Measure ewa_lanczossharp vs ewa_lanczos on real footage

[prekabreki]

Cheapest remaining unknown in the upscale chain. Both kernels are verified available in this build (ffmpeg 8.1.2 / libplacebo 7.360.1); the quality delta has never been measured, so we do not know whether the current upscaler=ewa_lanczos is leaving anything on the table.

The research doc recommended ewa_lanczossharp (docs/research/upscale_research.md §A.1), but that recommendation arrived bundled with three falsified claims, so it carries no weight on its own — it needs its own measurement.

Acceptance criteria

• HF-energy comparison (method in docs/research/upscale_research.md §Measured baseline) of ewa_lanczos vs ewa_lanczossharp vs ewa_lanczos4sharpest on >=3 frames of real 720p->1080p footage, encoder held constant.
• Speed measured for each on a >=30s clip.
• Visual 1:1 crop check for ringing/haloing, especially on high-contrast UI text.
• Result recorded in the research doc; if a kernel wins clearly, a follow-up issue to switch the default.

Files

• fetchforge/server.py:594-596 — upscaler=ewa_lanczos:downscaler=ewa_lanczos.

How to verify

• Numbers written into docs/research/upscale_research.md, replacing the "Open work" note under Tier 2.

⚠ Danger zones

• This interacts with Add FSRCNNX-8 neural upscale option via libplacebo custom_shader_path #37: if FSRCNNX-8 lands first, measure the kernel change underneath the shader, since the shader handles luma and the kernel then mostly governs chroma.

Out of scope

• Changing the default. This issue only produces the measurement.

Notes

• Needs a GPU and real source footage — not closeable from unit tests, so not scoped for a background executor.
• Original attempt was blocked when the test source (~/Downloads/Video/x2/) was deleted mid-session.