fix: stability batch - self-pipe busy loop, UIA walker, console flash, screenshot fallback, emoji crash

[Jeomon]

Six commits addressing reported stability bugs, each landing with regression tests. Full suite: 555 passed.

Closes #392, #382, #371, #369. Mitigates #301 / #385.

---

fix: rebuild the asyncio self-pipe instead of pinning a CPU core — #392

asyncio wakes its event loop through a loopback socketpair. A Windows session or display transition tears that idle pair down, and BaseSelectorEventLoop._read_from_self reads the resulting EOF, hits its break, and returns without unregistering the socket. It stays permanently readable, so select() returns instantly on every pass and the server burns a full core — silently, with no exception and nothing in the log — until the process is killed.

Reproduced on this stack (CPython 3.14.3, _WindowsSelectorEventLoop), killing the write half on an otherwise idle loop:

           healthy:  cpu=0.00s  wall=3.01s  ->   0.0% of one core
  self-pipe closed:  cpu=2.98s  wall=3.00s  ->  99.5% of one core   (zero stderr)
   self-pipe reset:  cpu=2.89s  wall=3.00s  ->  96.3% of one core   (84k tracebacks in 3s)

The reporter filed this upstream as python/cpython#156333 against the proactor loop and proposed a guard on _loop_self_reading. That patch would be dead code here: serve has forced WindowsSelectorEventLoopPolicy since 757940d, so we reach the same dead end by a different path. He explicitly predicted the selector loop would spin too — the numbers above confirm it.

install_selfpipe_guard() wraps _read_from_self: after the stock method runs, it peeks the socket (BlockingIOError = healthy, immediate empty read = dead peer) and rebuilds a dead pipe through the loop's own _close_self_pipe() / _make_self_pipe(). ConnectionResetError from the abortive-close variant is handled the same way. Idempotent, no-op on a healthy loop.

Rebuilding is deliberately unbounded. An earlier draft capped it and unregistered the reader after 20 attempts; the test for that path proved it unsurvivable, since select() with no registered descriptors raises WinError 10022 on Windows and under the stdio transport the self-pipe is frequently the only descriptor. A WARNING at 20 rebuilds keeps a pathological machine visible instead.

Verification: the suite was run against stock asyncio with the guard stubbed out — 5 of 10 tests fail, the spin measuring 0.94 of a core. test_wakeups_survive_the_rebuild asserts a call_soon_threadsafe from another thread still wakes the loop afterwards, which the watchdog's focus callbacks depend on. The real server was also driven end to end (serve --transport stdio, MCP initialize round-trip in 1.94s, clean exit).

Worth noting for the future: WindowsSelectorEventLoopPolicy is deprecated and slated for removal in 3.16. When it goes, the default returns to the proactor loop and the upstream-reported path becomes live for us again.

feat: desktop service module with screen state management and UIA integration — #301, #385

RawViewWalker.GetFirstChildElement returns NULL for every element on some configurations (Windows ARM64 under x86-emulated Python), while FindAll on the same element enumerates children normally — turning every window into an apparent leaf and emptying the accessibility tree. _ViewWalkerState cross-checks the first few empty walks with FindAll and latches onto whichever path works, so steady-state cost is one enumeration per element either way.

is_overlay_window also no longer treats "no children" alone as proof of an overlay; when child enumeration degrades, that heuristic filters out the entire desktop. A real app window always has a title, so both signals are now required.

These issues are hardware-specific and could not be reproduced on this machine, so the commit is a mitigation on reported symptoms rather than a confirmed fix — hence no closing keyword for either.

fix: stop PowerShell child processes flashing a console window — #369

A console child started from a server with no console of its own gets a brand new console allocated: a window flashes and steals keyboard focus on every tool call. Redirecting the streams does not prevent the allocation — only CREATE_NO_WINDOW does. Applied to the child and to the taskkill fallbacks in the timeout path, composed with the existing CREATE_NEW_PROCESS_GROUP so the CTRL_BREAK_EVENT graceful-stop path is unaffected.

fix: fall back to mss when a screenshot backend returns an empty frame — #371

On VM/RDP desktops the primary backend returns an undecodable frame after the first capture. Detect the empty frame and fall back rather than handing the caller a corrupt image. Reported against VM/RDP specifically, which is not reproducible here; the fallback logic itself is covered by tests.

fix: repair UTF-16 surrogate pairs before serializing the UI tree — #382

Snapshot crashed with UnicodeEncodeError whenever the UI tree contained emoji. Lone surrogates are repaired before serialization.

fix: import Display and SemanticNode in the desktop service

Missing imports in desktop/service.py.

---

Testing: pytest — 555 passed. ruff check and ruff format --check clean on all touched files.