FE-1552: Reuse sweep compute, stream the selection first, and pipeline the ladder

[kube]

Important

Experimental
Behind the Parameter sweeps and WebGPU feature flags.

Summary

Before this PR, opening the sweep drawer and moving its sliders was slow because compute was spawned, serialised and thrown away. 166 web workers filled one view, one experiment ran per surface cell, ladder rungs ran one at a time, and every batch created a fresh GPU device, shader and calibration. Terminating nets plotted an empty surface.

This PR reuses that compute with pooled workers, batched surface cells and one GPU backend per session. It streams the selected point before anything else, fills the surface in quad-tree levels and pipelines the ladder. Results are unchanged: same seeds, same trajectories.

Measured on the Bench / SweepDrawer story, base against this PR:

Case	Before	After
Workers to fill the view	166	20
Workers per slider commit	16–24	1
Navigator reaches 1,000 runs after page load	1,488 ms	1,312 ms (first frames stream at 927 ms)
11×11 surface complete	~15 s sequential, or empty on terminating nets	1.6 s, filling 4 → 25 → 81 → 121 cells
GPU 1,000-run ladder	1,512–1,582 ms	937 ms
GPU command encoders per commit	75–79	68–79

Links

• Ticket FE-1552
• Docs Parameter sweeps
• Docs Sweep surface
• Docs GPU capacity calibration
• Docs Worker sharding
• Docs Experiments

Changes

Sweep view

• Terminating nets no longer plot an empty surface

  sweepCellObjective reads a cell's value from its last frame that holds samples.

• Chart data applies once per animation frame

  Latest wins, with the crossfade's epoch change latched until a frame paints.

• Chart axes mount before any data
• Switching the shown metric repaints without simulating

  Surface caches every metric's value per cell.

• Adds the Bench / SweepDrawer Storybook fixture

  Numbers above are measured on it.

Compute reuse

• createReusableWorkerFactory in core hands out leases over an idle pool

  A lease returns its worker only once every cancelled acknowledgement it owes has arrived, so no stale reply reaches the next lease.
  drain() flushes and dispose() shuts down.

• Experiments provider routes every worker-creation site through one pool per mount
• sampleCellsBatch turns many surface cells into one experiment

  Each cell's parameter values are laid out as per-run draws with pinned per-run seeds, so a cell's value matches the navigator's own runs whatever chunk sampled it. Cells whose initial marking the swept parameters shape keep the per-cell path.

• One GPU backend is cached per experiment session and leased across batches

  Device, compiled shader and learned calibration are reused.
  Cache is keyed by what shapes the shader; per-run-buffered parameter values are excluded so range selections share it.
  Calibration is also keyed per marking.

• Streamed GPU runs open with a 128-run tile

  It precedes the full-size tiles so a usable picture lands early.

Scheduling

• whenSelectionStreamed() gates surface work behind the selection's first frames

  It re-arms on every slider move.

• quadTreeLevels fills the surface corners first

  Each level then splits every region in two per axis: 4, 5, 16, 56, 40 cells on an 11×11 grid, with three chunks in flight.

• Next ladder rung starts as soon as the current one streams its first frames

  At most two are live; folds stay ordered.

• Surface chunks take one worker while the ladder computes on the CPU

  They shard across the freed pool once it idles or runs on the GPU. Shards stream partial cells as they complete.

Known issues

• On the GPU backend the pipelined ladder holds two GPU experiments concurrently for the overlap window, doubling buffer residency then. Bounded and not observed to fail
• Preview tile is invisible on light models, where every tile finishes inside one 100 ms publish window. It pays on heavy models

Test coverage

• New reusable-worker-factory.test.ts:

  Reuse, listener detachment, the owed-acks gate, quarantine of a silent worker, maxIdle, drain and dispose semantics.

• gpu-backend-cache.test.ts:

  New suite.

• runner.test.ts, metric-windows.test.ts:

  planTiles and calibrationKey.

• sweep-session.test.ts:

  sampleCells draw layout and pinned seeds, whenSelectionStreamed, foregroundActive, pipelining order and aborts, partial streaming.

• contour-grid.test.ts:

  Objective scan-back and quadTreeLevels.

• provider.test.tsx:

  One worker serves two consecutive experiments and is terminated on unmount.

How to test

• Open Petrinaut preview on Vercel
• Viewport controls > Settings > Simulation > Parameter sweeps, WebGPU
• Load example > Café Queue
• Simulate > Experiments > Create
• Sweep two numeric scenario parameters, Run
• Expect parameter strip, surface container and metric boxes with axes painted before any data
• Expect charts streaming first, then the surface filling corners-first and sharpening level by level
• Drag and release a parameter slider
• Expect charts restreaming first, surface resuming, run counter climbing smoothly through the ladder
• Switch the surface metric
• Expect plot repainting from cache
• In a browser with WebGPU, create the experiment again with Backend > GPU
• Expect 1,000-run ladder completing in about a second and slider commits streaming without a per-batch setup pause

Demo

Screenshots pending (will be handed over for drag-drop).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 3, 2026 6:36pm UTC
petrinaut	Ready	Preview	Sep 3, 2026 6:36pm UTC
petrinaut-docs	Ready	Preview	Sep 3, 2026 6:36pm UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 3, 2026 6:36pm UTC

[cursor]

PR Summary

Medium Risk
Touches worker pooling races, concurrent GPU batch leases, and sweep scheduling across CPU/GPU backends; behavior is heavily tested but regressions could affect experiment lifecycle or sweep results timing.

Overview
Sweep experiments stop tearing down compute on every batch. CPU workers are leased from a reusable pool (reset via cancel with strict ack ordering), the WebGPU path shares one cached backend and calibration per session, and streamed GPU runs can open with a small preview tile before full tiles.

Scheduling changes how work is ordered: the navigator’s selection must stream before surface sampling (whenSelectionStreamed), the refine ladder pipelines up to two rungs once the current rung has frames, and the surface fills in quad-tree levels via batched sampleSurfaceCells (many cells per experiment when marking is shared). Background CPU batches narrow while the ladder holds the pool and widen when it idles or the GPU is in use.

UI correctness/perf: surface cells read metrics from the last frame with samples (fixing terminating nets), metric charts apply plot data once per animation frame, and axes can mount before data arrives; switching surface metrics repaints from cached per-cell values.

^{Reviewed by Cursor Bugbot for commit 777f383. Bugbot is set up for automated code reviews on this repo. Configure here.}

The merge-base changed after approval.