FE-1528: Skip combination enumeration for token-independent lambdas

[kube]

Summary

Before this PR, a lambda's only per-combination input is the token attributes it reads. When it reads none, its rate is the same for every combination. Engine still enumerated every combination for such a lambda, evaluating the same rate C(n, w) times per frame.

Engine draws one acceptance uniform per transition per frame and fires the first passing combination, so either the first combination fires or none does. Testing only the first combination is therefore exact: same RNG stream, same consumed tokens, meaning the first w in place order, same trajectories for every seed. Measured on benchmarks/coloured-enumeration.mjs with a never-firing weight-2 arc, per run-frame:

Case	Before (token-reading lambda, enumerates)	After (token-independent lambda)
50 tokens	40 µs	0.7 µs
400 tokens	2 543 µs	0.9 µs (flat)

Before the lazy enumeration of FE-1526, the eager implementation cost 13.8 ms per run-frame on the 400-token case.

Links

• Ticket FE-1528
• Docs Simulation performance

Changes

Core

• analyzeHir marks token-independent lambdas readsNoInputTokens

  Set on the compiled artifact when the lambda reads no token attributes or counts and is deterministic; analyzeHir already tracks token reads.
  Artifacts compiled before the field existed keep enumerating.

• Both engines test only the first combination when the flag is set

  A single first-combination marking replaces the enumerator; loop bodies are untouched. Engine-supplied constant lambdas, for transitions with no lambda code, take the same path.

Benchmarks

• benchmarks/coloured-enumeration.mjs gains a token-independent case

  Existing cases now read a token so they keep measuring enumeration.

Next steps

WebGPU pair scan could use the same flag to skip its unranking loop.

Test coverage

• hir/artifacts.test.ts:

  readsNoInputTokens on compiled artifacts: set for constant and parameters-only lambdas, absent when the lambda reads token attributes or a token count, or draws randomness.

• simulation/engine/compute-possible-transition.test.ts:

  One lambda call instead of C(4,2), first tokens consumed, identical RNG state and outcomes against the enumerating path.

• simulation/monte-carlo/token-independent-lambda.test.ts:

  End-to-end seeded run with identical results when the flag is stripped.

How to test

• turbo run build test:unit --filter '@hashintel/petrinaut-core'
• cd libs/@hashintel/petrinaut-core
• node benchmarks/coloured-enumeration.mjs, imports from ../dist
• Expect token-independent case near 1 µs per run-frame from 50 to 400 tokens
• Expect token-reading cases growing with token count

[vercel]

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

Project	Deployment	Actions	Updated
hash	Ready	Preview	Sep 3, 2026 2:12pm UTC
petrinaut	Ready	Preview	Sep 3, 2026 2:12pm UTC
petrinaut-docs	Ready	Preview	Sep 3, 2026 2:12pm UTC

1 Skipped Deployment

Project	Deployment	Actions	Updated
hashdotdesign-tokens	Ignored	Preview	Sep 3, 2026 2:12pm UTC

The merge-base changed after approval.

[TimDiekmann]

re-approval of pre-rebase