Build from COO triples. Columns are sorted ascending within each row, and duplicate `(row, col)` pairs are collapsed (max weight) so each row has uniq

Build CSR from COO (coordinate) format

Create from an existing 3-element state.

Look up A[row, col] (0.0 if absent). Columns are sorted, so binary search.

Gets an object from the cache with the given key. Returns `null` if the entry is not found.

Get the bit at (x, y)

Loads the dataset into the given table in the coordinator's database

Whether the metric uses bit-packed data

Leaf size for random projection trees.

Initial learning rate. Default: 1.0.

Local connectivity constraint. Default: 1.0.

Maximum number of candidates per iteration.

Maximum RP tree depth. Default: 200.

Multiply this matrix by a dense vector.

Number of NN-descent iterations. Default: auto.

Create a new logger with the given verbosity, optional progress callback, and stage definitions (pairs of stage name and estimated time fraction).

Create a new TauRng seeded with the given value. Matches Python: `rng_state.fill(seed + 0xFFFF)`

Create a new GPU context, uploading the data matrix and compiling the distance shader for the given metric. Returns `None` if no suitable GPU adapter