FE-1546: Stream range sweeps as one stochastic simulation

[kube]

🌟 What is the purpose of this PR?

A sweep's range selection sharded the region into quantized grid points and computed one small batch per point — slow (a scenario compile and batch per point, breadth-first leveling) and not what a range means. A range selection now runs as one stochastic simulation over the ranges: every run draws its own value per ranged parameter, the whole worker pool computes it, and the metric distribution over the region streams from the first frames.

🔗 Related links

• FE-1546 (internal)

🔍 What does this change?

• sweep-session.ts rebuilt around selections: a point runs at its value (ladder, GPU-eligible, unchanged); a range emits one batch per ladder rung whose runs carry per-run parameter values (sweepRangeRuns) — drawn continuously inside the selected value interval (the slider's quantized positions bound the interval, they do not grid it), low-discrepancy per axis and prefix-stable in the global run index, so a rung extends the exact sequence earlier rungs drew and the cache never goes stale. The cache keys whole selections; a point is a degenerate range, so the surface's per-point sampling shares it.
• ExperimentRequest.runs: per-run overrides now flow through the backend layer; the simulator, the experiment runtime, and per-shard slicing already supported them. The worker-pool backend forwards them; the WebGPU backend explicitly refuses requests with per-run values (its shader bakes parameters in) with a clear blocker.
• The provider routes range batches straight to a full-shard CPU pool — no backend walk, full parallelism — while point batches keep the chosen backend (GPU included).
• Scenario compilation needs one value per parameter, so an initial state derived from a ranged parameter holds at the range's value-space midpoint while runtime reads vary per run; documented in the user guide and arch page.
• The navigator status reads "N of M runs — sampling across the selected ranges"; the per-cell counters are gone from the sweep state.

14 sweep-session tests pin the new contract: one batch per rung with drawn runs inside the value bounds, prefix stability across rungs, point selections sending no per-run values, range revisits resuming mid-ladder with the identical drawn sequence, and distinct selections caching apart.

Pre-Merge Checklist 🚀

🚢 Has this modified a publishable library?

This PR:

• modifies an existing block or library which will be updated (@hashintel/petrinaut, @hashintel/petrinaut-core; patch changeset included)

📜 Does this require a change to the docs?

The changes in this PR:

• require changes to docs which are made as part of this PR

🕸️ Does this require a change to the Turbo Graph?

The changes in this PR:

• do not affect the execution graph

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
hash	Ready	Preview	Aug 28, 2026 8:21pm
petrinaut	Ready	Preview	Aug 28, 2026 8:21pm
petrinaut-docs	Ready	Preview	Aug 28, 2026 8:21pm

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
hashdotdesign-tokens	Ignored	Preview	Aug 28, 2026 8:21pm