v2.1.37

Pull Requests

• fix(F-1007): a CPU-starved host no longer gets a healthy browser healed (#246)

Commits

• 5bd650f release: 2.1.37
• 2e6fe2b test(F-1009): the net-armed pin fails cleanly when the net is neutered (T-5156)
• bdf4d2f test(F-1009): pin the shielded drop, its bound, and that the process-kill net is armed
• 1bf0af8 test(F-1009): the process-kill net shows an empty machine to process_iter; profile_lock's read-only scan runs on every launch
• 2bbcfd8 fix(F-1009): review - fail closed on an unidentifiable browser, log silent paths, drop the socket on cancel, guard tests against real process kills
• 553fdae fix(F-1009): ask the connected browser which process it is; a squatter on the DevTools port is refused, never registered
• 66b7cac fix(F-1009): keep the concrete DevTools port in argv; believe an answer only on Chrome's own proof
• c3cb7ef fix(F-1009): a spawn can no longer attach to a sibling's browser through a stolen DevTools port
• afcc9fb test(F-999): the capture step POSTs the exact /api/echo path (a query string is a 404 with no body, on the unguarded master too)
• 0fecd76 test(F-999): the capture e2e step runs on the unguarded master first, so a clone failure is the guard's and not the step's
• e134072 fix(F-999): review - the residual is stated in full; pins for service/shared workers and network capture; Network.* frames are skipped before parsing
• 8a8250b docs(F-999): round 10 - a Document-only Fetch.enable still flakes (16/216, nothing paused); the navigation gap is an accepted, documented residual
• 6c158e9 test(F-999): a frame navigation to a pattern URL is served under Network.setBlockedURLs (measured on CI); say so
• 57f0f7f fix(F-999): the rotation guard blocks with Network.setBlockedURLs, not Fetch.enable, so a clone page's CORS POST is no longer aborted
• b1c7848 docs(T-4836): review - coverage does not run on 3.14; reflow the contract paragraph
• ba0fde1 docs(T-4836): say Python 3.14 is tested; add the 3.14 classifier
• ad7e842 review(F-1008): say why any registered match is the launch; pin a process-less Browser keeping the marker
• f8bd7c9 review(F-1008): a scan miss frees a failed launch's profile only when its pid is gone; a nodriver start that left no process frees it at once; a delegated failure keeps its grace
• f60d592 review(F-1008): a failed launch holds until its Chrome is reaped; a second claim re-dates the marker; an unanswerable age is dated by the directory; the unlink is retried
• 8a27ac4 test(F-1008): the hidepid pin reaches main's old answer instead of failing on a missing name
• e5d806a fix(F-1008): a profile reads held while our own browser is still starting on it, and unreadable, hidden or oversized lock owners resolve toward held
• 9ce56e0 fix(F-1010): a window Windows judges covered is no longer backgrounded
• 81c0078 ci(T-4804): add a Linux unit-tests cell on Python 3.14
• d0de6a4 test(T-4709): the fixture hit list records cut-off requests; a settle failure names resource timing and the document's load state
• 6f449d3 test(T-4713): liveness witness inside the writer thread instead of a tick count
• 1f710e0 test(T-4713): freeze_capture tick floor is 1, not 5 - any tick proves the loop ran