fetch/XHR transport failures and unhandled rejections lose all diagnostic fidelity before reaching the embedding app #196

Applications embedding Babylon Native depend on field crash reports for ongoing maintenance, and today a fetch()/XMLHttpRequest failure arrives there with no usable signal: DNS failure, connection refused, TLS rejection, proxy auth failure, server outage, and a missing bundled app:/// asset are all byte-identical, and an un-caught rejection never reaches the host (BN) application at all. This issue tracks the chain of fixes; each hop is independently shippable.

The chain (where fidelity is lost today)

1. UrlLib swallows transport errors at the source. ✅ Fixed by Expose normalized transport-error detail (ErrorString/ErrorSymbol/ErrorCode) UrlLib#31 (open): the Apple backend discarded the NSError (long-standing in-code TODO) and the curl backend caught and dropped its own failure, leaving only StatusCode() == 0. UrlLib now exposes ErrorString() / ErrorSymbol() / ErrorCode(), normalized as "<domain>:<symbol>(<code>): <detail>" (e.g. curl:CURLE_COULDNT_RESOLVE_HOST(6): ..., nsurl:NSURLErrorCannotConnectToHost(-1004): ..., urllib:AppResourceNotFound(0): ...) — stable tokens for observability-pipeline filtering, purely additive (status-0 contract unchanged).

2. The fetch polyfill flattens every transport failure to a constant string. Fetch.cpp throws std::runtime_error{"fetch: network request failed"}, discarding result.error() and the URL/method context it has in scope; the rejection surfaces as a plain Error (browsers/Node/Bun: TypeError) with no cause, no code, and a .stack snapshotted inside the scheduler tick — i.e. zero user frames. Proposed shape, once the UrlLib pin includes Test CI #31:

   • reject with a TypeError whose message is stable ("fetch failed" style — keeps crash-report grouping intact), carrying the variable detail as properties: cause (message from ErrorString()), code (ErrorSymbol()), url;
   • capture the JS call-site stack synchronously inside fetch() before SendAsync() (create the rejection Error, or a stack carrier, while user frames are still on the stack — the undici approach) so crash reports can attribute the failing call;
   • same treatment for XMLHttpRequest's error path.

3. Unhandled promise rejections never reach UnhandledExceptionHandler. AppRuntime::Dispatch only catches synchronous Napi::Error throws from dispatched callbacks; no engine-level rejection tracker is wired anywhere in Core, so a fire-and-forget fetch() failure (or a throw inside any .then) vanishes silently — the embedder's handler never fires and the process exits 0. Proposal: an opt-in AppRuntime::Options handler (or routing into the existing one) fed per engine — Isolate::SetPromiseRejectCallback (V8), JsSetHostPromiseRejectionTracker (Chakra), JSGlobalContextSetUnhandledRejectionCallback (JavaScriptCore), and the JSI tracker — the per-engine seams already exist as AppRuntime_{V8,Chakra,JavaScriptCore,JSI}.cpp.

4. fetch ignores init.signal. The implementation passes arcana::cancellation::none() for both continuations, so abort never rejects (AbortError) and never cancels the transport. Two prerequisites: UrlRequest::Abort() currently only cancels the Windows backend (the curl/NSURLSession backends never observe m_cancellationSource), and the AbortSignal polyfill predates the modern spec (no reason/throwIfAborted(), writable aborted). Worth sequencing with Playground: link AbortController + TextEncoder polyfills from JsRuntimeHost BabylonNative#1708, which installs AbortController globally in the Playground — after which library feature-detection passes while signals silently no-op, and user-initiated cancellations become indistinguishable from network failures in telemetry.

Validation

Beyond unit tests per hop (UrlLib#31 already ships offline-deterministic transport-failure tests + CI), the highest-value conformance ports for this area: the WPT fetch/api/abort/general.any.js core cases (pre-aborted/mid-flight/post-settle, signal.reason), undici's fetch failed-with-cause assertions from test/fetch/client-fetch.js, a four-way failure-shape bank (refused / NXDOMAIN / bad TLS / missing local asset must be distinguishable and TypeError-shaped), and a native-side test that an unhandled rejection actually reaches the host handler.

Happy to submit PRs for hops 2–4 individually (2 is unblocked as soon as the UrlLib pin can include #31; 3 and 4 are independent).

Related: #188 (fetch polyfill), #195 (statusText — same UrlLib-consumption pattern hop 2 would follow), BabylonJS/BabylonNative#1707.