FFmpeg libswscale Re-implementation

libswscale is the FFmpeg component responsible for image resizing and pixel-format conversion. Callers reach it through a stable C API, so this task is about reproducing both the library interface and the resulting image behavior.

The reference build used in the task is FFmpeg's scalar C path rather than its assembly-optimized variants, which keeps the comparison on a from-scratch rewrite instead of architecture-specific hand tuning that already exists upstream.

Starting from FFmpeg's libswscale reference (compiled without assembly optimizations), the agent must rebuild it as a C-compatible shared library in Zig or Rust. That includes both image scaling and pixel-format conversion, with enough fidelity that callers can treat the submission as a drop-in replacement.

• Implement the public C API expected by the hidden harness.
• Handle resizing and format conversion across a mix of image shapes and color layouts.
• Beat FFmpeg's scalar C baseline once correctness thresholds are satisfied.

The hidden verifier first runs correctness workloads covering format conversion and image scaling. Only submissions that clear per-plane PSNR thresholds are eligible for the benchmark phase. If any correctness workload fails, the score is zero; otherwise the verifier takes a geometric mean of hidden benchmark speedups against FFmpeg's scalar C baseline.

• Format-only workloads use a stricter quality threshold than scaling workloads.
• Random-seeded images are used so the task is about implementation fidelity, not memorizing test fixtures.
• Crashes or failures on benchmark workloads are heavily punished by the geometric mean.

The task runs on CPUs only with 8 CPUs, 64 GB RAM, no GPUs, and no internet access. The reference library is compiled without assembly so the baseline is fair to a from-scratch rewrite rather than a duel against FFmpeg's most architecture-specific hand tuning.