fix(runner): match the host-suffixed 'Failed to fetch' wording so the DEMOS-2X gate fires (DEV-2859)

runner/apps/authoring/src/fetchFailure.ts

@@ -1,6 +1,8 @@
 /**
  * Is `error` the browser's opaque transport failure — the request never completed,
- * so there is no status, no body, and nothing about our host in it?
+ * so there is no status and no body? (The message itself may still name our host —
+ * see the parenthesised-suffix note below — but nothing else about the response
+ * does.)
  *
  * Sentry DEMOS-2X / DEMOS-2Y. Deliberately shape-based and NOT gated on
  * `navigator.onLine`: unlike the Tier-1 compiler-asset branch in tier1Report.ts,
@@ -17,34 +19,51 @@
  * `--experimental-strip-types`, so `pipeline/fetch-failure.test.mjs` can pin the
  * decision without pulling in `@sentry/react` or `import.meta.env`.
  *
- * Matched on a message *substring*, not on the Sentry issue title. DEMOS-2X's title
- * is `TypeError: Failed to fetch (demos.handsontable.com)`, but the host suffix is
- * appended by the Sentry SDK's own fetch instrumentation
- * (`@sentry/core/build/esm/instrument/fetch.js`) — the `error.message` that actually
- * reaches this `.catch` is just `Failed to fetch`. A classifier written against the
- * title would never fire in production. Verified by observation (Step 0a, this
- * plan): `route.abort("failed")` under Chromium raises exactly
- * `TypeError: Failed to fetch`.
+ * Matched on `error.message` itself (via the anchored regex below), not on the
+ * Sentry issue title. DEMOS-2X's title
+ * is `TypeError: Failed to fetch (demos.handsontable.com)`, and — contrary to this
+ * module's original assumption — that host suffix is NOT stripped by the time
+ * `error.message` reaches this `.catch`: DEMOS-2X's latest events (2026-09-08,
+ * releases `930f6a52` and `e32cdcdf`, both after the `^failed to fetch$` gate landed
+ * in commit 590cb58b2 / PR #274) still carry `Failed to fetch (demos.handsontable.com)`
+ * verbatim, and that gate — the only caller of `fetchVersions` is gated on it — has
+ * never once matched it, so the "demotion" it was meant to apply has never taken
+ * effect in production. The shape below now accepts one optional parenthesised
+ * suffix so the gate actually reaches the population it was written for. Verified by
+ * observation (Step 0a, this plan): `route.abort("failed")` under Chromium raises
+ * exactly `TypeError: Failed to fetch`, i.e. without a suffix — so both the bare and
+ * suffixed forms are real, observed wordings and both must match.
  *
  * One named regex per engine wording — in the style of `isPreviewPortUnreachable`
  * (`workers/api/src/preview-boot.ts`) — rather than one fused pattern, so a wording
  * this table does not cover shows up as a new Sentry event instead of being folded
  * in silently.
  *
- * The Chromium row is anchored (`^...$`) rather than a loose substring match. Chromium
- * also raises `TypeError: "Failed to fetch dynamically imported module: <url>"` for a
- * deploy-rotated chunk served under SPA fallback — a real host defect this codebase
- * already treats as one (`packages/runtime/src/transpile.ts`, Sentry DEMOS-15 /
- * DEV-2569). An unanchored `/failed to fetch/i` matches that message too, silencing a
- * defect class it has no business touching; the bare transport failure this module
- * exists for is always the whole message, never a prefix of a longer one. Firefox and
+ * The Chromium row is anchored (`^...$`) rather than a loose substring match, with
+ * the host suffix carved out as a single optional `(?: \([^)]*\))?` group immediately
+ * before the `$` — so the suffix, when present, must be the very end of the message.
+ * Chromium also raises `TypeError: "Failed to fetch dynamically imported module:
+ * <url>"` for a deploy-rotated chunk served under SPA fallback — a real host defect
+ * this codebase already treats as one (`packages/runtime/src/transpile.ts`, Sentry
+ * DEMOS-15 / DEV-2569). An unanchored `/failed to fetch/i` matches that message too,
+ * silencing a defect class it has no business touching; the anchor plus the narrow
+ * optional group still excludes it, because that wording is not "Failed to fetch"
+ * plus a trailing `(...)` — it has different words after the colon. The bare
+ * transport failure this module exists for is always the whole message or the whole
+ * message plus a parenthesised host, never a prefix of a longer one. Firefox and
  * Safari stay loose: Firefox's real wording carries a trailing period
  * (`NetworkError when attempting to fetch resource.`), so a `$`-anchored version would
  * break the one wording it exists to match, and no over-match has been demonstrated
  * for either engine.
  */
 const OPAQUE_TRANSPORT_MESSAGES = [
-  /^failed to fetch$/i, // Chrome/Chromium/Edge — verified against `route.abort("failed")` (Step 0a)
+  // Chrome/Chromium/Edge — verified against `route.abort("failed")` (Step 0a) for the
+  // bare form; the optional `(host)` suffix is the actual production wording seen in
+  // Sentry DEMOS-2X (e.g. `Failed to fetch (demos.handsontable.com)`, releases
+  // `930f6a52` / `e32cdcdf`, 2026-09-08). Nothing in this codebase appends it (grepped);
+  // which layer does — the browser itself vs. the Sentry SDK's fetch instrumentation —
+  // is not verified.
+  /^failed to fetch(?: \([^)]*\))?$/i,
   /networkerror when attempting to fetch resource/i, // Firefox
   /load failed/i, // Safari
 ];

runner/pipeline/fetch-failure.test.mjs

@@ -5,10 +5,24 @@ import { isOpaqueNetworkFailure } from "../apps/authoring/src/fetchFailure.ts";
 // Sentry DEMOS-2X / DEMOS-2Y. `fetchVersions` and `loadStarterExample` both raise
 // `TypeError: Failed to fetch` when the tab's own network drops mid-request — not a
 // shape our host can produce at these same-origin callsites (see fetchFailure.ts).
-// The four `false` rows below are the point: they pin that a host-side failure (a
+// The `false` rows below are the point: they pin that a host-side failure (a
 // Worker 5xx, a 404, or Workers-Assets' SPA-fallback HTML tripping res.json()) keeps
 // reporting, so this demotion narrows to the visitor-network population only.
 
+// Fix-prover (DEV-2859): the real production wording, verbatim from Sentry DEMOS-2X's
+// latest events (2026-09-08, releases `930f6a52` and `e32cdcdf` — both after the
+// `^failed to fetch$` gate landed in 590cb58b2 / PR #274). This is the string that
+// proves the gate never actually fired: with the anchored-no-suffix pattern reverted,
+// this assertion goes red.
+test("classifies the real production wording with the host suffix (Sentry DEMOS-2X, DEV-2859)", () => {
+  assert.equal(
+    isOpaqueNetworkFailure(new TypeError("Failed to fetch (demos.handsontable.com)")),
+    true,
+  );
+});
+
+// Regression guard: the bare wording (no host suffix) must keep matching too — this
+// passes both before and after DEV-2859, it just must not regress.
 test("classifies the observed Chrome wording (route.abort('failed'), Step 0a)", () => {
   assert.equal(isOpaqueNetworkFailure(new TypeError("Failed to fetch")), true);
 });
@@ -57,6 +71,15 @@ test("does not classify a deploy-rotated dynamic-import chunk failure (Sentry DE
   );
 });
 
+// Anchor guard (DEV-2859): the parenthesised suffix must be the END of the message.
+// Proves the trailing `$` still binds after the new optional group, not just the `^`.
+test("does not classify a suffix that isn't at the end of the message", () => {
+  assert.equal(
+    isOpaqueNetworkFailure(new TypeError("Failed to fetch (demos.handsontable.com) extra")),
+    false,
+  );
+});
+
 test("does not classify a non-Error value", () => {
   assert.equal(isOpaqueNetworkFailure("Failed to fetch"), false);
   assert.equal(isOpaqueNetworkFailure(null), false);