FE-1556: Constraint authoring prototypes with margins, sampling, and STL

[kube]

🌟 What is the purpose of this PR?

Stacked on #9371. Five playable Storybook prototypes for how optimization constraints get defined, so we can compare directions and pick one. Parameter constraints aim at sampling inside the safe region rather than pruning; state constraints aim at a margin-based robustness that feeds the objective as a continuous multiplier dropping to zero outside the safe region.

🔗 Related links

• FE-1556 (internal)
• RFC: Constraints and constrained optimisation (FE-1282, FE-1339) (internal)
• Base PR: FE-1518: Optimization constraints as shared HIR expressions #9371 (FE-1518 constraint groundwork)

🔍 What does this change?

Everything lives under src/ui/dev/constraint-prototypes/ (Storybook-only, nothing ships; see its README for the tour). Open Dev / Constraint Prototypes:

• 1 · Predicate with derived margin — constraints stay FE-1518's boolean expressions; margin, run robustness, and the smooth objective multiplier are derived (comparison slack, && = min, worst step). Includes the min-vs-mean-violation aggregation toggle that shows how min masks all but the worst constraint.
• 2 · Margin-first — the user authors the margin itself (80 - temperature, must stay ≥ 0), with the RFC's canonical rewrite offered when a comparison is typed, and the normalisation scale as an explicit control.
• 3 · Sentence builder — structured pickers compiling to the same expressions, judged over 24 seeded runs with an across-runs quorum (the RFC's chance constraint).
• 4 · Parameter sampling playground — one predicate, four samplers side by side (uniform / rejection / soft / by-construction) with a per-conjunct router: bounds fold into the box, orderings become a gap transform, the affine part is a hit-and-run polytope walk, nonlinear leftovers reject. Draw/kept/leaked counts per strategy.
• 5 · Temporal operators (extension) — always, eventually, during, within, until, atEnd as plain functions with STL quantitative robustness, plus a logsumexp smoothing temperature (exact at 0).
• Shared prototype core, unit-tested: a small expression parser/evaluator (kept syntax-compatible with the product surface), margin derivation mirroring the Python margin() on FE-1518, trace robustness, penalty multipliers, seeded samplers, and a linear-form extractor for the router.

Pre-Merge Checklist 🚀

🚢 Has this modified a publishable library?

This PR:

• does not modify any publishable blocks or libraries, or modifications do not need publishing

Files are inside @hashintel/petrinaut but Storybook-only (stories + a dev folder that no shipped entry point imports), so nothing publishes differently.

📜 Does this require a change to the docs?

The changes in this PR:

• are internal and do not require a docs change

🕸️ Does this require a change to the Turbo Graph?

The changes in this PR:

• do not affect the execution graph

⚠️ Known issues

• The prototypes evaluate with their own small expression interpreter rather than the HIR pipeline (kept synchronous so plots re-render per slider tick); the grammar matches the product surface so authored expressions lower unchanged.
• The toy cooling-tank model stands in for a real net; robustness numbers are illustrative, not calibrated.

🐾 Next steps

• Pick a direction (or a combination — e.g. 1 + 3 as two entry points over one representation, 4's router as the enforcement plan, 5 as a later extension) and promote it onto the FE-1518 schema.

🛡 What tests cover this?

• expr.test.ts, robustness.test.ts, sampling.test.ts (37 tests): parser/evaluator, margin signs vs boolean semantics, STL operators incl. windows and until, soft min/max bounds, penalty shapes, router classification, all four samplers, hit-and-run containment, toy-model determinism.

❓ How to test this?

1. yarn exec turbo run dev --filter @hashintel/petrinaut and open Storybook (port 6006).
2. Open Dev / Constraint Prototypes, stories 1–5.
3. In each: drag the run parameters to cross the safe boundary and watch robustness, the margin strip, and the scored objective respond; in 4, tighten the constraint and compare each strategy's draws/kept/leaked.

📹 Demo

Screenshots pending — placeholder for drag-drop.

[vercel]

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

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 2, 2026 2:07am UTC
petrinaut	Ready	Preview	Sep 2, 2026 2:07am UTC
petrinaut-docs	Ready	Preview	Sep 2, 2026 2:07am UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 2, 2026 2:07am UTC