fix(pet): keep desktop pet on the active display

[wjl2023]

Summary

• keep the desktop pet on a visible area of the correct display in multi-monitor setups
• track the display of the GenericAgent window and move the pet when the app window changes monitors
• exit the desktop pet automatically when the parent GenericAgent process closes
• avoid snapping the pet back after manual dragging by only repositioning when the target display actually changes
• keep the fallback to the legacy desktop_pet.pyw launcher path in stapp.py

Validation

• python -m py_compile frontends/stapp.py
• py_compile.compile('frontends/desktop_pet_v2.pyw', doraise=True)

[lsdefine]

Thanks for the PR! The multi-monitor tracking direction is solid, and I appreciate the careful handling of drag-vs-reposition (_rect_key dedup). A few things to address before merging:

---

🔴 Critical Bug: Singleton Pet + Single Parent PID = Regression

The pet uses a singleton pattern (port 51983) — only one instance runs regardless of how many GA sessions exist. But --parent-pid binds the pet to exactly one GA process.

GA's primary usage pattern is multiple concurrent sessions (several stapp tabs in parallel), not one-GA-one-pet. So:

1. GA-1 starts → spawns pet with --parent-pid=GA1_PID
2. GA-2 starts → port taken, reuses existing pet (singleton working correctly)
3. GA-1 closes → pet detects parent dead → pet exits
4. GA-2 still running, but pet is gone

This is a regression — before this PR, closing one GA session didn't kill the pet. Now it does.

Suggested fix: Track a set of parent PIDs. Each GA that connects (or spawns) should register its PID via the existing HTTP server (e.g. POST /register?pid=xxx). The pet only exits when all registered PIDs are dead.

---

🔴 os.kill(pid, 0) Is Unreliable on Windows

def _parent_alive(parent_pid):
    try:
        os.kill(parent_pid, 0)
        return True
    except OSError:
        return False

On Windows, os.kill(pid, 0) behavior varies across Python versions — some call OpenProcess, others call TerminateProcess. Also susceptible to PID reuse.

Safer alternative:

if sys.platform == 'win32':
    kernel32 = ctypes.windll.kernel32
    h = kernel32.OpenProcess(0x100000, False, pid)  # SYNCHRONIZE
    if h:
        kernel32.CloseHandle(h)
        return True
    return False

---

🟡 Suggestions

• _find_streamlit_screen_rect is ~70 lines with two unrelated platform paths — Consider splitting into _find_screen_rect_darwin() and _find_screen_rect_win32(), each ~30 lines and independently readable.

• ctypes structs (RECT, MONITORINFO) defined inside the function — They get recreated every call. Moving them to module level avoids redundant class creation every 2 seconds.

• import re added but appears unused in this diff — please remove if not needed.

---

🟢 What's Good

• Pure helper functions (_clamp, _pet_position, _rect_key) — clean, no side effects
• argparse for CLI params — clean approach
• _rect_key preventing snap-back after manual drag — thoughtful UX detail
• stapp.py change is minimal (just passing args) — good change radius control
• PetBase abstraction with platform subclasses — good composability

---

Please fix the singleton + parent PID regression first (the critical one), then this should be good to go. The other items can be addressed here or in a follow-up. Thanks!