perf(#862): clip the Local2D composite to its non-identity pixels

[dfattal]

Implements #862, generalising #858 from a whole-frame special case to a per-region one — and correcting the premise I filed #862 on.

What it does

The composite reads the weave from a scratch that is a copy of the target, so wherever the pass reduces to weave it rewrites bytes already present:

mode	identity where
ALPHA_OVER / implicit LERP	outside the 2D rects
ZONES	inside a zone interior with no 2D over it — outside every zone it genuinely clears, and a feather ramp is not the identity either
explicit authored mask	never clips

So the pass is covered with up to 8 disjoint rects: the four bands around the largest feather-inset zone interior, plus the part of the 2D bbox inside it. One clipped draw per rect, and the weave snapshot copies only those rects. Disjointness is required by vkCmdCopyImage; the draws would be idempotent regardless (blending off, output is a pure function of the sampled inputs). Bails out if the cover exceeds 80% of the region. DXR_L2D_CLIP=0 restores the old path for A/B.

The feather insetting is what makes a feathered zone clippable at all — the leftover is a ring, not a box, which is why a single scissor rect could never fire on the shipping avatar config (default feather is 16 px, 24 px via env).

Measured — and an important correction

Avatar, 811×1421, 2D band + feathered zone. Cover resolves to 4 rects = 32% of the region:

	GPU busy
full region (DXR_L2D_CLIP=0)	53.87%
clipped (#862)	53.09%

Mean of 3 alternating pairs, clipped lower in every pair — so ~0.8 GPU points, real but modest.

This is much less than #862 predicted, and the prediction was wrong for an instructive reason: I read composite=5.9 ms from [FRAME_STAGES] as the cost of the composite. That stage actually spans the composite + HUD recording and the submit and whole-frame GPU drain — at a paced 60 fps it mostly measures the wait. Cutting the fill by 68% correctly did nothing to it. This PR relabels the stage composite+wait= so the next person doesn't optimise the wrong thing, and the issue is updated with the correction.

Where it should matter more than on the avatar: apps whose Local2D overlay is small relative to a large region (non-zones path clips straight to the 2D bbox) — modelviewer's toast/HUD, gauss, Unity zones apps.

Verification

• Rendering eyeball-verified twice by David during development: rounded corners, bubble, weave and punch-through all correct.
• Composite tap inspected: bubble crisp over the weave, bake underneath, no seams at the band boundaries.
• 60 fps, zero errors/validation messages across runs.

🤖 Generated with Claude Code