[ex3ndr]

WebKit ignores a page's ProxyConfiguration for the phone's own loopback. On an iPhone 12 (iOS 26.1), a live preview or chat link to http://127.0.0.1:PORT or http://[::1]:PORT connected straight to the phone, and localhost bypassed the proxy too. The page came from the phone, not from the agent's computer. This fix is client-only. The server and agent are unchanged: the tunnel still receives a literal loopback CONNECT target.

• Click handler / open: a page start or main-frame navigation or redirect to 127.0.0.1, [::1] or localhost (normalized) opens on ipv4|ipv6|name.happy-loopback.invalid at the same port and scheme (BrowserPolicy.Loopback).
• App proxy: BrowserProxy tunnels an alias as CONNECT <loopback>:<port>. BrowserRelay.loopback frames each request and rewrites Host, plus Origin/Referer on the alias, back to the loopback, so Vite/Next host checks pass. TLS, upgrades and responses pass through byte for byte.
• Fail closed: every BrowserPage has a WKContentRuleList that blocks anything else it loads from the phone's loopback. The exact spellings are listed in docs/browser-proxy.md.
• The address bar, errors and Share show the original address.

Tests:

• Core BrowserTests: new alias mapping and request-rewrite tests, 17/17.
• Hosted BrowserStates + Browser: 23/23 on the simulator, including the unchanged numeric-127 test testALiveArtifactWhoseComputerIsGoneSaysSoAndOpensWhenItIsBack and testComputerUnreachable….
• New testTheLoopbackOpensOnTheComputerAndNothingReachesThePhones: listeners on the phone's 127.0.0.1/[::1] get 0 hits for 127.0.0.1, [::1] and localhost. No-cors fetches to 9 loopback spellings are blocked. 127.example.com still goes through the proxy. The dev server sees Host: <loopback>:<port>.
• Physical iPhone 12: 5/5. As a negative control, removing the alias and the blocker makes the listeners record direct hits.

Reviewed by ufcvos (2 passes, nothing blocking).

Known limits, not fixed here:

• A main-frame form POST to the loopback reloads on the alias as a GET.
• Only the alias of the current CONNECT has its Origin rewritten, so a page on name. fetching ipv4. sends the alias Origin.
• Content-blocker coverage of <link rel=preconnect> hasn't been verified.
• WebRTC ICE/UDP bypasses the proxy, as it did before this change.
• A subresource on the exact loopback (e.g. fetch('http://localhost:3000') from another origin) is blocked, not rewritten.

[ex3ndr]

Hosted CI, run 42557795138: uikit and hosted3 passed. Every failure in hosted1 and hosted2 has failed before on runs without this change:

• BrowserStates testButtonsFollowHistory…: "Timed out waiting for the page to load" from the shared loaded helper. It's line 237 here because the new test fixture adds 40 lines; it was line 197 in the fix(ios): repair pairing and media reconnect; isolate UI test readiness #2000, fix(ios): discard connection snapshots after page hide #2010 and feat(ios): add read-only Cloud compute monitoring #2011 hosted2 runs, which fail the same way. It passes in this branch's sim runs and on the iPhone 12.
• AttachMenu 346 testOpenExpandClose…: seen on the providers-rebased hosted1 run.
• SendFlight 315–319/344: main 42475559210, plus the cleanup2012 hosted1 run.
• ChatNavigation 359: the feat(ios): add read-only Cloud compute monitoring #2011 and fix(ios): repair pairing and media reconnect; isolate UI test readiness #2000 hosted2 runs, and cleanup2012.
• ChiefOfStaffAsk 81, InstantSignOut 318, SystemMonitor 334 and NavigationTiming 152 are already on the known list.