must not run. The in-worker
+  // createBrowserModuleHarness would refuse it too, but a rejection that
+  // arrives after the whole rig exists reads like a worker fault rather than
+  // the declaration it actually is.
+  assertBrowserRuntimeTarget(wasmModule, options.harnessOptions?.manifest);
   const host = options.host ?? createBrowserHost(options.hostOptions);
   const dispatch = options.dispatchHost ?? createAsyncHostDispatcher(host);
   const buffer = createSabHostcallBuffer({

package/src/index.d.ts

@@ -2337,8 +2337,55 @@ export interface FlowRuntimeHostOptions {
   engineLink?: {
     exports: WebAssembly.Exports & { memory: WebAssembly.Memory };
   } | null;
+  /**
+   * The runtime leg this host stands for. A composed flow DERIVES its
+   * runtimeTargets from its parts, so a single-leg artifact exists and must be
+   * refused by name on any other leg. Omit and a real browser is detected and
+   * gated; pass `null` to opt out explicitly.
+   */
+  runtimeTarget?: string | null;
+  /** Caller-side manifest; the artifact's embedded record wins on conflict. */
+  manifest?: PluginManifest | Record<string, unknown>;
+}
+
+export declare class RuntimeTargetError extends Error {
+  code: "runtime-target-out-of-scope";
+  declaredTargets: string[];
+  leg: string;
+  declarationSource: "embedded" | "caller";
 }
 
+/**
+ * Does a declared target set admit this leg? `wasi` is the portability
+ * baseline and admits either leg, but only when `capabilities` carries nothing
+ * that leg cannot serve.
+ */
+export function runtimeTargetSatisfies(
+  targets: string[] | undefined,
+  leg: string,
+  capabilities?: Array<string | { capabilityId?: string; name?: string }>,
+): boolean;
+
+/** The artifact's own declared targets, read from its embedded `$PLG`. */
+export function embeddedRuntimeTargets(wasmModule: WebAssembly.Module): string[];
+
+/**
+ * Which declaration refuses this leg, if either does. The artifact's embedded
+ * record wins over a caller-supplied manifest.
+ */
+export function resolveRuntimeTargetRefusal(options: {
+  wasmModule?: WebAssembly.Module;
+  manifest?: PluginManifest | Record<string, unknown>;
+  embeddedManifest?: PluginManifest | Record<string, unknown> | null;
+  leg: string;
+}): { targets: string[]; source: "embedded" | "caller" } | null;
+
+/** Throws a RuntimeTargetError when the artifact does not target the browser. */
+export function assertBrowserRuntimeTarget(
+  wasmModule: WebAssembly.Module,
+  manifest?: PluginManifest | Record<string, unknown>,
+): void;
+
 export const FLOW_INVALID_INDEX: number;
 export function createFlowRuntimeHost(
   options: FlowRuntimeHostOptions,

package/src/index.js

@@ -16,6 +16,17 @@ export * from "./deployment/index.js";
 export * from "./app/index.js";
 export { FLOW_INVALID_INDEX, createFlowRuntimeHost } from "./flow/flowRuntimeHost.js";
 export { createIsomorphicFlowRuntimeHost } from "./flow/isomorphicFlowHost.js";
+// The runtime-target gate: a composed flow derives its runtimeTargets, so a
+// single-leg artifact exists and every loader refuses one that is not its own.
+// Exported so a consumer can catch the refusal by CLASS rather than by
+// matching a message string, and can ask the same question before offering a
+// module to a leg.
+export {
+  RuntimeTargetError,
+  runtimeTargetSatisfies,
+  resolveRuntimeTargetRefusal,
+  embeddedRuntimeTargets,
+} from "./host/runtimeTargetGate.js";
 export {
   DefaultInvokeExports,
   DefaultManifestExports,

package/src/standards/catalogCore.js

@@ -62,8 +62,35 @@ function collectFlatbufferTableFields(idl, tableName) {
     .filter(Boolean);
 }
 
+// A schema has TWO written forms in this ecosystem and they mean the same file:
+//
+//   "SCV.fbs"       — the catalog's canonical form, built as `<CODE>.fbs` from
+//                     the standards manifest (parseStandardsEntry below).
+//   "SCV/main.fbs"  — the PATH form, which is how the schema actually lives in
+//                     spacedatastandards.org (`schema/SCV/main.fbs`) and is
+//                     therefore what modules, tests and generated bindings across
+//                     the stack write.
+//
+// Matching them as raw strings made every manifest using the path form fail
+// resolution with `standards-type-identity-mismatch` — "mixes a known
+// schemaName ... with a different standards entry" — because the `$SCV` file
+// identifier DID resolve while the name did not. That is not a manifest defect
+// and not a schema defect; it is this matcher refusing a spelling the standards
+// repository itself uses. The rest of the toolchain already treats the two as
+// one schema (see the closed-modules builder's
+// `^([A-Z][A-Z0-9]{2})(?:\/main)?\.fbs$` header-injection regex), so the
+// catalog must too, or a correct module cannot be compiled.
+//
+// Deliberately narrow: ONLY the exact `<CODE>/main.fbs` shape collapses. Any
+// other path is left alone so two genuinely different schemas can never be
+// merged by a loose rule.
+const SCHEMA_PATH_FORM = /^([A-Za-z][A-Za-z0-9_]*)\/main\.fbs$/;
+
 function normalizeSchemaName(value) {
-  return value === undefined || value === null ? "" : String(value);
+  if (value === undefined || value === null) return "";
+  const text = String(value);
+  const pathForm = SCHEMA_PATH_FORM.exec(text);
+  return pathForm ? `${pathForm[1]}.fbs` : text;
 }
 
 function normalizeFileIdentifier(value) {

package/src/testing/parityBrowserRunner.js

@@ -17,6 +17,7 @@ import { createBrowserModuleHarness } from "../host/browserModuleHarness.js";
 const OK = "ok";
 const GUEST_ERROR = "guest-error";
 const TRAP = "trap";
+const OUT_OF_SCOPE = "out-of-declared-scope";
 
 function base64ToBytes(value) {
   const binary = atob(String(value ?? ""));
@@ -42,6 +43,16 @@ function classifyBrowserError(error) {
   if (error?.name === "WasiExitError") {
     return { exitClass: GUEST_ERROR, exitDetail: `exit=${error.code}` };
   }
+  // The harness refused the artifact because the ARTIFACT declares it does not
+  // run here. That is the contract working. Classing it TRAP would score a
+  // correct refusal as a P1 cross-runtime divergence — the gate failing the
+  // very artifacts the compiler now legitimately emits.
+  if (error?.name === "RuntimeTargetError") {
+    return {
+      exitClass: OUT_OF_SCOPE,
+      exitDetail: `declared runtimeTargets [${(error.declaredTargets ?? []).join(", ")}]`,
+    };
+  }
   return {
     exitClass: TRAP,
     exitDetail: `${error?.name ?? "Error"}: ${error?.message ?? String(error)}`,