fix: honor byteOffset when deserializing L2 coordinate cache (Node Buffer) — v0.7.1

[guidocerqueira]

Problem

Coordinate/time arrays served from the shared L2 coordinateCache (ManagedDataset.decodedArray) came back corrupted whenever the cache returned a Node Buffer.

bytesToFloat64 rebuilt the Float64Array via:

const copy = bytes.slice(0, usable);            // Buffer.slice → VIEW, not a copy
return new Float64Array(copy.buffer, 0, usable / 8); // reads offset 0 of the POOL

ioredis getBuffer carves Buffers from a shared pool, so the returned Buffer is a view with a non-zero byteOffset. Reading from offset 0 of the underlying ArrayBuffer (ignoring byteOffset) returns shifted/garbage values.

Downstream impact (observed in nautilus-api compare-models)

A garbled, non-ascending decoded time axis broke timeWindow's binary search → the requested window collapsed to the full range with a single repeated timestamp. Every series point rendered with the same timestamp. Same corruption silently affected lat/lon coordinate arrays (nearest-point resolution). Only manifested once the L2 cache was populated (first read = L2 miss = correct), which is why it looked like "just cache".

Fix

Copy the view's own byte range with ArrayBuffer.prototype.slice, honoring byteOffset:

const ab = bytes.buffer.slice(bytes.byteOffset, bytes.byteOffset + usable);
return new Float64Array(ab);

Robust for plain Uint8Array views and pooled Node Buffers alike.

Tests

• Added a regression test in dataset-registry.test.ts that serves L2 through a cache returning a pooled Node Buffer (non-zero byteOffset). Fails on the old code ([0, 10, 20] shifted), passes with the fix.
• Full suite: 335 passing; prepublishOnly chain (typecheck/lint/test/build/validate-exports/cjs/esm interop) green.

Release

Bumps to 0.7.1 + CHANGELOG entry. After merge, tag v0.7.1 to trigger the publish workflow.