VSE Gamma Cross

A runnable follow-up to vse-cut-list: the check renders tiny frames across a GAMMA_CROSS between signal-orange and azure strips and asserts every sample against the fade's actual math — because AI-generated sequencer code assumes the cross is the naive linear mix, and it is not.

What it witnesses: the fade math and the frame convention behind it.

• The cross blends in a gamma-0.5 space. Not (1-t)·A + t·B but ((1-t)·√A + t·√B)² — the mid-cross dips below the sRGB lerp. The endpoints are chosen to make that dip as large as a cross can: orange (1.0, 0.36, 0.0) and azure (0.0, 0.16, 1.0) have per-channel square roots summing to 1, so the gamma midpoint is the neutral (0.25, 0.25, 0.25) while the lerp midpoint is the violet (0.5, 0.26, 0.5): 0.25 darker on red and blue, the per-channel maximum. The check renders and asserts nine samples (t = 0, 1/8, …, 7/8, 31/32) within 5e-3 (2× the 8-bit quantization step + fit residual; measured 2.71e-3) and that the mid lerp deviation is material (≥0.05; measured 0.250).
• t = (frame − start) / duration, and it never reaches 1 inside the effect. The last frame of the span blends at (duration−1)/duration; B arrives only when the effect ends. That final frame (frame 32, t = 31/32) is one of the nine asserted samples, so the convention is pixel-backed end to end. An endpoint-inclusive convention is off by a full frame-step (probe: exit 5 at t 0.129 vs 0.125).
• The pixel witness demands view_transform = 'Standard' — the factory default AgX tone-maps the samples and poisons the fit (measured 0.146 on the red channel during authoring), the same class of silent failure the gallery's render standard exists to prevent. Also caught: deleting a consumed input strip orphans-and-deletes the effect (Strip 'GC' not in scene) — remove effects before their inputs.

What each check catches on failure: asserting the naive lerp as the expectation (exit 6, deviation 0.2490 at mid), the endpoint-inclusive t convention (exit 5), and swapped cross inputs (exit 4, input1=T2).

Version witness: the blend math is identical on Blender 4.5 LTS, 5.1 and 5.2 LTS — every sample matches to the quantization step. The creation contract from vse-cut-list still gates the timeline: strips (never .sequences), and new_effect ending in length= on 5.x vs frame_end= on 4.5.

The render is the evidence itself, mounted. All 32 frames of the cross are rendered by the sequencer and laid side by side as a filmstrip: the GAMMA_CROSS strip on top, and directly beneath it the same two strips crossed by the sequencer's linear CROSS effect, frame for frame. Both filmstrips are authentic sequencer pixels shown unaltered (closest-texel sampling, emission only, Standard view) on a hooded grading monitor beside a three-ball control surface. The gamma strip sinks into a dark neutral valley at mid-cross where the linear strip passes through violet; three small warm ticks mark the t = 1/2 column. A GAMMA_CROSS that was really a lerp would make the two strips identical, so the still visibly breaks with the contract.

Before shooting, the render path guards its own comparison (exit 9): every frame must fill its tile edge to edge — a 5.2 COLOR strip bakes the scene size into its width/height at creation, so strips built before the tile resolution is set letterbox into black bars (probe: in-frame spread 1.0) — and each strip's mid frame must match its closed form, so the lower strip really is the lerp (probe: wiring GAMMA_CROSS in its place fails at 0.249).

Run#

# Correctness check (tiny per-frame sample renders) — the CI check:
blender --background --python vse_gamma_cross.py --

# Falsifier: GC T2 -> T1. Must exit non-zero (inputs).
blender --background --python vse_gamma_cross.py -- --swap-inputs

# Also render the grading-monitor still (EEVEE on a GPU host; cycles on GPU-less):
blender --background --python vse_gamma_cross.py -- --output bench.png
blender --background --python vse_gamma_cross.py -- --output bench.png --engine cycles

Exit codes#

Per-script sequential checks. 9 is a valid check code; there is no rule against it.

The blender-smoke workflow runs the check on Blender 5.2 LTS and 4.5 LTS (5.1 on the weekly cron, the needs-5.1 PR label, or manual dispatch). Smoke does not pass --output. Its catalog falsifier is --swap-inputs (expects exit 4).