Load Sparkle without PyObjC's class scan (~150 MB less memory on macOS)

[ThomasWaldmann]

Description

get_updater() loads the Sparkle framework with objc.loadBundle('Sparkle', globals(), bundle_path). PyObjC's default scan_classes=True does not just register Sparkle: it calls objc_getClassList(), realizes every Objective-C class in the process (~67,000 on macOS 15 once Qt's Cocoa plugin and the system frameworks are loaded) and wraps each one as a Python class that is kept in this module's globals for the lifetime of the app. That costs ~150 MB of physical footprint and only happens in the frozen app bundle (the updater is only set up when sys.frozen), which is why the shipped Vorta.app idles at ~220–250 MB while a venv install of the same code idles at ~80–120 MB.

This PR loads only the framework (scan_classes=False) and looks up the one class Vorta uses explicitly with objc.lookUpClass('SUUpdater'). Nothing else changes: SUUpdater was the only bare name taken from the scan (hence the old # noqa: F821), other Sparkle classes remain reachable through objc.lookUpClass, and objects returned from Sparkle are wrapped lazily by PyObjC either way.

Related Issue

• Why is vorta using 500MB+ memory when in tray? #337 (Vorta using several hundred MB in the tray)
• What are the current stable GUIs? (Currently using Vorta) borgbackup/borg#10328 (asks for a lighter Borg GUI; Vorta's ~300–400 MB on macOS is the concern)

Motivation and Context

Vorta is a tray-resident app, so its idle footprint is what users see all day. Measured on the shipped 0.11.5 bundle right after launch (scratch $HOME, no backup run yet): 218 MB with the main window hidden, 250 MB with it shown; heap reported 66,870 realized ObjC classes and 71,919 Python objects (94 MB) allocated by the class scan. A pip install of the same version on the same machine: 91 MB / 120 MB. A 150-cycle borg create run showed no per-backup growth, so the class scan is the dominant cost.

How Has This Been Tested?

• New unit test tests/unit/test_updater.py (macOS only): asserts objc.loadBundle is called with scan_classes=False and SUUpdater is obtained via objc.lookUpClass.

• pytest tests/unit: 265 passed, 7 skipped. ruff check / ruff format --check clean.

• Two app bundles built with PyInstaller from this tree (macOS 15.7.9, Apple M3, Python 3.11, PyQt6 6.6.1), launched under a scratch $HOME with the main window hidden and measured 35 s after launch:

  	physical footprint	realized ObjC classes
  master (ebf7571)	228.8 MB	66,928
  this branch	78.6 MB	2,322

• Isolated check with PyObjC 12 and the real Sparkle.framework: the default scan costs +62 MB and adds 15,559 names to the module, of which 57 come from Sparkle; scan_classes=False + lookUpClass costs +1 MB and returns the same class.

Side note for maintainers: building from master needed the spec's datas to collect assets/UI recursively (the dialogs' .ui files moved into subdirectories), otherwise the bundle fails at startup on assets/UI/dialogs/exception.ui. That is unrelated to this change and not included here.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Checklist:

• I have read the CONTRIBUTING guide.
• My code follows the code style of this project.
• My change requires a change to the documentation.
• I have updated the documentation accordingly.
• I have added tests to cover my changes.
• All new and existing tests passed.

I provide my contribution under the terms of the license of this repository and I affirm the Developer Certificate of Origin.

🤖 Generated with Claude Code

[ThomasWaldmann]

@m3nu Have a look!

Looks like Claude Fable 5.1 found a nice tweak to reduce memory usage significantly.

Also, it found some other issues while doing that, see the other PRs.