Fix: prevent fontconfig deadlock in forked processes

[ghost]

Summary

• Implements lazy fontconfig initialization in PixelRendererProcessor to prevent deadlock when using fork-based multiprocessing
• Adds comprehensive tests for single fork, double fork, and spawn scenarios
• Fixes deadlock issues with PyTorch DataLoader and TRL's pack_dataset()

Problem

When using PixelRendererProcessor with PyTorch's DataLoader (num_workers > 0) or TRL's pack_dataset(), processes would deadlock and hang indefinitely.

Root Cause:

• Fontconfig uses internal mutexes to protect its data structures
• When fork() creates a child process, it copies the mutex state but not the threads
• If fontconfig was initialized in the parent process, child processes inherit broken mutex state
• Child processes deadlock when trying to acquire mutexes owned by non-existent threads

Reproduction: Double fork scenario (most common in practice)

1. Create PixelRendererProcessor in main process (fontconfig initialized)
2. Call pack_dataset() which forks → First child inherits locked mutexes
3. First child creates DataLoader with num_workers>0 → Second fork inherits locked mutexes
4. Second child tries to render text → Deadlock when accessing fontconfig

Solution

Implement lazy initialization of fontconfig:

• Parent process: Only stores configuration parameters, never initializes fontconfig
• Child processes: Each child initializes fontconfig independently on first render_text() call
• Fork detection: Tracks process ID to detect forks and re-initialize when needed

Changes

pixel_renderer/processor.py:

• Store fontconfig parameters instead of initializing in __init__
• Add _ensure_fontconfig_initialized() method called before rendering
• Track initialization per-process using PID comparison
• Convert fontconfig_path to property with lazy initialization

tests/pixel_renderer/test_no_deadlock.py:

• test_no_deadlock_with_dataloader_single_fork: Basic single fork test
• test_no_deadlock_with_dataloader_double_fork: Reproduces the issue scenario (main test)
• test_no_deadlock_with_dataloader_spawn: Validates spawn still works (baseline)

Test Plan

• Run new deadlock tests: pytest tests/pixel_renderer/test_no_deadlock.py -v
• All 3 tests pass (single fork, double fork, spawn)
• Run existing test suite to ensure no regressions
• Test with real-world PyTorch DataLoader usage
• Test with TRL's pack_dataset() function

References

Fixes #13

🤖 Generated with Claude Code