arcane-os 0.2.2 → 0.3.0

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

@@ -1,97 +1,228 @@
 # Browser speech providers
 
-`arcane-os/ai/browser-speech` is the shipped browser-only provider package for
-caller-supplied Whisper speech-to-text and Kokoro text-to-speech runtimes. It
-implements the same provider-neutral role contract used by
-`AIProviderRuntime.js`; it does not ship model weights, runtime adapter bytes,
-voices, download URLs, a catalog, or a cloud fallback.
+`arcane-os/ai/browser-speech` is the browser-only, event-neutral SDK boundary
+for caller-supplied Whisper speech-to-text and Kokoro text-to-speech runtimes.
+It supplies immutable artifact-graph validation, DBOPFS admission, role Workers,
+provider/2 adapters, audio normalization, cancellation, and cleanup. It does
+not ship or select runtime modules, model weights, voice bytes, a default
+catalog, credentials, a native provider, or a cloud fallback.
+The package contains no speech runtime, model, or voice payload and
+never downloads one before explicit `load()`.
 
-Use this page when an application deliberately owns browser-local speech model
-and runtime selection. For ordinary application AI lifecycle and status, start
-with the [normalized AI guide](../README.md#normalized-ai) and
-[`AIProviderRuntime.js`](../runtime-modules.md#aiproviderruntimejs).
+Use this entrypoint when an application deliberately owns every browser-local
+speech choice. Higher-level selection, startup policy, shared state, and event
+projection belong to the application runtime; this package exposes only
+promises, `AbortSignal`, and the caller's progress callback.
 
 ## Availability
 
 | Host | Availability | Notes |
 | --- | --- | --- |
-| Browser | Shipped | Requires DBOPFS/OPFS, Web Locks, Workers, Fetch or an injected fetch function, Blob/File, and object URLs. |
-| Native WebView | Conditional | Works only when the WebView exposes the same browser APIs and the application admits every artifact. This does not invoke Core speech. |
-| Node | Importable, execution unavailable | The ESM subpath is exported, but `0.2.1` publishes no Node speech storage, Worker, audio-decoder, or execution adapter. |
-| Cloud | Not provided | A cloud speech provider can implement `arcane-ai-provider/2`, but this package never selects one. |
+| Browser | Shipped | Requires DBOPFS/OPFS, Web Locks, Workers, Fetch, Blob/File, and object URLs. Authenticated graphs additionally require `MessageChannel`; shared Blob/File STT requests require the browser audio decoder. |
+| Native WebView | Conditional | Available only when the WebView exposes the same browser APIs and the application admits every artifact. It does not invoke Core speech. |
+| Node | Importable, execution unavailable | The ESM subpath can be imported, but the SDK supplies no Node speech storage, Worker, or audio-decoder host. |
+| Cloud | Not provided | A cloud speech adapter may separately implement `arcane-ai-provider/2`; this package never selects it. |
 
-Browser speech is independent from `arcane-os/ai/browser-wasm`: the Wllama
-entrypoint owns the LLM role, while this entrypoint owns optional STT and TTS
-providers. Each role loads, unloads, reports status, cancels, and recovers
-independently.
+Whisper `stt` and Kokoro `tts` each own an independent provider lifecycle. A
+failure or cancellation in either role does not disable LLM use and never
+authorizes a different local, native, Core, or cloud provider.
 
-## Import
+## Public exports
 
 ```javascript
 import {
+  BROWSER_SPEECH_ARTIFACT_GRAPH_PROTOCOL,
   BROWSER_SPEECH_ARTIFACT_PROTOCOL,
   createBrowserKokoroProvider,
+  createBrowserSpeechArtifactGraph,
   createBrowserSpeechAuthority,
   createBrowserWhisperProvider,
   createDbopfsSpeechArtifactStore
 } from 'arcane-os/ai/browser-speech';
 ```
 
-The entrypoint exports exactly those five names. Importing it downloads no
-model and starts no Worker.
+The entrypoint exports exactly those seven names. Importing it downloads no
+artifact, opens no cache, creates no Worker, and publishes no event.
 
-## `BROWSER_SPEECH_ARTIFACT_PROTOCOL`
+## Protocol and enum registry
 
-The constant is exactly `"arcane-ai-browser-speech-artifacts/1"`. It identifies
-an SDK-created artifact store; it is not model authority, a capability grant,
-or evidence that a cache is complete.
+| Subject | Exact value or closed set |
+| --- | --- |
+| Artifact-store protocol | `arcane-ai-browser-speech-artifacts/1` |
+| Authenticated artifact-graph protocol | `arcane-ai-browser-speech-artifact-graph/1` |
+| Graph `kind` and prepared `runtime.moduleGraph` | `browser-speech-authenticated-artifact-graph` |
+| Legacy prepared `runtime.moduleGraph` | `self-contained` |
+| Legacy authority protocol | `arcane-ai-model-authority/1` |
+| Provider protocol | `arcane-ai-provider/2` |
+| Worker protocol | `arcane-ai-speech-worker/1` |
+| Worker error-envelope protocol | `arcane-ai-speech-worker-error/1` |
+| Nested artifact module-Worker protocol | `arcane-ai-browser-speech-artifact-module-worker/1` |
+| Nested module-Worker rejection event | `artifact-module-worker-bootstrap-rejected` |
+| Runtime graph-guard protocol | `arcane-ai-browser-speech-artifact-graph-runtime/1` |
+| Roles | `stt`, `tts` |
+| Role operations | `transcribe` for `stt`; `synthesize` for `tts` |
+| Public Worker operations | `load`, `use`, `status`, `unload`, `dispose` |
+| Worker transport-only control operation | `cancel` |
+| TTS response format | `wav` |
+| Runtime adapters | `transformers-whisper` for `stt`; `kokoro-js` for `tts` |
+| ONNX namespace identifiers | `transformers-env-backends-onnx-wasm` for `stt`; `kokoro-env-wasm-paths` for `tts` |
+| Edge policies | `artifact-targets-admitted`, `inactive-runtime-branch-rejected` |
+| Import target matches | `exact-runtime-specifier`, `materialized-module-url`; module-Worker targets additionally admit `self-module-url` |
+| Transform kinds | `function-return-this-to-global-this`, `typed-array-constructor` |
+| Graph descriptor status | `artifact-graph-descriptor-verified` |
+| Graph admission | `artifact-graph-network-dbopfs-verified`, `artifact-graph-dbopfs-cache-verified`, `artifact-graph-offline-dbopfs-cache-verified` |
+
+These values are exact, case-sensitive public contract values. Unknown values
+fail closed; a newer-looking value is not treated as compatible.
+
+## `createBrowserSpeechArtifactGraph()`
 
-## `createBrowserSpeechAuthority()`
+This is the operational authority for a real auxiliary ESM/WASM/model/voice
+closure. It creates one frozen, caller-selected graph and computes its SHA-256
+identity from a canonical JSON projection. The inventory is dynamic and has no
+fixed file or edge count; every declared member must still belong to the one
+complete reachable closure.
+
+The top-level shape is:
 
 ```text
-createBrowserSpeechAuthority({
-  providerId,
+createBrowserSpeechArtifactGraph({
+  kind?,
+  identitySha256?,
+  providerId?,
   role,
-  model,
-  runtime,
-  security
-} = {})
+  model: { ... },
+  runtime: { ... },
+  files: [fileDescriptor, ...],
+  edges: {
+    staticImports: [...],
+    dynamicImports: [...],
+    moduleWorkers: [...],
+    fetches: [...],
+    cacheOpens: [...]
+  },
+  transforms: [...]
+})
 ```
 
-Validates and freezes one caller-owned model/runtime declaration. Construction
-binds the declared identities and descriptor rules; actual downloaded or cached
-runtime-graph closure is validated by `store.prepare()` before a Worker starts.
+### Top-level and model fields
 
-### Authority input
+| Field | Contract |
+| --- | --- |
+| `kind` | Omit it or use exactly `browser-speech-authenticated-artifact-graph`. |
+| `identitySha256` | Optional 64-character lowercase SHA-256 assertion. If supplied, it must equal the SDK's canonical descriptor digest. The returned graph always contains the computed value. |
+| `providerId` | `null`/omitted, or one trimmed 1–128 character identity. A non-null value must equal the provider constructor's `id`. |
+| `role` | Exactly `stt` or `tts`. |
+| `model.id` | Trimmed 1–128 character caller-owned model identity. |
+| `model.repository` | Trimmed 1–128 character logical repository identity. The SDK passes it to the admitted adapter but does not resolve it independently. |
+| `model.revision` | Trimmed 1–128 character immutable revision. Every model and voice file must use this same revision. |
+| `model.dtype` | Trimmed 1–128 character caller-owned dtype. There is no SDK-selected graph default. |
+| `model.inputSampleRate` | Required positive safe integer for `stt`; caller-owned. STT must not declare voices. |
+| `model.outputSampleRate` | Required positive safe integer for `tts`; caller-owned. |
+| `model.defaultVoice` | Required for `tts`; must name one declared `voices[].id`. |
+| `model.voices` | Required nonempty TTS array of unique `{id,path}` records. Every path must name one unique `voice-style-binary`, and every such file must be in this inventory. |
+
+### Runtime fields
+
+| Field | Contract |
+| --- | --- |
+| `runtime.adapter` | `transformers-whisper` for `stt`; `kokoro-js` for `tts`. |
+| `runtime.version` / `runtime.revision` | Trimmed 1–128 character caller-pinned identities. The sole entrypoint file revision must equal `runtime.revision`. |
+| `runtime.entrypoint` | Canonical path of the sole `runtime-entrypoint-javascript`. `runtime.entry` is accepted as an input alias; the normalized graph exposes `runtime.entry`. |
+| `runtime.onnxWasm.namespace` | `transformers-env-backends-onnx-wasm` for STT or `kokoro-env-wasm-paths` for TTS. |
+| `runtime.onnxWasm.mjsPath` | Path of a declared `runtime-auxiliary-javascript`. |
+| `runtime.onnxWasm.wasmPath` | Path of a declared `runtime-wasm-binary`. |
+| `runtime.onnxWasm.numThreads` | Optional positive safe integer for Transformers STT only. Kokoro does not expose a verified thread setting and rejects this field with reason `kokoro-env-num-threads-field-not-exposed`. No hardware-derived default is chosen. |
+| `runtime.negativeRuntimeRequestUrls` | Optional unique absolute HTTPS routes that a declared fetch edge may intentionally resolve as a local `404`. They may not overlap a positive source or runtime route and must be referenced by an edge. |
+
+### File descriptor
+
+Every `files[]` entry contains all of these fields:
+
+```text
+{
+  kind,
+  path,
+  sourceUrl,
+  revision,
+  license,
+  mediaType,
+  sourceMediaType?,
+  bytes,
+  sha256,
+  runtimeRequestUrls?,
+  redirectFinalOrigins?
+}
+```
 
 | Field | Contract |
 | --- | --- |
-| `providerId` | Trimmed 1–128 character provider identity. |
-| `role` | Exactly `"stt"` or `"tts"`. |
-| `model.id` | Caller-owned model identity. |
-| `model.repository` | Immutable logical repository identity; the SDK does not resolve it. |
-| `model.revision` | Caller-pinned revision. Every model URL must contain this revision or its supplied SHA-256. |
-| `model.files` | Nonempty unique `{path,url,bytes?,sha256?,mediaType?}[]`. Paths are normalized relative paths; URLs are immutable HTTPS or same-origin URLs without credentials/fragments. |
-| `model.defaultVoice` | Required only for TTS. |
-| `runtime.adapter` | Exactly `"transformers-whisper"` for STT or `"kokoro-js"` for TTS. |
-| `runtime.version` / `runtime.revision` | Caller-pinned runtime identity. |
-| `runtime.entry` | Path of one JavaScript module in `runtime.files`. |
-| `runtime.files` | Nonempty immutable runtime closure. The entry module must be self-contained: undeclared imports, re-exports, executable strings, alternate transports, and child workers are rejected. |
-| `security` | Closed `{secure?,checks?:{byteLength?,sha256?}}` artifact policy. Disabled checks are disclosed rather than silently claimed. |
-
-The result is a frozen `arcane-ai-model-authority/1` record. The SDK owns the
-validation rules; the application owns every model/runtime choice, license,
-revision, URL, hash, voice, and business policy.
-
-Provider security resolves each field independently in this order: SDK default,
-`appSecurity`, provider `security`, then `load({security})`. The SDK default is
-`secure:false`. After resolution, each omitted check defaults to the effective
-`secure` value; an explicit `checks.byteLength` or `checks.sha256` boolean
-overrides that default. An enabled check requires the corresponding `bytes` or
-`sha256` field on every admitted artifact and fails closed when evidence is
-missing or mismatched. A disabled check never becomes a claim that bytes were
-verified. Direct `store.prepare()` has only its supplied security scope above
-the SDK default because it is not bound to a provider.
+| `kind` | One exact kind from the table below. |
+| `path` | Nonempty NFC-normalized relative path. Leading/trailing slash, backslash, empty/`.`/`..` segment, percent escape, query/fragment delimiter, control character, duplicate spelling, or case-folded collision is rejected. |
+| `sourceUrl` | Immutable starting HTTPS or same-origin authority without credentials or fragment. The pathname must not name `main`, `master`, `latest`, `refs/heads/main`, `refs/heads/master`, `resolve/main`, or `resolve/master`, or an `@latest`/`@next` channel. The URL must contain the declared revision or SHA-256 and remains authoritative even when redirect following is opted in. `url` is accepted as an input alias. |
+| `revision` | Nonempty 1–128 character immutable file revision. Runtime entry revision equals the runtime revision; all model/voice revisions equal the model revision. |
+| `license` | Nonempty declaration up to 256 characters. It is part of graph identity and the DBOPFS manifest authority, but it is only a declaration: it is not provenance, a composite notice, corresponding-source evidence, or a license inferred from bytes. The caller remains responsible for complete immutable license/notice evidence. |
+| `mediaType` | Exact lowercase `type/subtype` without parameters. JavaScript is `application/javascript` or `text/javascript`; WASM is `application/wasm`; `*-json` kinds are `application/json`. |
+| `sourceMediaType` | Optional exact lowercase `type/subtype` without parameters for the cold HTTP response. It defaults to `mediaType`; when different, it is exposed in the normalized descriptor and graph identity while `mediaType` remains the authenticated materialized Blob type. |
+| `bytes` | Positive safe integer; always required and always verified for graphs. |
+| `sha256` | Exactly 64 lowercase hexadecimal characters; always required and always verified for graphs. |
+| `runtimeRequestUrls` | Optional unique absolute HTTPS aliases used only inside the authenticated Worker. An alias may describe a third-party runtime's hard-coded mutable request, but it is never a download authority: the Worker maps it to this already verified local file. |
+| `redirectFinalOrigins` | Optional nonempty array that opts this file into Fetch redirect following. Every member must canonicalize to a unique HTTPS origin with no credentials, path, query, or fragment. The normalized array is lexically sorted, frozen, and graph-identity-bound. Omit the field for redirect rejection; an empty array is rejected. |
+
+The closed file kinds are:
+
+| Runtime kinds | Model/data kinds |
+| --- | --- |
+| `runtime-entrypoint-javascript` | `model-configuration-json` |
+| `runtime-auxiliary-javascript` | `model-generation-configuration-json` |
+| `runtime-wasm-binary` | `model-onnx-binary` |
+| `runtime-opaque-data` | `model-onnx-external-data` |
+|  | `model-preprocessor-json` |
+|  | `model-tokenizer-json` |
+|  | `model-opaque-data` |
+|  | `voice-style-binary` |
+
+Paths and immutable source URLs are globally unique within a graph. Positive
+runtime request aliases are also unique and cannot overlap a source URL or a
+negative route. Every non-entry file must be reachable from the declared ONNX
+pair, TTS voice inventory, or one exact edge; unreachable files and routes are
+rejected rather than silently retained.
+
+### Edge and transform descriptors
+
+Edge `occurrence` values are positive, one-based occurrences in the named
+`modulePath`. A declaration must match the scanner's exact source occurrence;
+missing, duplicate, extra, occurrence-mismatched, or undeclared runtime edges
+fail closed. Descriptor collections are canonically sorted for graph identity,
+so caller array order is not semantic.
+
+| Collection | Exact record |
+| --- | --- |
+| `edges.staticImports` | `{modulePath,occurrence,specifier,targetPath}`. The literal specifier and declared JavaScript target must match. |
+| `edges.dynamicImports` | `{modulePath,occurrence,edgePolicy,targets}`. Each target is `{match,targetPath,exactSpecifier?}`. `exactSpecifier` is required only for `exact-runtime-specifier`. |
+| `edges.moduleWorkers` | Same shape as dynamic imports; `self-module-url` is additionally available and must target the declaring module. Only module Workers are admitted. |
+| `edges.fetches` | `{modulePath,occurrence,edgePolicy,methods:["GET"],targetPaths,negativeRuntimeRequestUrls,allowMaterializedUrls}`. `methods` defaults to the same one-element `GET` array; the two target arrays default empty; `allowMaterializedUrls` is `true` only when explicitly set and otherwise normalizes to `false`. No other method is admitted, and a fetch target cannot be JavaScript. |
+| `edges.cacheOpens` | `{modulePath,occurrence,edgePolicy,cacheName,targetPaths}`. `artifact-targets-admitted` requires at least one exact non-JavaScript file path; `inactive-runtime-branch-rejected` requires none. The returned cache can read only those files. |
+| `transforms` | `{kind,modulePath,occurrence}`. `function-return-this-to-global-this` replaces the one audited `Function("return this")()` pattern. `typed-array-constructor` rewrites only a scanner-recognized typed-array `.constructor(...)` call and later proves the receiver uses an intrinsic typed-array prototype. |
+
+An omitted `edgePolicy` normalizes to `artifact-targets-admitted`, which
+requires a declared target. An
+`inactive-runtime-branch-rejected` edge has no target and proves only that the
+branch must remain inactive; execution of that occurrence rejects. The source
+scanner rejects undeclared imports, fetches, Workers, executable-string
+construction or constructor access, Cache Storage access outside a direct
+declared `caches.open(...)`, computed capability access, ambiguous tokens, and
+unmatched transforms. It also reserves the SDK guard name so artifact code
+cannot address the guard directly. Static import cycles are rejected so
+materialization order remains deterministic; runtime isolation separately
+denies every raw transport listed below.
+
+The returned graph is deeply frozen and has exactly
+`{protocol,kind,providerId,role,model,runtime,files,edges,transforms,identitySha256,artifactGraphStatus}`.
+The status is `artifact-graph-descriptor-verified`. Its computed identity binds
+every preceding descriptor field, including paths, revisions, licenses,
+runtime aliases, edge policy, and transform occurrences; the status itself is
+the result of that computation rather than identity input.
 
 ## `createDbopfsSpeechArtifactStore()`
 
@@ -104,226 +235,1012 @@ createDbopfsSpeechArtifactStore({
 } = {})
 ```
 
-Adapts an existing DBOPFS instance into an authority-scoped store. The returned
-frozen object exposes `{protocol,tableName,prepare,remove}`.
+The frozen store exposes `{protocol,tableName,prepare,remove}`. It accepts both
+the authenticated graph above and an SDK-created legacy authority.
+`objectUrlFactory` remains a legacy compatibility seam. Graph preparation
+ignores it and uses module-captured native Blob URL creation, revocation, and
+fetch so a caller cannot substitute the executable materialization boundary.
 
-`prepare(authority,{signal,onProgress,offline=false,security})`:
+For a graph, `prepare(graph,{signal,onProgress,offline=false,security})` follows
+the shared model-security flag and requires explicit `secure:true`. The graph is
+the opt-in strict path: it enables byte-length, SHA-256, and undeclared-
+capability enforcement. Ordinary warn-first operation uses the direct
+`model`/`runtime` authority described below and does not construct or admit an
+artifact graph.
 
-1. obtains an exclusive Web Lock for the exact authority;
-2. admits a complete cached manifest when every record still matches;
-3. otherwise, unless `offline:true`, downloads every declared file without
-   credentials;
-4. enforces enabled byte-length and SHA-256 checks;
-5. validates the closed runtime module graph;
-6. commits `arcane.ai.browser-speech.assets.v1` only after every file succeeds;
-7. returns materialized object URLs plus a `release()` callback.
+### Cold admission
 
-An installation or cache-validation failure removes incomplete stored state for
-that authority. A request or provider-use failure does not imply cache removal.
-`offline:true` never falls back to network and rejects with
-`ARCANE_AI_ARTIFACT_OFFLINE_MISS` when no admitted cache exists.
+1. Acquire the exact authority's exclusive Web Lock.
+2. Delete incomplete prior state.
+3. Fetch only each declared immutable `sourceUrl` using `credentials:"omit"`,
+   `cache:"no-store"`, `mode:"cors"`, and `referrerPolicy:"no-referrer"`.
+   The default is `redirect:"error"`; only a file with declared
+   `redirectFinalOrigins` uses `redirect:"follow"`.
+4. Without a followed redirect, require the response URL to remain exactly the
+   immutable `sourceUrl`. After a followed redirect, require a readable final
+   HTTPS URL without credentials or a fragment and require its canonical origin
+   to occur in that file's declared final-origin inventory.
+5. Require the response Content-Type to equal `sourceMediaType` (which defaults
+   to `mediaType`), then stream exact byte-length and SHA-256 verification into
+   DBOPFS.
+6. Reopen and rehash every persisted file.
+7. Decode and scan every runtime JavaScript file and prove exact edge/transform
+   closure.
+8. Persist `arcane.ai.browser-speech.authenticated-artifact-graph.v1` only
+   after every file and the complete runtime graph pass.
 
-## `createBrowserWhisperProvider()`
+The returned admission is
+`artifact-graph-network-dbopfs-verified`. A failure before manifest completion
+removes the graph's incomplete records.
 
-```text
-createBrowserWhisperProvider({
-  id = 'arcane-browser-whisper',
-  localOnly = true,
-  model,
-  runtime,
-  appSecurity,
-  security,
-  store,
-  offline = false
-} = {})
-```
+`redirectFinalOrigins` admits only a final origin, not a final path, query, or
+signed/expiring URL. The browser may follow such an implementation-specific
+final URL, but that URL is neither persisted nor accepted as source, revision,
+signature, graph identity, or future download authority. Trust remains bound to
+the immutable starting `sourceUrl` and the exact authenticated length and
+SHA-256 of the received bytes.
 
-Returns a frozen `arcane-ai-provider/2` object for role `stt`. The provider
-surface is `{protocol,role,id,localOnly,catalog,inspect,status,load,request,
-unload,dispose}`.
+Browser Fetch exposes the final CORS response, not an inspectable list of every
+intermediate redirect hop. The SDK therefore cannot authenticate intermediate
+hop origins or headers. CORS must succeed for the browser-managed chain, and
+the SDK can enforce only the immutable start, the declared final HTTPS origin,
+the final response metadata, and the end-to-end bytes. A caller that requires
+every hop to be independently pinned must use a direct immutable source rather
+than this redirect opt-in.
 
-The only request operation is `transcribe`:
+### Warm and offline admission
 
-```javascript
-async function transcribeAfterUserChoice(provider, pcmFloat32, signal) {
-  // The admitted Whisper provider must already be ready.
-  const result = await provider.request({
-    role: 'stt',
-    operation: 'transcribe',
-    signal,
-    payload: {
-      audio: pcmFloat32,
-      sampleRate: 16_000
-    }
-  });
-
-  console.log(result.text);
+A warm prepare requires the exact manifest authority and file count, reopens
+every DBOPFS record, checks the manifest's observed bytes and declared SHA-256,
+rehashes every file, and rescans the runtime graph before returning
+`artifact-graph-dbopfs-cache-verified`.
+
+`offline:true` runs the same cache verification and never calls the source
+fetch function. It returns
+`artifact-graph-offline-dbopfs-cache-verified`, or rejects with code
+`ARCANE_AI_ARTIFACT_GRAPH_OFFLINE_CACHE_MISS` and reason
+`artifact-graph-offline-cache-miss`. There is no network repair, runtime CDN,
+private Cache Storage fallback, or partial-cache admission.
+
+The manifest authority binds each file's `sourceMediaType` and
+`redirectFinalOrigins` through the graph identity. A valid warm or offline
+admission does not resolve the starting URL again, follow a redirect, or reuse a
+prior final URL; it authenticates only the complete cached bytes and graph.
+
+After every cold, warm, or offline admission, the store materializes a fresh
+set of unique native `blob:` URLs. Each URL is fetched back with omitted
+credentials and redirect rejection, and its exact URL identity, media type,
+byte length, and SHA-256 are compared with the Blob that was created. Only then
+does the SDK apply the already-scanned deterministic module rewrites. A fresh
+cryptographically random 32-byte lowercase-hex guard capability binds every
+rewritten dynamic import, fetch, cache-open, child-Worker, and typed-array call
+for that materialization. The capability is ephemeral: it is neither supplied
+by the caller nor persisted in the manifest or graph identity.
+
+Successful graph preparation returns:
+
+```text
+{
+  cache,                    // the exact graph admission value
+  artifactGraphId,
+  artifactGraphAdmission,
+  runtime: {
+    ...normalizedRuntime,
+    files, edges, transforms,
+    guardCapability,
+    artifactGraphId,
+    artifactGraphAdmission
+  },
+  model: {...normalizedModel, files},
+  release
 }
 ```
 
-`audio` must be a `Float32Array` sampled at exactly 16,000 Hz. The provider-native
-provider result is a structured `{text}` record transferred by Worker
-`postMessage`; the client does not re-freeze that cloned record. The same
-provider also accepts the shared AI.js STT
-payload `{audio:Blob|File,mimeType,model}`. It verifies `model`, decodes the
-browser audio to authoritative 16 kHz mono PCM, and then enters the same provider-native
-provider operation. Decode support and failure remain explicit.
+`release()` revokes the materialized object URLs after Worker termination. It
+does not delete the caller-owned DBOPFS cache. `remove(graph)` is the explicit
+cache-deletion operation.
+
+## Authenticated Worker host
+
+Graph loading creates one dedicated Worker for the selected role. Its first
+load envelope transfers a new private `MessagePort`; all graph responses,
+progress, cancellation settlement, and later operations stay on that port.
+Graph loading rejects if `MessageChannel`/`MessagePort` is unavailable, if a
+legacy global transport was already selected, or if the Worker receives a graph
+load without the private port. The published single-module compatibility path
+continues to use the original Worker-global message transport.
+
+Before importing the entrypoint, the graph Worker installs these fail-closed
+boundaries:
+
+- exact graph routes map declared source URLs, runtime aliases, and materialized
+  URLs to already authenticated object URLs;
+- the ephemeral guard capability must match at every rewritten call;
+- global `fetch` rejects unless the scanned source was rewritten to one
+  declared fetch edge;
+- global Cache Storage `open` and `match` reject unless the scanned
+  `caches.open(...)` was rewritten to one exact declared cache-open edge; its
+  returned read-only cache serves only `targetPaths`, while `put`, `add`, and
+  `addAll` reject, so DBOPFS remains the sole durable store;
+- a `typed-array-constructor` rewrite returns only the intrinsic constructor of
+  a verified typed-array receiver; an own or non-intrinsic `constructor`
+  rejects;
+- the `Function`, `AsyncFunction`, `GeneratorFunction`, and
+  `AsyncGeneratorFunction` prototype constructor escape is replaced before
+  import, and string callbacks to `setTimeout` or `setInterval` reject;
+- `indexedDB` and `navigator.storage`/OPFS are made unavailable;
+- raw `BroadcastChannel`, `EventSource`, `Function`, `RTCPeerConnection`,
+  `ShadowRealm`, `SharedWorker`, `WebSocket`, `WebSocketStream`, `WebTransport`,
+  `Worker`, `XMLHttpRequest`, `eval`, and `importScripts` capabilities are
+  denied; and
+- an admitted child module Worker is created through the SDK's role Worker,
+  receives the authenticated graph configuration, installs the same guard, and
+  imports only its declared materialized target.
+
+An immutable `sourceUrl` may be an explicitly caller-selected HTTPS origin,
+but runtime execution cannot escalate to that or another network origin. Its
+declared request is answered from authenticated local object bytes, and any
+other request or import edge rejects.
+
+### ONNX Runtime Web configuration
+
+The Worker accepts only the two mechanically verified namespace shapes:
 
-## `createBrowserKokoroProvider()`
+- Kokoro: `namespace.env.wasmPaths = {mjs,wasm}`. No Kokoro cache setting or
+  thread setting is invented. A missing property rejects as
+  `kokoro-env-wasm-paths-unavailable`; a property that rejects or does not
+  retain the exact assignment rejects as
+  `kokoro-env-wasm-paths-assignment-rejected`.
+- Transformers: `namespace.env.backends.onnx.wasm.wasmPaths = {mjs,wasm}`.
+  The Worker also requires the verified outer `env` fields, sets
+  `allowLocalModels:false`, `allowRemoteModels:true`, `useBrowserCache:false`,
+  and `useFSCache:false`. With no admitted `cacheOpens` edge named exactly
+  `transformers-cache`, it assigns `useCustomCache:false` and `customCache:null`.
+  With exactly one such edge it assigns `useCustomCache:true` and the
+  target-limited read-only graph facade; more than one rejects as ambiguous.
+  Only a caller-declared STT `numThreads` is assigned. The
+  `allowRemoteModels` value permits the audited library code path to issue its
+  declared request; the graph guard still prevents network access and serves
+  only exact local graph routes.
+
+For Transformers, the exact missing-field reasons are
+`transformers-env-backends-onnx-wasm-unavailable`,
+`transformers-env-allow-local-models-unavailable`,
+`transformers-env-allow-remote-models-unavailable`,
+`transformers-env-browser-cache-unavailable`,
+`transformers-env-fs-cache-unavailable`,
+`transformers-env-custom-cache-toggle-unavailable`, and
+`transformers-env-custom-cache-unavailable`. The corresponding exact
+assignment-rejection reasons are
+`transformers-env-allow-local-models-assignment-rejected`,
+`transformers-env-allow-remote-models-assignment-rejected`,
+`transformers-env-browser-cache-assignment-rejected`,
+`transformers-env-fs-cache-assignment-rejected`,
+`transformers-env-custom-cache-toggle-assignment-rejected`, and
+`transformers-env-custom-cache-assignment-rejected`. The optional, creatable
+ONNX fields fail assignment as
+`transformers-env-wasm-paths-assignment-rejected` or
+`transformers-env-num-threads-assignment-rejected`; their defensive
+unavailable reasons are `transformers-env-wasm-paths-unavailable` and
+`transformers-env-num-threads-unavailable`. Every case rejects with
+`ARCANE_AI_PROVIDER_UNAVAILABLE`; the Worker never guesses a different
+namespace shape.
+
+Kokoro's audited bundle hard-codes a mutable
+`.../resolve/main/voices/${voice}.bin` request and opens `kokoro-voices` Cache
+Storage. That URL is never accepted as `sourceUrl`. An operational graph must
+declare the exact resolved voice file as an immutable authenticated source,
+bind the hard-coded URL only as that file's `runtimeRequestUrls` alias, and
+declare the exact `cacheOpens` occurrence with `cacheName:"kokoro-voices"` and
+the caller-owned voice paths. Every scanned fetch occurrence still requires its
+own explicit admitted or inactive policy. The Worker satisfies the cache read
+from verified local voice bytes; it never opens Kokoro's private durable cache.
+If the graph does not prove those exact edges, Kokoro fails closed before the
+mutable request can reach the network.
+
+That graph closure is an optional caller-owned secure/offline configuration,
+not the public runtime's default and not a publication gate. In warn-first mode
+the SDK loads the caller-selected version-pinned upstream package and provider
+assets from their publishers. The SDK neither republishes those bytes nor
+requires their optional provenance or legal metadata before the provider can
+operate. When `secure:true` is selected, incomplete graph metadata rejects that
+specific secure load without disabling ordinary warn-first operation.
+
+## Providers
 
 ```text
+createBrowserWhisperProvider({
+  id = 'arcane-browser-whisper',
+  localOnly = true,
+  graph?,
+  model?, runtime?,
+  store,
+  offline = false,
+  appSecurity,
+  security
+} = {})
+
 createBrowserKokoroProvider({
   id = 'arcane-browser-kokoro',
   localOnly = true,
-  model,
-  runtime,
-  appSecurity,
-  security,
+  graph?,
+  model?, runtime?,
   store,
-  offline = false
+  offline = false,
+  appSecurity,
+  security
 } = {})
 ```
 
-Returns the same provider/2 lifecycle for role `tts`. The only request
-operation is `synthesize`:
+`graph` is mutually exclusive with the legacy `model`/`runtime` options and may
+load only when effective security explicitly selects `secure:true`. Both
+constructors require an SDK-created DBOPFS speech store and return one frozen
+`arcane-ai-provider/2` object:
 
-```javascript
-async function synthesizeAfterUserChoice(provider, signal) {
-  // The admitted Kokoro provider must already be ready.
-  const result = await provider.request({
-    role: 'tts',
-    operation: 'synthesize',
-    signal,
-    payload: {
-      text: 'Hello from Arcane.',
-      voice: 'caller-owned-voice-id',
-      speed: 1
-    }
-  });
-
-  console.log(result.audio, result.sampleRate, result.voice);
+```text
+{
+  protocol, role, id, localOnly,
+  catalog, inspect, status, load, request, unload, dispose
 }
 ```
 
-`speed` must be greater than zero and at most four. Omitted `voice` selects the
-caller-supplied `model.defaultVoice`. The provider-native result is the structured
-`{audio:Float32Array,sampleRate:24000,voice}` record transferred by Worker
-`postMessage`; the client does not re-freeze that cloned record. The shared
-AI.js TTS payload is
-`{model,input,responseFormat,voice?,speed?}`. Only `responseFormat:"wav"` is
-admitted; the provider maps it to `{text,voice,speed}` and returns frozen
-`{audio:Uint8Array,contentType:"audio/wav"}` containing 24 kHz mono PCM.
-Shared AI.js synthesis reads the selected model's catalog default and never
-replaces it with a voice saved for a different provider route.
+`localOnly` must remain `true`. `offline:true` makes every later explicit load
+use strict offline graph/cache admission; it does not trigger a load itself.
 
-## Lifecycle and status
+`catalog()` exposes only the caller's selected model/runtime/files. A graph
+catalog also reports `artifactGraphId`, caller dtype, exact sample rate, and the
+caller voice inventory. Its one exact catalog record is
+`{id,providerId,role,localOnly,defaultVoice?,repository,revision,dtype?,runtime,files,artifactGraphId?,voices?,speech}`;
+the `speech` record is `{inputSampleRate}` for STT or
+`{outputSampleRate,responseFormats:["wav"],defaultResponseFormat:"wav"}` for
+TTS.
 
-Whisper and Kokoro use the same independent state machine:
+`inspect()` admits only the exact provider/model/local selection and, when
+supplied, the matching role. A successful graph inspection returns
+`{available:true,authority}`. Its frozen authority has exactly
+`{protocol,admitted,graph,artifactGraphProtocol,providerId,role,modelId,repository,revision,dtype,defaultVoice,voices,inputSampleRate,outputSampleRate,runtime,files,security,artifactGraphId}`;
+`protocol` is `arcane-ai-model-authority/1` and `artifactGraphId` is the graph
+identity. An unavailable record uses
+`ARCANE_AI_MODEL_AUTHORITY_REQUIRED` and the exact role-specific reason
+`stt-provider-inspection-selection-authority-mismatch` or
+`tts-provider-inspection-selection-authority-mismatch`. `load()` is explicit;
+construction, catalog, status, and inspection never start a download or Worker.
 
-| Operation | Behavior |
-| --- | --- |
-| `catalog()` | Returns the single caller-admitted model descriptor. TTS exposes its admitted `defaultVoice`; STT has no voice field. |
-| `inspect(selection,{signal})` | Returns `{available:true,authority}` only for the exact provider/model/local selection; mismatch returns an unavailable record. |
-| `status()` | Returns `{role,providerId,modelId,state,loaded,busy,generation,errorCode,cache}`. |
-| `load(context)` | Requires `{role,selection,progress,signal?,security?}`, admits/cache-materializes artifacts, creates one role Worker, and loads the adapter. Compatible repeated loads coalesce. |
-| `request(context)` | Runs exactly one STT or TTS operation. A provider rejects concurrent use with `ARCANE_AI_PROVIDER_BUSY`. |
-| `unload()` | Aborts active load/use, terminates the Worker, releases object URLs, and returns `unloaded`. |
-| `dispose()` | Completes unload and permanently returns `disposed`. |
-
-Cancellation after a Worker request begins terminates that Worker slot and
-returns the provider to `unloaded`; a later request requires a fresh load. A
-shared STT request can instead be aborted while the browser is still decoding
-its `Blob` or `File`, before Worker use begins. That request rejects with
-`ARCANE_AI_REQUEST_ABORTED`, but the loaded Worker remains intact and the
-provider stays `ready`. Worker crash/message failures become explicit error
-state. No result is retried through another provider.
-
-Register the providers with the projected
-[`AIProviderRuntime`](../runtime-modules.md#aiproviderruntimejs) to normalize
-selection, startup, application calls, and
-[`AIRuntimeState`](../runtime-modules.md#airuntimestatejs) observation. The
-provider objects themselves do not expose an event target.
+### Whisper STT
+
+The provider-native operation is:
 
 ```javascript
-async function loadSpeechProviderAfterUserChoice(provider) {
-  const selection = {
-    providerId: provider.id,
-    modelId: provider.catalog()[0].id,
-    localOnly: true
-  };
-  return provider.load({
-    role: provider.role,
-    selection,
-    progress(update) {
-      console.info(update.phase, update.completed, update.total);
-    }
-  });
+const result = await whisper.request({
+  role: 'stt',
+  operation: 'transcribe',
+  signal,
+  payload: {
+    audio: pcmFloat32,
+    sampleRate: graph.model.inputSampleRate
+  }
+});
+```
+
+The native payload is exact `Float32Array` mono PCM at the graph's
+caller-selected input sample rate. The result is `{text}`. The shared AI
+payload `{audio:Blob|File,mimeType,model}` verifies the exact model, uses the
+browser decoder to normalize to that same rate, and then enters the same Worker
+operation. Multi-channel decoded audio is averaged to one mono `Float32Array`.
+Cancellation during Blob/audio decoding rejects with
+`ARCANE_AI_REQUEST_ABORTED` / `stt-transcription-cancelled` while the already
+loaded Worker remains ready.
+
+### Kokoro TTS
+
+```javascript
+const result = await kokoro.request({
+  role: 'tts',
+  operation: 'synthesize',
+  signal,
+  payload: {
+    text: 'Hello from Arcane.',
+    voice: 'caller-voice-id',
+    speed: 1
+  }
+});
+```
+
+`voice` must belong to the graph's caller-declared inventory; omission uses
+only that graph's `model.defaultVoice`. `speed` must be greater than zero and at
+most four. The provider-native result is
+`{audio:Float32Array,sampleRate,voice}` at the caller-selected output sample
+rate. The shared AI payload `{model,input,responseFormat,voice?,speed?}` accepts
+only `responseFormat:"wav"` and returns frozen
+`{audio:Uint8Array,contentType:"audio/wav"}` containing mono signed 16-bit PCM
+at the caller-selected output rate. No saved voice from another route is
+substituted.
+
+## Lifecycle, progress, cancellation, and cleanup
+
+Provider `status()` returns:
+
+```text
+{
+  role,
+  providerId,
+  modelId,
+  state,
+  lifecycleStatus,
+  lifecycleReason,
+  activeOperation,
+  loaded,
+  busy,
+  generation,
+  errorCode,
+  cache,
+  artifactGraphId,
+  artifactGraphAdmission
 }
 ```
 
-The example only defines the operation; invoke it from an explicit user action
-after showing the admitted runtime/model and download policy. Requests are
-valid only after that load resolves successfully. Applications
-register and configure browser-speech providers on the shared runtime. `AI.js`
-obtains that singleton and translates its existing browser speech request shapes
-at this one provider boundary; it does not register providers, choose
-authorities, or trigger a hidden model download.
+The provider states are exactly `unloaded`, `loading`, `ready`, `unloading`,
+`error`, and `disposed`. `lifecycleStatus` is therefore exactly one of
+`stt-provider-unloaded`, `stt-provider-loading`, `stt-provider-ready`,
+`stt-provider-unloading`, `stt-provider-error`, `stt-provider-disposed`,
+`tts-provider-unloaded`, `tts-provider-loading`, `tts-provider-ready`,
+`tts-provider-unloading`, `tts-provider-error`, or `tts-provider-disposed`.
+
+`activeOperation` is `null` or one of `stt-provider-load`,
+`tts-provider-load`, `stt-provider-transcription`, `tts-provider-synthesis`,
+`stt-provider-unload`, `tts-provider-unload`, `stt-provider-dispose`, and
+`tts-provider-dispose`.
+
+`loaded` is true only for `ready`. `busy` reports an active transcription or
+synthesis request; load, unload, and dispose remain visible through `state` and
+`activeOperation`. `cache` is `null`, a graph admission value, or the legacy
+`installed`/`cached` value. Graph identity and admission are otherwise `null`.
+
+Provider-owned lifecycle reasons are:
+
+- creation and load: `stt-provider-created`, `tts-provider-created`,
+  `stt-load-started`, `tts-load-started`, `stt-load-completed`,
+  `tts-load-completed`, `stt-load-cancelled`, `tts-load-cancelled`,
+  `stt-provider-load-rejected`, `tts-provider-load-rejected`,
+  `stt-load-rejected-during-unload`,
+  `tts-load-rejected-during-unload`,
+  `stt-load-superseded-by-security-change`,
+  `tts-load-superseded-by-security-change`,
+  `stt-load-superseded-by-unload`, `tts-load-superseded-by-unload`,
+  `stt-load-progress-callback-threw`, and
+  `tts-load-progress-callback-threw`;
+- use: `stt-transcription-started`, `stt-transcription-completed`,
+  `stt-transcription-cancelled`,
+  `stt-transcription-engine-operation-rejected`,
+  `tts-synthesis-started`, `tts-synthesis-completed`,
+  `tts-synthesis-cancelled`, `tts-synthesis-engine-operation-rejected`,
+  `stt-transcription-superseded-by-unload`, and
+  `tts-synthesis-superseded-by-unload`;
+- teardown: `stt-unload-started`, `tts-unload-started`,
+  `stt-unload-completed`, `tts-unload-completed`, `stt-dispose-started`,
+  `tts-dispose-started`, `stt-dispose-completed`, `tts-dispose-completed`,
+  `stt-worker-terminated-by-unload`, `tts-worker-terminated-by-unload`,
+  `stt-worker-terminated`, and `tts-worker-terminated`; and
+- Worker failure: `stt-worker-crashed`, `tts-worker-crashed`,
+  `stt-worker-message-rejected`, `tts-worker-message-rejected`,
+  `stt-worker-private-message-rejected`,
+  `tts-worker-private-message-rejected`,
+  `stt-worker-progress-envelope-rejected`,
+  `tts-worker-progress-envelope-rejected`,
+  `stt-worker-response-envelope-shape-rejected`,
+  `tts-worker-response-envelope-shape-rejected`,
+  `stt-worker-error-envelope-rejected`, `tts-worker-error-envelope-rejected`,
+  `stt-worker-protocol-mismatch`, and
+  `tts-worker-protocol-mismatch`.
+
+When an active load or use fails with one of the stable reasons below, the
+provider preserves that exact reason as `lifecycleReason`; it does not collapse
+the boundary to a generic failure label.
+
+Worker statuses use exactly `stt-worker-unloaded`, `stt-worker-ready`,
+`stt-worker-disposed`, `tts-worker-unloaded`, `tts-worker-ready`, and
+`tts-worker-disposed`. The nested bootstrap status is exactly
+`stt-artifact-module-worker-awaiting-initialization` or
+`tts-artifact-module-worker-awaiting-initialization`.
+
+Worker `activeOperation` is `null`, `stt-load`, `tts-load`,
+`stt-transcription`, `tts-synthesis`, `stt-status`, `tts-status`, `stt-unload`,
+`tts-unload`, `stt-dispose`, or `tts-dispose`. Its own lifecycle starts with
+`stt-worker-created` or `tts-worker-created`. Load uses exactly
+`stt-load-started`, `stt-load-completed`, `stt-load-cancelled`,
+`stt-worker-runtime-configuration-rejected`,
+`stt-worker-runtime-import-rejected`, `stt-worker-model-load-rejected`,
+`tts-load-started`, `tts-load-completed`, `tts-load-cancelled`,
+`tts-worker-runtime-configuration-rejected`,
+`tts-worker-runtime-import-rejected`, and
+`tts-worker-model-load-rejected`. Use uses exactly
+`stt-transcription-started`, `stt-transcription-completed`,
+`stt-transcription-cancelled`, `stt-transcription-engine-operation-rejected`,
+`tts-synthesis-started`, `tts-synthesis-completed`,
+`tts-synthesis-cancelled`, and `tts-synthesis-engine-operation-rejected`; a stable underlying
+reason replaces only the matching failure fallback. Orderly Worker teardown
+additionally uses `stt-unload-completed`,
+`tts-unload-completed`, `stt-worker-engine-dispose-rejected`,
+`tts-worker-engine-dispose-rejected`, `stt-dispose-unloaded-worker`,
+`tts-dispose-unloaded-worker`, `stt-dispose-completed`, and
+`tts-dispose-completed`. A Worker protocol cancel result is
+`stt-cancel-target-not-active` or `tts-cancel-target-not-active` when its target
+is already settled; otherwise it names the exact cancelled load, transcription,
+synthesis, status, unload, or dispose operation.
+
+The Worker client admits only `load`, `use`, `status`, `unload`, and `dispose`.
+The transport host additionally admits only its internal `cancel` control
+message. Any other `op` rejects as `ARCANE_AI_INVALID_REQUEST` with
+message `The speech worker operation is not part of its protocol.` and reason
+`stt-worker-operation-unknown` or `tts-worker-operation-unknown` before it can
+enter active-operation, cancellation, or serialized-error naming.
+
+Progress records have exactly `{phase,completed,total,unit,heartbeat}`. They do
+not add a `role` field; Worker phase names carry the role and the provider
+forwards the same provider-neutral record. Graph store phases are exactly
+`artifact-graph-network-download`,
+`artifact-graph-dbopfs-persisted-rehash`, and
+`artifact-graph-dbopfs-cache-rehash`. Worker phases are
+`stt-runtime-import-started`, `tts-runtime-import-started`,
+`stt-model-load-started`, `tts-model-load-started`,
+`stt-model-load-progress`, `tts-model-load-progress`,
+`stt-provider-ready`, and `tts-provider-ready`.
+
+`load(context)` requires the exact role/selection and a `progress` function.
+Compatible concurrent loads coalesce behind one underlying preparation/Worker
+operation while each caller keeps its own progress callback and cancellation
+observation. Cancelling or throwing from the last observer aborts that shared
+load; one observer cannot cancel another observer that remains attached. A
+security change unloads and reloads; graph verification itself can never be
+weakened. `request(context)` admits one role operation at a time and otherwise
+rejects with `ARCANE_AI_PROVIDER_BUSY`.
+
+An `AbortSignal` before or during graph preparation/load rejects that exact
+operation. Once a Worker request has started, cancellation terminates the whole
+role Worker, rejects every pending operation, closes private ports, terminates
+nested Workers, discards that isolated Worker's guards and runtime settings,
+revokes object URLs, and returns the provider to `unloaded`; a later request
+requires explicit load. `unload()` uses the same destructive boundary for
+active load/use before
+releasing state. `dispose()` completes unload and permanently enters
+`disposed`. Cancelling an `unload()` or `dispose()` caller rejects only that
+caller's wait with the corresponding stable reason; the already-started shared
+teardown continues to its terminal state. Late results are generation-checked
+and cannot settle a superseded operation.
+
+The provider owns no listener registry or event source. Applications may
+project its promises, status, and progress records into the one SDK event/state
+authority, and DOM components may project presentation/intent events, but this
+package creates no `EventPubSub`, exposes no public `EventTarget` lifecycle
+source, and owns no competing listener registry or event bus.
+
+The provider also has no implicit mute flag. A higher-level shared state owner
+can implement mute by awaiting the role's explicit `unload()` and unmute by
+performing an explicit `load()`; it projects the returned status and progress
+without inventing a second provider, listener registry, or background load.
+
+## Stable errors and reasons
+
+Operational graph, provider, client, and Worker rejections expose a stable
+`code` and a concise `reason` naming the failed operation or boundary.
+`ARCANE_AI_REQUEST_ABORTED` errors use `name:"AbortError"`; an arbitrary
+caller `AbortSignal.reason` is retained only as `cause` and cannot replace the
+stable public reason. Immediate JavaScript API-shape misuse can instead throw a
+plain `TypeError` before an operation exists.
+
+### Compatibility and provider codes
+
+Worker failures cross the private port only in the exact envelope
+`{protocol:"arcane-ai-speech-worker-error/1",code,message,reason}`. Its own-key
+set must be exactly `code,message,protocol,reason`, every field must be a data
+property rather than an accessor, `protocol` and `message` must match the
+registered code, and the finite registered code/reason pair must match the
+pending role and operation. A missing field, extra string or symbol key,
+foreign value, accessor, or cross-role/cross-operation pairing is rejected and
+the role Worker is terminated; arbitrary third-party error fields never become
+public authority.
+
+| Code | Exact caller-visible reasons |
+| --- | --- |
+| `ARCANE_AI_MODEL_AUTHORITY_REQUIRED` | `stt-provider-inspection-selection-authority-mismatch`, `tts-provider-inspection-selection-authority-mismatch`, `stt-load-selection-authority-mismatch`, `tts-load-selection-authority-mismatch`, `stt-provider-request-selection-authority-mismatch`, `tts-provider-request-selection-authority-mismatch`, `stt-provider-unload-selection-authority-mismatch`, `tts-provider-unload-selection-authority-mismatch`, `stt-provider-dispose-selection-authority-mismatch`, `tts-provider-dispose-selection-authority-mismatch`, `stt-transcription-model-authority-missing`, `stt-transcription-model-authority-mismatch`, `tts-synthesis-model-authority-missing`, `tts-synthesis-model-authority-mismatch`, `tts-synthesis-voice-not-declared` |
+| `ARCANE_AI_INVALID_REQUEST` | The exact context/payload and Worker reasons enumerated immediately below this table. |
+| `ARCANE_AI_NOT_READY` | `stt-provider-request-not-ready`, `tts-provider-request-not-ready`, `stt-transcription-rejected-before-load`, `tts-synthesis-rejected-before-load` |
+| `ARCANE_AI_PROVIDER_BUSY` | `stt-provider-request-already-active`, `tts-provider-request-already-active` |
+| `ARCANE_AI_PROVIDER_DISPOSED` | `stt-provider-load-rejected-after-dispose`, `tts-provider-load-rejected-after-dispose`, `stt-load-rejected-after-dispose`, `tts-load-rejected-after-dispose` |
+| `ARCANE_AI_REQUEST_ABORTED` | `stt-provider-inspection-cancelled`, `tts-provider-inspection-cancelled`, `stt-load-cancelled`, `tts-load-cancelled`, `stt-transcription-cancelled`, `tts-synthesis-cancelled`, `stt-unload-cancelled`, `tts-unload-cancelled`, `stt-dispose-cancelled`, `tts-dispose-cancelled`, `stt-status-cancelled`, `tts-status-cancelled`, `artifact-graph-preparation-cancelled` |
+| `ARCANE_AI_OPERATION_SUPERSEDED` | `stt-load-rejected-during-unload`, `tts-load-rejected-during-unload`, `stt-load-superseded-by-security-change`, `tts-load-superseded-by-security-change`, `stt-load-superseded-by-unload`, `tts-load-superseded-by-unload`, `stt-transcription-superseded-by-unload`, `tts-synthesis-superseded-by-unload`, `stt-worker-terminated-by-unload`, `tts-worker-terminated-by-unload`, `stt-worker-terminated`, `tts-worker-terminated`, `stt-worker-already-terminated`, `tts-worker-already-terminated` |
+| `ARCANE_AI_AUDIO_DECODE_UNAVAILABLE` | `stt-browser-offline-audio-context-unavailable`, `stt-browser-offline-audio-context-construction-rejected`, `stt-browser-audio-decode-method-unavailable` |
+| `ARCANE_AI_AUDIO_DECODE_FAILED` | `stt-browser-audio-decode-operation-rejected`, `stt-browser-decoded-audio-not-object`, `stt-browser-decoded-audio-sample-rate-mismatch`, `stt-browser-decoded-audio-frame-length-not-safe-integer`, `stt-browser-decoded-audio-empty`, `stt-browser-decoded-audio-channel-count-not-safe-integer`, `stt-browser-decoded-audio-channel-count-zero`, `stt-browser-decoded-audio-get-channel-data-not-function`, `stt-browser-decoded-audio-channel-read-rejected`, `stt-browser-decoded-audio-channel-not-float32-array`, `stt-browser-decoded-audio-channel-length-mismatch`, `stt-browser-decoded-audio-sample-non-finite` |
+| `ARCANE_AI_UNSUPPORTED_RESPONSE_FORMAT` | `tts-synthesis-response-format-not-wav` |
+| `ARCANE_AI_INVALID_PROVIDER_RESULT` | `stt-transcription-result-text-not-string`, `tts-synthesis-result-not-object`, `tts-synthesis-result-audio-not-float32-array`, `tts-synthesis-result-sample-rate-mismatch`, `tts-synthesis-result-audio-empty`, `tts-synthesis-result-wav-byte-length-overflow`, `tts-synthesis-pcm-result-not-float32array`, `tts-synthesis-pcm-sample-non-finite` |
+| `ARCANE_AI_PROVIDER_LOAD_FAILED` | `stt-provider-load-rejected`, `tts-provider-load-rejected` |
+| `ARCANE_AI_PROVIDER_REQUEST_FAILED` | `stt-worker-runtime-configuration-rejected`, `tts-worker-runtime-configuration-rejected`, `stt-worker-runtime-import-rejected`, `tts-worker-runtime-import-rejected`, `stt-worker-model-load-rejected`, `tts-worker-model-load-rejected`, `stt-transcription-engine-operation-rejected`, `tts-synthesis-engine-operation-rejected`, `stt-worker-engine-dispose-rejected`, `tts-worker-engine-dispose-rejected`, `stt-worker-dispose-rejected`, `tts-worker-dispose-rejected`, `stt-worker-status-rejected`, `tts-worker-status-rejected` |
+| `ARCANE_AI_PROVIDER_UNAVAILABLE` | `speech-worker-fetch-unavailable`, `artifact-graph-fetch-constructor-unavailable`, `artifact-graph-negative-response-constructor-unavailable`, `artifact-graph-module-worker-constructor-unavailable`, `artifact-graph-onnx-wasm-pair-not-materialized`, `transformers-env-backends-onnx-wasm-unavailable`, `transformers-env-allow-local-models-unavailable`, `transformers-env-allow-local-models-assignment-rejected`, `transformers-env-allow-remote-models-unavailable`, `transformers-env-allow-remote-models-assignment-rejected`, `transformers-env-browser-cache-unavailable`, `transformers-env-browser-cache-assignment-rejected`, `transformers-env-fs-cache-unavailable`, `transformers-env-fs-cache-assignment-rejected`, `transformers-env-custom-cache-toggle-unavailable`, `transformers-env-custom-cache-toggle-assignment-rejected`, `transformers-env-custom-cache-unavailable`, `transformers-env-custom-cache-assignment-rejected`, `transformers-env-wasm-paths-unavailable`, `transformers-env-wasm-paths-assignment-rejected`, `transformers-env-num-threads-unavailable`, `transformers-env-num-threads-assignment-rejected`, `transformers-whisper-pipeline-export-missing`, `stt-transcription-method-unavailable`, `kokoro-env-wasm-paths-unavailable`, `kokoro-env-wasm-paths-assignment-rejected`, `kokoro-tts-constructor-export-missing`, `tts-synthesis-method-unavailable` |
+| `ARCANE_AI_LOAD_PROGRESS_CALLBACK_THREW` | `stt-load-progress-callback-threw`, `tts-load-progress-callback-threw` at the public provider observer boundary |
+| `ARCANE_AI_PROGRESS_CALLBACK_THREW` | `stt-load-progress-callback-threw`, `tts-load-progress-callback-threw` at the internal Worker-client callback boundary |
+| `ARCANE_AI_ADAPTER_PROTOCOL_MISMATCH` | `stt-worker-client-authority-mismatch`, `tts-worker-client-authority-mismatch`, `stt-worker-protocol-mismatch`, `tts-worker-protocol-mismatch`, `artifact-graph-private-message-port-established-too-late`, `stt-worker-runtime-transport-mode-mismatch`, `tts-worker-runtime-transport-mode-mismatch` |
+| `ARCANE_AI_WORKER_CRASHED` | `stt-worker-crashed`, `tts-worker-crashed` |
+| `ARCANE_AI_WORKER_MESSAGE_ERROR` | `stt-worker-message-rejected`, `tts-worker-message-rejected`, `stt-worker-private-message-rejected`, `tts-worker-private-message-rejected`, `stt-worker-progress-envelope-rejected`, `tts-worker-progress-envelope-rejected`, `stt-worker-response-envelope-shape-rejected`, `tts-worker-response-envelope-shape-rejected`, `stt-worker-error-envelope-rejected`, `tts-worker-error-envelope-rejected`, `stt-load-message-rejected`, `tts-load-message-rejected`, `stt-transcription-message-rejected`, `tts-synthesis-message-rejected`, `stt-status-message-rejected`, `tts-status-message-rejected`, `stt-unload-message-rejected`, `tts-unload-message-rejected`, `stt-dispose-message-rejected`, `tts-dispose-message-rejected`, `artifact-graph-private-message-channel-unavailable`, `artifact-graph-private-message-port-unavailable`, `artifact-graph-module-worker-error-envelope-rejected`, `artifact-graph-module-worker-initialization-message-rejected` |
+| `ARCANE_AI_WORKER_MESSAGE_REJECTED` | `stt-worker-error-envelope-rejected`, `tts-worker-error-envelope-rejected` when the provider rejects a foreign or missing SDK error brand |
+| `ARCANE_AI_UNDECLARED_ARTIFACT` | `speech-worker-artifact-request-method-rejected`, `speech-worker-artifact-request-undeclared`, `speech-worker-cache-open-rejected`, `speech-worker-cache-match-rejected`, `artifact-graph-cache-write-rejected`, `artifact-graph-runtime-request-url-malformed` |
+| `ARCANE_AI_ARTIFACT_GRAPH_CONFIGURATION_INVALID` | The exact Worker graph-configuration reasons enumerated below. |
+| `ARCANE_AI_ARTIFACT_GRAPH_FETCH_EDGE_UNDECLARED` | `artifact-graph-fetch-edge-undeclared`, `artifact-graph-fetch-edge-undeclared-inactive-runtime-branch-entered`, `artifact-graph-fetch-method-undeclared`, `artifact-graph-fetch-target-undeclared`, `artifact-graph-fetch-guard-bypassed` |
+| `ARCANE_AI_ARTIFACT_GRAPH_IMPORT_EDGE_UNDECLARED` | `artifact-graph-dynamic-import-edge-undeclared`, `artifact-graph-dynamic-import-edge-undeclared-inactive-runtime-branch-entered`, `artifact-graph-dynamic-import-target-undeclared` |
+| `ARCANE_AI_ARTIFACT_GRAPH_CACHE_EDGE_UNDECLARED` | `artifact-graph-cache-open-edge-undeclared`, `artifact-graph-cache-open-edge-undeclared-inactive-runtime-branch-entered`, `artifact-graph-cache-name-mismatch`, `artifact-graph-cache-read-target-undeclared`, `artifact-graph-cache-open-guard-bypassed`, `artifact-graph-cache-match-guard-bypassed` |
+| `ARCANE_AI_ARTIFACT_GRAPH_WORKER_EDGE_UNDECLARED` | `artifact-graph-module-worker-edge-undeclared`, `artifact-graph-module-worker-edge-undeclared-inactive-runtime-branch-entered`, `artifact-graph-module-worker-target-undeclared`, `artifact-graph-module-worker-type-mismatch` |
+| `ARCANE_AI_ARTIFACT_GRAPH_ISOLATION_UNAVAILABLE` | The exact guard, typed-array, dynamic-code, timer, storage, and denied-capability reasons enumerated below. |
+
+The exact provider context reasons under `ARCANE_AI_INVALID_REQUEST` are:
+
+| Boundary | Exact reasons |
+| --- | --- |
+| Context is not an object | `stt-provider-inspection-context-not-object`, `tts-provider-inspection-context-not-object`, `stt-provider-load-context-not-object`, `tts-provider-load-context-not-object`, `stt-provider-request-context-not-object`, `tts-provider-request-context-not-object`, `stt-provider-unload-context-not-object`, `tts-provider-unload-context-not-object`, `stt-provider-dispose-context-not-object`, `tts-provider-dispose-context-not-object` |
+| Role does not match | `stt-provider-inspection-role-mismatch`, `tts-provider-inspection-role-mismatch`, `stt-provider-load-role-mismatch`, `tts-provider-load-role-mismatch`, `stt-provider-request-role-mismatch`, `tts-provider-request-role-mismatch`, `stt-provider-unload-role-mismatch`, `tts-provider-unload-role-mismatch`, `stt-provider-dispose-role-mismatch`, `tts-provider-dispose-role-mismatch` |
+| Signal is not an `AbortSignal` | `stt-provider-inspection-signal-not-abort-signal`, `tts-provider-inspection-signal-not-abort-signal`, `stt-provider-load-signal-not-abort-signal`, `tts-provider-load-signal-not-abort-signal`, `stt-provider-request-signal-not-abort-signal`, `tts-provider-request-signal-not-abort-signal`, `stt-provider-unload-signal-not-abort-signal`, `tts-provider-unload-signal-not-abort-signal`, `stt-provider-dispose-signal-not-abort-signal`, `tts-provider-dispose-signal-not-abort-signal` |
+| Lifecycle context property read rejects | `stt-provider-unload-context-read-rejected`, `tts-provider-unload-context-read-rejected`, `stt-provider-dispose-context-read-rejected`, `tts-provider-dispose-context-read-rejected` |
+
+The other exact provider request reasons under `ARCANE_AI_INVALID_REQUEST` are
+`stt-load-progress-callback-not-function`,
+`tts-load-progress-callback-not-function`, `stt-provider-operation-mismatch`,
+`tts-provider-operation-mismatch`, `stt-transcription-payload-not-object`,
+`tts-synthesis-payload-not-object`, and these eight structural payload reasons:
+`stt-transcription-payload-not-plain-object`,
+`stt-transcription-payload-field-unknown`,
+`stt-transcription-payload-accessor-rejected`,
+`stt-transcription-payload-required-field-missing`,
+`tts-synthesis-payload-not-plain-object`,
+`tts-synthesis-payload-field-unknown`,
+`tts-synthesis-payload-accessor-rejected`, and
+`tts-synthesis-payload-required-field-missing`.
+
+The remaining exact provider input reasons are
+`stt-transcription-blob-constructor-unavailable`,
+`stt-transcription-audio-not-blob-or-file`,
+`stt-transcription-audio-blob-empty`,
+`stt-transcription-mime-type-not-string`,
+`stt-transcription-mime-type-malformed`,
+`stt-transcription-audio-blob-mime-type-not-string`,
+`stt-transcription-audio-blob-mime-type-malformed`,
+`stt-transcription-audio-blob-mime-type-mismatch`,
+`stt-transcription-audio-not-float32-array`,
+`stt-transcription-pcm-input-empty`,
+`stt-transcription-sample-rate-mismatch`,
+`stt-transcription-pcm-sample-non-finite`, `tts-synthesis-text-empty`,
+`tts-synthesis-voice-empty`, and `tts-synthesis-speed-out-of-range`.
 
-## Errors
+The exact Worker configuration/input reasons under
+`ARCANE_AI_INVALID_REQUEST` are `speech-worker-configuration-missing`,
+`speech-worker-runtime-selection-mismatch`, `speech-worker-model-id-empty`,
+`speech-worker-model-repository-empty`, `speech-worker-model-revision-empty`,
+`speech-worker-model-dtype-empty`, `speech-worker-runtime-entry-empty`,
+`speech-worker-materialized-files-missing`,
+`speech-runtime-entrypoint-not-materialized`,
+`speech-model-sample-rate-not-positive-safe-integer`,
+`tts-default-voice-empty`, `artifact-graph-materialized-file-not-object`,
+`artifact-graph-materialized-path-empty`,
+`artifact-graph-materialized-source-url-empty`,
+`artifact-graph-materialized-module-url-empty`,
+`artifact-graph-materialized-media-type-empty`,
+`artifact-graph-runtime-request-routes-not-array`,
+`artifact-graph-materialized-path-ambiguous`,
+`artifact-graph-entrypoint-not-materialized`,
+`stt-worker-message-envelope-shape-rejected`,
+`tts-worker-message-envelope-shape-rejected`,
+`stt-worker-operation-unknown`, `tts-worker-operation-unknown`,
+`stt-transcription-audio-not-float32array`,
+`stt-transcription-sample-rate-mismatch`, `tts-synthesis-text-empty`,
+`tts-synthesis-voice-empty`, `tts-synthesis-voice-not-declared`, and
+`tts-synthesis-speed-out-of-range`.
 
-| Code | Meaning |
+The exact `ARCANE_AI_ARTIFACT_GRAPH_CONFIGURATION_INVALID` reasons are
+`artifact-graph-worker-configuration-incomplete`,
+`artifact-graph-onnx-wasm-configuration-mismatch`,
+`artifact-graph-kokoro-voice-inventory-missing`,
+`artifact-graph-negative-request-routes-not-array`,
+`artifact-graph-runtime-request-route-ambiguous`,
+`artifact-graph-module-worker-target-not-materialized`,
+`artifact-graph-transform-kind-not-admitted`,
+`artifact-graph-transform-module-path-empty`,
+`artifact-graph-transform-occurrence-not-positive-safe-integer`,
+`artifact-graph-transform-identity-ambiguous`,
+`artifact-graph-typed-array-constructor-transform-undeclared`,
+`artifact-graph-transformers-cache-edge-ambiguous`,
+`kokoro-env-num-threads-field-not-exposed`, and
+`transformers-env-num-threads-not-positive-safe-integer`. For each exact edge
+subject `cache-open`, `dynamic-import`, `fetch`, and `module-worker`, that code
+also admits the six exact endings `edges-not-array`, `edge-not-object`,
+`edge-module-path-empty`, `edge-occurrence-not-positive-safe-integer`,
+`edge-policy-not-admitted`, and `edge-identity-ambiguous`, joined as
+`artifact-graph-{subject}-{ending}`.
+
+`browser-speech-provider-error-reason-unmapped` is a fail-closed internal
+sentinel for a provider error constructed without a mapped reason. No current
+public operation deliberately selects it.
+
+The published legacy authority/store code-to-reason pairs remain:
+
+| Code | Exact compatibility reason |
 | --- | --- |
-| `ARCANE_AI_MODEL_AUTHORITY_REQUIRED` | Selection or authority does not match the provider. |
-| `ARCANE_AI_INVALID_REQUEST` | Role, operation, audio, text, voice, or speed is invalid. |
-| `ARCANE_AI_NOT_READY` | Use was requested before a successful load. |
-| `ARCANE_AI_PROVIDER_BUSY` | The role already has an active request. |
-| `ARCANE_AI_PROVIDER_DISPOSED` | Load/use was requested after disposal. |
-| `ARCANE_AI_REQUEST_ABORTED` | The caller's AbortSignal cancelled work. |
-| `ARCANE_AI_OPERATION_SUPERSEDED` | A newer unload/load/dispose generation replaced the operation. |
-| `ARCANE_AI_ARTIFACT_SOURCE_INVALID` | Artifact bytes or body shape are not readable. |
-| `ARCANE_AI_ARTIFACT_SOURCE_UNAVAILABLE` | Fetch is unavailable. |
-| `ARCANE_AI_ARTIFACT_DOWNLOAD_FAILED` | Network or HTTP download failed. |
-| `ARCANE_AI_ARTIFACT_SOURCE_CHANGED` | Redirect/final URL differs from the admitted immutable URL. |
-| `ARCANE_AI_ARTIFACT_SIZE_MISMATCH` | Enabled byte-length evidence differs. |
-| `ARCANE_AI_ARTIFACT_DIGEST_MISMATCH` | Enabled SHA-256 evidence differs. |
-| `ARCANE_AI_ARTIFACT_CACHE_REJECTED` | DBOPFS did not preserve the completed file. |
-| `ARCANE_AI_ARTIFACT_OFFLINE_MISS` | Strict offline use has no admitted complete cache. |
-| `ARCANE_AI_STORAGE_BUSY` | Another context owns the same authority lock. |
-| `ARCANE_AI_STORAGE_UNAVAILABLE` / `ARCANE_AI_STORAGE_READ_FAILED` / `ARCANE_AI_STORAGE_DELETE_FAILED` | The selected DBOPFS store is unavailable or could not read/delete the authority. |
-| `ARCANE_AI_RUNTIME_MODULE_GRAPH_UNDECLARED` | Runtime code leaves the closed self-contained module grammar. |
-| `ARCANE_AI_WORKER_CRASHED` / `ARCANE_AI_WORKER_MESSAGE_ERROR` | The isolated adapter Worker failed. |
-| `ARCANE_AI_ADAPTER_PROTOCOL_MISMATCH` | The Worker/client/provider protocol is not the SDK-owned contract. |
-| `ARCANE_AI_PROVIDER_LOAD_FAILED` / `ARCANE_AI_PROVIDER_REQUEST_FAILED` / `ARCANE_AI_PROVIDER_UNAVAILABLE` | Adapter load/request failed or the admitted runtime did not expose the requested provider. |
-| `ARCANE_AI_INVALID_PROVIDER_RESULT` | Whisper/Kokoro returned a result outside the normalized contract. |
-| `ARCANE_AI_UNDECLARED_ARTIFACT` | The runtime requested a file outside its admitted closure. |
-| `ARCANE_AI_AUDIO_DECODE_UNAVAILABLE` / `ARCANE_AI_AUDIO_DECODE_FAILED` | Shared Blob/File STT cannot obtain a browser decoder or cannot produce nonempty 16 kHz mono PCM. |
-| `ARCANE_AI_UNSUPPORTED_RESPONSE_FORMAT` | Shared TTS requested a format other than `wav`. |
-
-## Security and ownership
-
-- The SDK contains no speech model or adapter runtime bytes in this entrypoint.
-- The application must verify the license and authority for every supplied byte.
-- URLs never carry credentials; fetch uses `credentials: "omit"`.
-- Mutable `main`, `master`, and `latest` paths are rejected.
-- The completion manifest is consistency evidence for the admitted DBOPFS
-  bytes. It is not publisher authenticity by itself.
-- The runtime entry is isolated in a Worker and limited to its declared
-  artifact closure.
-- Tool calls are an LLM concern; speech providers execute no application tool.
-- A native `Arcane.speech` service is a different privileged host path. This
-  browser package neither calls it nor inherits its capability admission.
-
-<details>
-<summary>Protocol details</summary>
-
-The artifact store identifies itself as
-`arcane-ai-browser-speech-artifacts/1`. Authorities use
-`arcane-ai-model-authority/1`. Providers implement `arcane-ai-provider/2` with
-roles `stt` and `tts`; their only operations are `transcribe` and `synthesize`.
-The Worker client additionally binds exact role, operation, request IDs, and
-generation so stale or cross-role messages cannot settle current work.
-
-</details>
+| `ARCANE_AI_REQUEST_ABORTED` | `browser-speech-artifact-preparation-cancelled` |
+| `ARCANE_AI_STORAGE_BUSY` | `browser-speech-artifact-dbopfs-write-lock-unavailable` |
+| `ARCANE_AI_STORAGE_UNAVAILABLE` | `browser-speech-artifact-dbopfs-table-unavailable` |
+| `ARCANE_AI_STORAGE_DELETE_FAILED` | `browser-speech-artifact-dbopfs-delete-rejected` |
+| `ARCANE_AI_STORAGE_READ_FAILED` | `browser-speech-artifact-dbopfs-read-rejected` |
+| `ARCANE_AI_ARTIFACT_SOURCE_INVALID` | `browser-speech-artifact-source-body-unreadable` |
+| `ARCANE_AI_RUNTIME_MODULE_GRAPH_UNDECLARED` | `browser-speech-runtime-module-graph-undeclared` |
+| `ARCANE_AI_ARTIFACT_SOURCE_UNAVAILABLE` | `browser-speech-artifact-fetch-unavailable` |
+| `ARCANE_AI_ARTIFACT_DOWNLOAD_FAILED` | `browser-speech-artifact-fetch-rejected` |
+| `ARCANE_AI_ARTIFACT_SOURCE_CHANGED` | `browser-speech-artifact-source-redirected` |
+| `ARCANE_AI_ARTIFACT_SIZE_MISMATCH` | `browser-speech-artifact-byte-length-mismatch` |
+| `ARCANE_AI_ARTIFACT_DIGEST_MISMATCH` | `browser-speech-artifact-sha256-mismatch` |
+| `ARCANE_AI_ARTIFACT_CACHE_REJECTED` | `browser-speech-artifact-dbopfs-cache-rejected` |
+| `ARCANE_AI_ARTIFACT_OFFLINE_MISS` | `browser-speech-artifact-offline-cache-miss` |
+
+### Graph reason/code rule
+
+Graph construction and store errors use
+`code = "ARCANE_AI_" + reason.toUpperCase().replaceAll("-", "_")`. The
+following closed reason groups therefore define their exact matching codes.
+
+The redirect and source-response descriptor/runtime pairs are exactly:
+
+| Reason | Code |
+| --- | --- |
+| `artifact-graph-file-source-media-type-missing` | `ARCANE_AI_ARTIFACT_GRAPH_FILE_SOURCE_MEDIA_TYPE_MISSING` |
+| `artifact-graph-file-source-media-type-format-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_FILE_SOURCE_MEDIA_TYPE_FORMAT_MISMATCH` |
+| `artifact-graph-source-redirect-final-origins-not-array` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGINS_NOT_ARRAY` |
+| `artifact-graph-source-redirect-final-origin-inventory-empty` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_INVENTORY_EMPTY` |
+| `artifact-graph-source-redirect-final-origin-text-required` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_TEXT_REQUIRED` |
+| `artifact-graph-source-redirect-final-origin-whitespace-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_WHITESPACE_REJECTED` |
+| `artifact-graph-source-redirect-final-origin-not-absolute` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_NOT_ABSOLUTE` |
+| `artifact-graph-source-redirect-final-origin-protocol-not-https` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_PROTOCOL_NOT_HTTPS` |
+| `artifact-graph-source-redirect-final-origin-credentials-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_CREDENTIALS_REJECTED` |
+| `artifact-graph-source-redirect-final-origin-path-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_PATH_REJECTED` |
+| `artifact-graph-source-redirect-final-origin-query-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_QUERY_REJECTED` |
+| `artifact-graph-source-redirect-final-origin-fragment-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_FRAGMENT_REJECTED` |
+| `artifact-graph-source-redirect-final-origin-duplicate` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_DUPLICATE` |
+| `artifact-graph-source-response-url-unreadable` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_RESPONSE_URL_UNREADABLE` |
+| `artifact-graph-source-redirected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECTED` |
+| `artifact-graph-source-response-url-protocol-not-https` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_RESPONSE_URL_PROTOCOL_NOT_HTTPS` |
+| `artifact-graph-source-response-url-credentials-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_RESPONSE_URL_CREDENTIALS_REJECTED` |
+| `artifact-graph-source-response-url-fragment-rejected` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_RESPONSE_URL_FRAGMENT_REJECTED` |
+| `artifact-graph-source-redirect-final-origin-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_REDIRECT_FINAL_ORIGIN_MISMATCH` |
+| `artifact-graph-source-response-url-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_SOURCE_RESPONSE_URL_MISMATCH` |
+
+When an explicit `sourceMediaType` differs from `mediaType`, a cold response
+mismatch uses the exact file-kind-specific pair below. When the two fields are
+equal, the existing `*-media-type-mismatch` pair applies instead.
+
+| Reason | Code |
+| --- | --- |
+| `artifact-graph-entrypoint-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_ENTRYPOINT_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-runtime-auxiliary-javascript-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_RUNTIME_AUXILIARY_JAVASCRIPT_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-runtime-wasm-binary-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_RUNTIME_WASM_BINARY_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-runtime-opaque-data-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_RUNTIME_OPAQUE_DATA_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-configuration-json-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_CONFIGURATION_JSON_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-generation-configuration-json-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_GENERATION_CONFIGURATION_JSON_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-onnx-binary-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_ONNX_BINARY_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-onnx-external-data-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_ONNX_EXTERNAL_DATA_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-opaque-data-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_OPAQUE_DATA_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-preprocessor-json-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_PREPROCESSOR_JSON_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-model-tokenizer-json-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_MODEL_TOKENIZER_JSON_SOURCE_MEDIA_TYPE_MISMATCH` |
+| `artifact-graph-voice-style-binary-source-media-type-mismatch` | `ARCANE_AI_ARTIFACT_GRAPH_VOICE_STYLE_BINARY_SOURCE_MEDIA_TYPE_MISMATCH` |
+
+Descriptor and identity reasons:
+
+- `artifact-graph-kind-mismatch`, `artifact-graph-role-not-stt-or-tts`,
+  `artifact-graph-provider-id-missing`,
+  `artifact-graph-provider-id-length-exceeded`,
+  `artifact-graph-model-descriptor-missing`,
+  `artifact-graph-runtime-descriptor-missing`,
+  `artifact-graph-file-inventory-missing`,
+  `artifact-graph-file-descriptor-not-object`,
+  `artifact-graph-file-kind-missing`,
+  `artifact-graph-file-kind-not-admitted`,
+  `artifact-graph-file-path-missing`,
+  `artifact-graph-file-path-noncanonical`,
+  `artifact-graph-file-revision-missing`,
+  `artifact-graph-file-revision-length-exceeded`,
+  `artifact-graph-file-byte-length-positive-safe-integer-required`,
+  `artifact-graph-file-sha256-missing`,
+  `artifact-graph-file-sha256-format-mismatch`,
+  `artifact-graph-file-license-missing`,
+  `artifact-graph-file-license-whitespace-rejected`,
+  `artifact-graph-file-license-not-nfc`,
+  `artifact-graph-file-license-control-character-rejected`,
+  `artifact-graph-file-license-length-exceeded`,
+  `artifact-graph-file-media-type-missing`,
+  `artifact-graph-file-media-type-format-mismatch`,
+  `artifact-graph-file-source-media-type-missing`,
+  `artifact-graph-file-source-media-type-format-mismatch`,
+  `artifact-graph-file-identity-ambiguous`,
+  `artifact-graph-source-url-missing`,
+  `artifact-graph-source-url-mutable`,
+  `artifact-graph-source-revision-unbound`,
+  `artifact-graph-source-redirect-final-origins-not-array`,
+  `artifact-graph-source-redirect-final-origin-inventory-empty`,
+  `artifact-graph-source-redirect-final-origin-text-required`,
+  `artifact-graph-source-redirect-final-origin-whitespace-rejected`,
+  `artifact-graph-source-redirect-final-origin-not-absolute`,
+  `artifact-graph-source-redirect-final-origin-protocol-not-https`,
+  `artifact-graph-source-redirect-final-origin-credentials-rejected`,
+  `artifact-graph-source-redirect-final-origin-path-rejected`,
+  `artifact-graph-source-redirect-final-origin-query-rejected`,
+  `artifact-graph-source-redirect-final-origin-fragment-rejected`,
+  `artifact-graph-source-redirect-final-origin-duplicate`,
+  `artifact-graph-javascript-media-type-mismatch`,
+  `artifact-graph-wasm-media-type-mismatch`,
+  `artifact-graph-json-media-type-mismatch`,
+  `artifact-graph-entrypoint-path-missing`,
+  `artifact-graph-entrypoint-path-noncanonical`,
+  `artifact-graph-entrypoint-file-kind-mismatch`,
+  `artifact-graph-entrypoint-count-mismatch`,
+  `artifact-graph-entrypoint-revision-mismatch`,
+  `artifact-graph-runtime-adapter-missing`,
+  `artifact-graph-runtime-adapter-role-mismatch`,
+  `artifact-graph-runtime-version-missing`,
+  `artifact-graph-runtime-version-length-exceeded`,
+  `artifact-graph-runtime-revision-missing`,
+  `artifact-graph-runtime-revision-length-exceeded`,
+  `artifact-graph-onnx-wasm-descriptor-missing`,
+  `artifact-graph-onnx-wasm-namespace-missing`,
+  `artifact-graph-onnx-wasm-namespace-role-mismatch`,
+  `artifact-graph-onnx-wasm-file-kind-mismatch`,
+  `artifact-graph-onnx-wasm-num-threads-positive-safe-integer-required`,
+  `artifact-graph-negative-runtime-routes-not-array`,
+  `artifact-graph-negative-runtime-route-duplicate`,
+  `artifact-graph-negative-runtime-route-ambiguous`,
+  `artifact-graph-model-id-missing`,
+  `artifact-graph-model-id-length-exceeded`,
+  `artifact-graph-model-repository-missing`,
+  `artifact-graph-model-repository-length-exceeded`,
+  `artifact-graph-model-revision-missing`,
+  `artifact-graph-model-revision-length-exceeded`,
+  `artifact-graph-model-dtype-missing`,
+  `artifact-graph-model-dtype-length-exceeded`,
+  `artifact-graph-sample-rate-missing`,
+  `artifact-graph-sample-rate-positive-safe-integer-required`,
+  `artifact-graph-default-voice-missing`,
+  `artifact-graph-default-voice-length-exceeded`,
+  `artifact-graph-default-voice-undeclared`,
+  `artifact-graph-stt-voice-authority-declared`,
+  `artifact-graph-stt-voice-file-declared`,
+  `artifact-graph-voice-inventory-missing`,
+  `artifact-graph-voice-descriptor-not-object`,
+  `artifact-graph-voice-id-missing`,
+  `artifact-graph-voice-id-length-exceeded`,
+  `artifact-graph-voice-file-kind-mismatch`,
+  `artifact-graph-voice-inventory-ambiguous`,
+  `artifact-graph-voice-file-undeclared`,
+  `artifact-graph-model-file-inventory-missing`,
+  `artifact-graph-model-file-revision-mismatch`,
+  `artifact-graph-file-unreachable`,
+  `artifact-graph-runtime-request-routes-not-array`,
+  `artifact-graph-runtime-request-url-text-required`,
+  `artifact-graph-runtime-request-url-not-absolute`,
+  `artifact-graph-runtime-request-url-protocol-not-https`,
+  `artifact-graph-runtime-request-url-credentials-rejected`,
+  `artifact-graph-runtime-request-url-fragment-rejected`,
+  `artifact-graph-runtime-request-route-duplicate`,
+  `artifact-graph-runtime-request-route-ambiguous`,
+  `artifact-graph-runtime-request-route-unreachable`,
+  `artifact-graph-negative-runtime-route-unreachable`,
+  `artifact-graph-identity-sha256-text-required`,
+  `artifact-graph-identity-sha256-format-mismatch`, and
+  `artifact-graph-identity-sha256-mismatch`.
+
+The lower-level field helpers retain the exact reasons
+`artifact-graph-field-text-required`, `artifact-graph-identifier-missing`,
+`artifact-graph-identifier-length-exceeded`, and
+`artifact-graph-positive-safe-integer-required`. A Kokoro thread declaration
+raises `kokoro-env-num-threads-field-not-exposed` at graph construction; the
+Worker retains that same reason under its defensive graph-configuration
+boundary if construction is bypassed.
+
+Edge and transform reasons:
+
+- `artifact-graph-edges-not-object`, `artifact-graph-edge-kind-not-admitted`,
+  `artifact-graph-edge-list-not-array`,
+  `artifact-graph-edge-not-object`, `artifact-graph-edge-module-path-missing`,
+  `artifact-graph-edge-module-path-noncanonical`,
+  `artifact-graph-edge-module-path-not-runtime-javascript`,
+  `artifact-graph-edge-occurrence-positive-safe-integer-required`,
+  `artifact-graph-edge-occurrence-duplicate`,
+  `artifact-graph-edge-policy-not-admitted`,
+  `artifact-graph-edge-targets-not-array`,
+  `artifact-graph-edge-target-not-object`,
+  `artifact-graph-edge-target-match-missing`,
+  `artifact-graph-edge-target-match-not-admitted`,
+  `artifact-graph-edge-target-path-missing`,
+  `artifact-graph-edge-target-path-noncanonical`,
+  `artifact-graph-edge-target-path-undeclared`,
+  `artifact-graph-edge-target-specifier-missing`,
+  `artifact-graph-edge-target-duplicate`,
+  `artifact-graph-static-import-specifier-missing`,
+  `artifact-graph-dynamic-import-policy-target-mismatch`,
+  `artifact-graph-module-worker-policy-target-mismatch`,
+  `artifact-graph-module-worker-self-target-mismatch`,
+  `artifact-graph-fetch-method-not-get`,
+  `artifact-graph-fetch-targets-not-array`,
+  `artifact-graph-fetch-target-duplicate`,
+  `artifact-graph-fetch-javascript-target-rejected`,
+  `artifact-graph-fetch-negative-routes-not-array`,
+  `artifact-graph-fetch-negative-route-undeclared`,
+  `artifact-graph-fetch-targets-incomplete`,
+  `artifact-graph-fetch-policy-target-mismatch`,
+  `artifact-graph-cache-name-missing`,
+  `artifact-graph-cache-targets-not-array`,
+  `artifact-graph-cache-target-duplicate`,
+  `artifact-graph-cache-javascript-target-rejected`,
+  `artifact-graph-cache-policy-target-mismatch`,
+  `artifact-graph-transforms-not-array`,
+  `artifact-graph-transform-not-object`,
+  `artifact-graph-transform-kind-not-admitted`, and
+  `artifact-graph-transform-occurrence-duplicate`.
+
+Runtime scan and materialization reasons:
+
+- `artifact-graph-runtime-javascript-file-missing`,
+  `artifact-graph-runtime-javascript-utf8-decode-rejected`,
+  `artifact-graph-javascript-block-comment-unterminated`,
+  `artifact-graph-javascript-escape-unterminated`,
+  `artifact-graph-javascript-hexadecimal-escape-malformed`,
+  `artifact-graph-javascript-unicode-code-point-escape-malformed`,
+  `artifact-graph-javascript-unicode-code-point-out-of-range`,
+  `artifact-graph-javascript-unicode-escape-malformed`,
+  `artifact-graph-javascript-quoted-string-line-break-rejected`,
+  `artifact-graph-javascript-quoted-string-unterminated`,
+  `artifact-graph-javascript-regexp-line-break-rejected`,
+  `artifact-graph-javascript-regexp-unterminated`,
+  `artifact-graph-javascript-template-expression-unterminated`,
+  `artifact-graph-javascript-template-literal-unterminated`,
+  `artifact-graph-javascript-escaped-identifier-rejected`,
+  `artifact-graph-runtime-fetch-direct-call-required`,
+  `artifact-graph-runtime-fetch-receiver-not-global`,
+  `artifact-graph-runtime-cache-open-direct-call-required`,
+  `artifact-graph-runtime-cache-open-receiver-not-global`,
+  `artifact-graph-runtime-worker-constructor-call-required`,
+  `artifact-graph-runtime-worker-receiver-not-global`,
+  `artifact-graph-runtime-guard-reference-reserved`,
+  `artifact-graph-runtime-dynamic-code-undeclared`,
+  `artifact-graph-runtime-computed-dynamic-code-undeclared`,
+  `artifact-graph-runtime-constructor-dynamic-code-undeclared`,
+  `artifact-graph-runtime-edge-undeclared`,
+  `artifact-graph-runtime-edge-declaration-unmatched`,
+  `artifact-graph-runtime-edge-occurrence-noncanonical`,
+  `artifact-graph-static-import-specifier-mismatch`,
+  `artifact-graph-static-import-specifier-unresolved`,
+  `artifact-graph-static-import-target-unmaterialized`,
+  `artifact-graph-runtime-static-import-cycle`,
+  `artifact-graph-module-transform-overlap`,
+  `artifact-graph-guard-capability-unavailable`,
+  `artifact-graph-object-url-platform-unavailable`,
+  `artifact-graph-object-url-scheme-not-blob`,
+  `artifact-graph-object-url-identity-ambiguous`,
+  `artifact-graph-object-url-readback-unavailable`,
+  `artifact-graph-object-url-readback-http-status-rejected`,
+  `artifact-graph-object-url-readback-identity-mismatch`,
+  `artifact-graph-object-url-media-type-mismatch`,
+  `artifact-graph-object-url-byte-length-mismatch`, and
+  `artifact-graph-object-url-sha256-mismatch`.
+
+Source, cache, and security reasons:
+
+- `artifact-graph-load-security-contract-rejected`,
+  `artifact-graph-source-fetch-rejected`,
+  `artifact-graph-source-http-response-rejected`,
+  `artifact-graph-source-redirected`,
+  `artifact-graph-source-response-url-unreadable`,
+  `artifact-graph-source-response-url-protocol-not-https`,
+  `artifact-graph-source-response-url-credentials-rejected`,
+  `artifact-graph-source-response-url-fragment-rejected`,
+  `artifact-graph-source-redirect-final-origin-mismatch`,
+  `artifact-graph-source-response-url-mismatch`,
+  `artifact-graph-offline-cache-miss`, and
+  `artifact-graph-preparation-cancelled`.
+
+For an exact file verification failure, `reason` is
+`artifact-graph-{subject}-{boundary}`. `{subject}` is `entrypoint` for
+`runtime-entrypoint-javascript`; otherwise it is the exact file kind from the
+closed kind table. `{boundary}` is one of `media-type-mismatch`,
+`source-media-type-mismatch`, `byte-length-mismatch`, `sha256-mismatch`,
+`dbopfs-persisted-byte-length-mismatch`, or
+`dbopfs-persisted-sha256-mismatch`. `source-media-type-mismatch` occurs only
+when an explicit `sourceMediaType` differs from `mediaType`; otherwise a source
+Content-Type mismatch uses `media-type-mismatch`. This expansion produces the
+exact public code by the rule above; for example,
+`artifact-graph-entrypoint-sha256-mismatch` becomes
+`ARCANE_AI_ARTIFACT_GRAPH_ENTRYPOINT_SHA256_MISMATCH`.
+
+The other exact isolation reasons are
+`artifact-graph-private-message-port-missing`,
+`artifact-graph-guard-global-collision`,
+`artifact-graph-guard-global-definition-rejected`,
+`artifact-graph-guard-capability-mismatch`,
+`artifact-graph-fetch-isolation-unavailable`,
+`artifact-graph-cache-isolation-unavailable`,
+`speech-worker-fetch-isolation-unavailable`,
+`speech-worker-cache-isolation-unavailable`,
+`artifact-graph-typed-array-validation-unavailable`,
+`artifact-graph-typed-array-constructor-receiver-not-typed-array`,
+`artifact-graph-typed-array-constructor-intrinsic-mismatch`,
+`artifact-graph-dynamic-code-constructor-rejected`,
+`artifact-graph-dynamic-code-constructor-isolation-unavailable`,
+`artifact-graph-setinterval-isolation-unavailable`,
+`artifact-graph-settimeout-isolation-unavailable`,
+`artifact-graph-setinterval-string-callback-rejected`,
+`artifact-graph-settimeout-string-callback-rejected`,
+`artifact-graph-indexeddb-isolation-unavailable`, and
+`artifact-graph-opfs-isolation-unavailable`.
+
+The denied-capability Worker reasons are
+`artifact-graph-{capability}-capability-undeclared` and
+`artifact-graph-{capability}-isolation-unavailable`, where `{capability}` is
+exactly `broadcastchannel`, `eventsource`, `function`, `rtcpeerconnection`,
+`shadowrealm`, `sharedworker`, `websocket`, `websocketstream`, `webtransport`,
+`worker`, `xmlhttprequest`, `eval`, or `importscripts`.
+
+## `createBrowserSpeechAuthority()` upstream package mode
+
+`BROWSER_SPEECH_ARTIFACT_PROTOCOL` remains exactly
+`arcane-ai-browser-speech-artifacts/1`, and
+`createBrowserSpeechAuthority({providerId,role,model,runtime,security})`
+loads a caller-selected browser bundle from npm or another upstream package
+authority. The SDK stores only the downloaded application-selected entrypoint;
+the runtime uses its normal provider fetch and browser cache behavior. Its
+optional byte/SHA checks still resolve independently from SDK default to app,
+provider, and load scopes. Its cache values remain `installed` and `cached`,
+and its offline miss remains `ARCANE_AI_ARTIFACT_OFFLINE_MISS`.
+
+The model descriptor is
+`{id,repository,revision,defaultVoice?,files?}`; `defaultVoice` is required only
+for `tts`. In default warn-first mode `files` may be omitted so the selected
+provider downloads its own model and voice assets. The runtime descriptor is
+`{adapter,version,revision,entry,wasmPaths?,files}` and its adapter is exactly
+`transformers-whisper` for `stt` or `kokoro-js` for `tts`. Every legacy file is
+`{path,url,bytes?,sha256?,mediaType?}` with a normalized relative path, a
+unique immutable URL containing the descriptor revision or SHA-256 identity,
+and optional positive byte length and SHA-256. Legacy SHA input is normalized
+to lowercase before validation and projection. `runtime.entry` must
+name the one `text/javascript` `.js` or `.mjs` runtime file. The frozen result
+is
+`{protocol,providerId,modelId,admitted,role,repository,revision,defaultVoice,runtime,files,security}`.
+
+`runtime.wasmPaths` may name a version-pinned upstream npm CDN directory in
+warn-first mode. `secure:true` rejects remote `wasmPaths`, requires explicit
+model files, applies the closed-module scan, and replaces ordinary fetch/cache
+access with the admitted file map. The authenticated artifact graph remains an
+advanced strict-control option; it is not required for normal speech use.
+
+## Security and ownership summary
+
+- The SDK redistributes no speech runtime, model, voice, third-party license,
+  or corresponding-source payload. Explicit `load()` resolves them from the
+  caller-selected npm/package/provider authorities.
+- Default direct-authority operation is warn-first and preserves ordinary
+  upstream fetch/cache behavior. An artifact graph is admitted only when
+  `secure:true` explicitly opts into strict graph/file verification and
+  capability isolation.
+- Every graph byte is caller-selected, immutable, exact-length, SHA-256 bound,
+  revision bound, media-type bound, license-declaration bound, and reachable
+  through one closed graph identity.
+- Redirects are rejected by default. An undeclared redirect, an undeclared or
+  non-HTTPS final origin, a mutable starting source authority, an ambiguous
+  path/route, an undeclared code/data edge, a raw network transport, a cache
+  write, or an incomplete offline closure fails closed.
+- DBOPFS is the sole durable artifact store; the Worker sees only authenticated
+  local object bytes and an exact read-only cache facade.
+- Runtime/model/sample-rate/default-voice/voice inventory and optional
+  Transformers thread count remain caller authority. No hardware heuristic,
+  hidden fallback, startup download, native/Core call, or cloud retry is added.
+- A completion manifest and SHA-256 prove consistency with the caller's graph;
+  they are not independent publisher authenticity or complete license evidence.
+- The component resolution record is
+  `browser-runtime/ai/ARCANE_AI_BROWSER_SPEECH_COMPONENTS.json`; the detailed
+  package/model audit is `docs/reference/ai/browser-speech-package-authority.json`.
+- Upstream publishers remain responsible for their distributed package and
+  provider assets. Applications own selected model/voice policy and integration
+  with the one shared SDK state/event owner.
 
 ## Related
 
@@ -331,6 +1248,5 @@ generation so stale or cross-role messages cannot settle current work.
 - [Browser-WASM LLM](browser-wasm.md)
 - [AIProviderRuntime.js](../runtime-modules.md#aiproviderruntimejs)
 - [AIRuntimeState.js](../runtime-modules.md#airuntimestatejs)
-- [PersistentAIChatSession.js](../runtime-modules.md#persistentaichatsessionjs)
 - [Availability and normalization](../availability-and-normalization.md)
 - [Protocol architecture](../protocols.md#portable-ai-provider-runtime)