[mvaligursky]

Adds unit tests for the FrameGraph compile step, which merges neighbouring passes on the same render target into a single GPU render pass and asks a pass to store the attachments the next pass on its target loads. Only tests, no engine changes.

Changes:

• New FrameGraph#compile tests, alongside the ones added in fix: Do not merge a frame pass which does not execute with its neighbour #9784:
  • no merge of passes on different targets, or of passes without a target
  • no merge into a pass which clears its color, depth or stencil
  • no merge of a pass with its own after-pass, before-pass or internal before-pass, while the neighbouring passes they do not own still merge
  • the store request of each attachment the next pass on the target loads, per color attachment of a multisampled target with two of them, and for the depth and the stencil

[github-actions]

Build size report

This PR does not change the size of the minified bundles.

Bundle	Minified	Gzip	Brotli
playcanvas.min.js	2664.2 KB —	690.8 KB —	535.6 KB —
playcanvas.min.mjs	2661.0 KB —	689.5 KB —	535.4 KB —

[mvaligursky]

Automated PR review by Codex (OpenAI AI agent), posted through the account holder's authorized automated review process.

Reviewed ace162838. No actionable findings. This is test-only, with no runtime, public API, or bundle changes.

The new cases exercise the actual frame-graph scheduling/compilation path and check both forbidden merges and permitted neighboring merges. The multisampled two-color-attachment case first verifies that store flags start false, then distinguishes the attachment that is loaded from the one that is cleared, so the assertions are not passing merely because of default store settings.

Validation:

• Full unit suite: 4,385 passing, 43 pending.
• Full WebGPU/Metal suite: 4,384 passing, 39 pending.
• Changed-file lint passed; PR CI is green.
• All 19 FrameGraph tests pass independently.
• Additional in-memory mutation checks deliberately broke 11 merge/store rules: target matching, targetless passes, each clear guard, after/before/internal-before dependencies, color-attachment selection, depth storage, and stencil-clear storage handling. The new tests detected every fault. These probes did not modify engine files.