Eve library

Phase 161 local completion evidence

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Date: 2026-07-15

Completion audit#

Phase 161 contains 150 implementation requirements across ten runtime and authoring areas, plus ten validation requirements. The implementation is present as eleven focused Rust workspaces under libs/neith/vfx-*. The source audit found no todo!, unimplemented!, placeholder, or ignored-test path.

Phase area Authoritative implementation Passing tests
161.1 asset model vfx-core 52
161.2 spawn vfx-spawn 85
161.3 attributes vfx-attribute 44
161.4 modules vfx-modules 69
161.5 renderers vfx-renderers 61
161.6 simulation vfx-sim 66
161.7 data interfaces vfx-data-interfaces 101
161.8 editor model vfx-editor 65
161.9 performance vfx-perf 60
161.10 preset library vfx-library 48
161.11 validation vfx-testing 19

All 670 tests pass with zero failures and zero ignored cases. Every workspace also passes rustfmt, strict all-target clippy, and a release build. The editor crate is a serializable authoring-domain model rather than a browser UI, so no web surface or Playwright path is affected.

The seven unchecked checklist items were audited separately. All locally executable supporting work is now present:

  • vfx-testing exposes named 1M@60 and 5M@30 stress scenarios and a release stress executable with machine-readable JSON output.
  • vfx-testing::external represents every open item as a typed fail-loud gate. Mismatched evidence cannot pass a gate. FPS evidence requires a named physical adapter and at least 120 frames; import evidence requires a fixture SHA-256 and full system equivalence; multi-GPU, dogfood, and study evidence each have explicit minimum acceptance fields.
  • Unit tests prove all seven gates are unavailable by default, reject wrong or incomplete evidence, and accept complete matching evidence.

Local stress evidence#

The release executable was run directly, rather than inferred from a scaled unit test:

bash
cargo run --release --manifest-path libs/neith/vfx-testing/Cargo.toml \
  --bin stress -- 1m60 4
cargo run --release --manifest-path libs/neith/vfx-testing/Cargo.toml \
  --bin stress -- 5m30 4
Scenario Particle updates CPU updates/s Mean CPU frame GPU certified
1M @ 60 4,000,000 45,647,020 21.907235 ms no
5M @ 30 20,000,000 46,894,424 106.62248425 ms no

These are honest CPU-reference measurements over the real attribute and spawn stack. They prove the exact particle populations allocate and update, but do not pretend to prove physical-GPU frame rate.

Open external and optional gates#

Gate Why it remains open
161.11.1 Needs a physical-GPU 1M-particle capture at or above 60 FPS
161.11.2 Needs a physical-GPU 5M-particle capture at or above 30 FPS
161.11.6 Optional; needs paired Niagara fixture and equivalence oracle
161.11.7 Optional; Unity/VFX-Graph editor and paired fixture are absent
161.11.8 Needs at least two physical GPUs for a timed soak
161.11.9 Needs a runnable Phase-42 demo embedding the VFX runtime
161.11.10 Needs at least three human Phase-27 artist participants

Hardware discovery found one Red Hat Virtio VGA adapter and one DRM render node, with no NVIDIA tooling or Vulkan inspection utility. The local Unreal 5.5 command-line editor and Niagara plugin content are present, but the optional Niagara gate calls for an importer equivalence result, not merely possession of engine-internal module assets. No Unity editor, second physical GPU, embedded Phase-42 demo, or human study cohort is available. Marking these items complete from a CPU proxy, a scheduling unit test, or an unpaired asset would weaken the acceptance criteria, so they remain open.

Verification#

Run sequentially for every libs/neith/vfx-*/Cargo.toml manifest:

bash
cargo fmt --manifest-path <manifest> --all -- --check
cargo clippy --manifest-path <manifest> --all-targets -- -D warnings
cargo test --manifest-path <manifest> --all-targets
cargo build --release --manifest-path <manifest>

All commands passed. The verification was serialized and supervised on the 15 GiB host; available memory stayed near 11 GiB and no broad workspace gate was run.