[krisnye]

Makes an editor's first tick after a bulk load faster: 31.4–33.2 ms → 20.0–20.4 ms at 50K nodes. up load improves from 221–223 ms to 182–186 ms.

This PR is stacked on #220 (base knye/render-order), so merge #220 first. There's no public contract change. It's in 0.2.0 if it merges before 0.2.0 is published.

Benchmark

New editor load tick 50K. It spawns 50,201 nodes in a tree (a document, 200 groups, 2 subgroups each, 124 shapes and images each), then times one tick of:

• two filtered up systems;
• a layout looking up children through a hash index;
• a paint system ordered by a sorted index.

Changes

Each change was measured alternating against the version without it.

• Whole order after a rebuild: a filtered up run whose tree was just rebuilt runs the whole order without marking rows one by one. Every row already passed after a rebuild. Tick 32 → 25.1 ms.
• Parent column read once: the tree build reads each archetype's parent column once instead of looking it up per entity. One walk per entity sets every level during a build, and the order scratch is shared by every up on a thread.
• Presorted sorted index: a sorted index rebuild skips its sort when entries are already in key order, as a load in document order lists them. Paint 5.6 → 5.1 ms.
• Ordered rows rebound in place: an ordered run binds its row in place when the archetype changes. Paint 4.85 → 4.5 ms.

One trade, recorded in perf.md: up one root changed goes from 28 µs to 33 µs at 100 rounds, because its sparse path warms up later. After warm-up both take 14 µs. The other up, index and order benchmarks are within noise.

Tests

up.test.ts adds a filtered up overflowing the journal with 5,000 spawns, then one change, then an idle run. pnpm run build and pnpm test in packages/ecs pass: 134 files and 1011 tests.

🤖 Generated with Claude Code