@noiselang/core@0.8.0

Minor Changes

• 7744ce7: plot::line(xs, ys) — the two-argument form plots a vector of values (numbers, estimates, or random variables, banded by ±1 sd like the one-argument form) against an explicit deterministic numeric x-axis, with the axis named after its source array. Histogram/introspection draw collection now runs on the parallel sampling driver (bit-identical stream, multicore wall-clock).
• 7744ce7: Slider values are now true runtime uniforms in every backend (PLAN-UNIFORM-INPUTS P1): the WGSL and WASM emitters lower input:: cones instead of declining them, and the value rides the dispatch (a GPU params block / a host-written memory region) rather than the compiled artifact. Dragging a slider re-dispatches cached pipelines and cached kernel instances with zero recompiles, and slider-heavy documents GPU-accelerate — turboquant.noise drops from ~14 s to under 0.1 s on the GPU CLI.

@noiselang/core@0.7.0

Minor Changes

• d38da29: Sliders work everywhere a number does, plus onehot and "did you mean?" hints.

  input::real evaluates to a symbolic value (Value::Sym), a fourth scalar shape beside
  Num/Est/Dist, and most builtins had never grown an arm for it — every math:: ufunc,
  norm/normsq/normalize/mse, max/min/cummax/cummin/quantize, count/any/all,
  and poisson/geometric/normal_complex all rejected a slider. They now defer symbolically, so
  the value stays an input (no recompile on drag) and gives the same answer as the constant it
  currently holds. A pair of tests locks this in: each builtin is exercised with a slider against the
  literal, and any new math/vec/rand name that is neither probed nor explicitly exempt fails
  the suite.

  New onehot(v, width) — the length-width indicator row of v: 1 at index v, 0 elsewhere.
  A random v yields a row of indicator RVs; a constant v folds to a constant array.

  Unknown module calls now suggest the nearest registered name (math::sqtr → "did you mean
  math::sqrt?"), with an edit-distance budget that stays quiet for genuinely unrelated names.

@noiselang/core@0.6.0

Minor Changes

• b1729c1: Array slicing: indexing with a deterministic array of indices selects into a new array. Since
  a..b is already an array, xs[0..r] takes the first r elements (Rust-style half-open), and
  any index list works — xs[[2, 0, 0]] reorders and repeats, xs[perm] applies a deterministic
  permutation. Each index obeys the existing scalar rules (constant non-negative integer, bounds
  checked element-wise); random indices inside the array are an error. It is pure build-time sugar
  for [for i in inds { xs[i] }], so nothing changes for codegen. The secretary example's
  observation phase collapses from a fold loop to bar = max(quality[0..r]).

@noiselang/core@0.5.1

Patch Changes

• cd83e72: Fix the engine worker failing to load in bundled production builds. The worker URL was hoisted into
  a variable, which stopped bundlers from recognizing new Worker(new URL('./worker.js', import.meta.url))
  as a worker — so worker.js was copied verbatim instead of bundled, and its import './gpu-protocol.js'
  resolved to a file that was never emitted. Every run then hung forever waiting on a worker that had
  died on load. Dev servers were unaffected, so this only appeared in production bundles.