FE-1536: Contour surface for optimization studies with local recompute

[kube]

Important

Experimental
Behind the Optimization surface feature flag.

Summary

Before this PR, Optuna-style contour surface existed only in the Experiments drawer. An optimization study is where it matters most. It already declares bounded parameters and an objective. Its trials stream back with parameter and objective values.

This adds a Surface section to the optimization detail drawer. Study trials draw as markers over a contour computed locally. Sliders and clicks navigate it. Objective mean and median stream at the selected point.

Links

• Ticket FE-1536
• Docs Sweep surface
• Docs Optimization

Changes

UI

• optimization-surface.tsx renders in the drawer after Steps

  Gated on two non-boolean optimized parameters. X/Y pickers and one slider per optimized parameter, initialized from the best trial, else the domain midpoint.

• Coarse-to-fine local walk feeds the contour

  Positions are cached with ladder seeds. Common random numbers follow the sweep rule.

• Readout streams objective mean, median, and runs at the selected point
• Completed trials draw as rings, best highlighted

  Projected the way Optuna's plot_contour projects.
  Clicking the surface navigates to the clicked position.

• ds Slider gains step and onChangeEnd pass-throughs

Core

• sampleDetachedObjective runs local compute against any net snapshot

  Lives on the experiments context, on a serialized single-worker lane.
  Compilation is cached per study, HIR artifacts plus scenario HIR.
  Objective runs as an expression metric.
  Booleans coerce to the engine's 0/1 encoding.

• Study runs its frozen input.model.definition

  Not the live editor net. Both drift the moment the user edits after launching.

• react/optimizations/surface-grid.ts builds quantized axes from the study's domains

  Log-scale domains quantize in log space. Integer domains snap to their declared step.

Known issues

Local batches queue on one serialized lane. On heavy nets the selected point's readout can wait behind walk cells. It should jump the queue. Tracked in the ticket.

Test coverage

• surface-grid.test.ts:

  Log-space quantization and integer step snapping.

How to test

• Open Petrinaut preview on Vercel
• Viewport controls > Settings > Simulation > Optimization surface
• Optimizations tab needs a deployment with the optimizer service on the /optimization route
• Local fallback: turbo run dev --filter @apps/petrinaut-website -- --with-optimizer-service
• Load example > Supply Chain Profit
• Simulate > Optimizations > Create
• Pick scenario, Optimize two numeric parameters with a search range, choose objective, Run
• Open study drawer
• Expect Surface section with sliders at best trial values and trial rings across the plane
• Expect contour filling from local simulations of the frozen snapshot
• Expect readout streaming mean and median for the selected point
• Click surface
• Expect sliders move to the clicked position

[vercel]

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

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 3, 2026 4:26pm UTC
petrinaut	Ready	Preview	Sep 3, 2026 4:26pm UTC
petrinaut-docs	Ready	Preview	Sep 3, 2026 4:26pm UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 3, 2026 4:26pm UTC

[TimDiekmann]

re-approval of pre-rebase

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 339f88e. Configure here.}