[sergioperezcheco]

v1({ random: Buffer }) and v6({ random: Buffer }) must not mutate caller-owned random bytes when setting the generated node's multicast bit. Capture the six node bytes in scalar locals before writing the output, preserving ordinary-array random/rng inputs and correct bytes when random or explicit-node views overlap the output. This avoids allocating another node array, following LinusU's review; explicit node multicast bits are unchanged.

Regression coverage includes Buffer/Uint8Array and ordinary-array sources, random/rng callbacks, input/output offsets and overlap, deterministic reuse, explicit nodes and separate output buffers. With the same final tests on Node 22/macOS, previous PR production gives 13 failures / 116 passes; the update gives 129 passes. Build, Node examples (ES modules/Jest/TypeScript), full Biome lint, direct changed-file formatting, strict publint and 140 built-export public checks pass. The unchanged format:check script fails on both versions because Biome rejects --check; lint retains an informational schema-version mismatch. Node 20/24, BrowserStack, other operating systems, performance benchmarks and a fresh root frozen-lock install were not run for this revision. The earlier copy-implementation benchmark does not describe this revision.

Prepared and checked with autonomous AI assistance through Hermes Agent; no human code review is claimed.

[LinusU]

Potentially it's faster to just modify the loop a bit to xor the value when assigning the output, that way you don't have to make a copy of the entire buffer.

e.g. from

  // `node`
  for (let n = 0; n < 6; ++n) {
    buf[offset++] = node[n];
  }

to

  // Set multicast bit
  // https://www.rfc-editor.org/rfc/rfc9562.html#section-6.10-3
  buf[offset++] = node[0] | 0x01;

  // `node`
  for (let n = 1; n < 6; ++n) {
    buf[offset++] = node[n];
  }

(this should only be done if node was null and was seeded from random data)

[LinusU]

Maybe something like:

  if (node == null) {
    node = rnds.slice(10, 16);

    // Set multicast bit
    // https://www.rfc-editor.org/rfc/rfc9562.html#section-6.10-3
    buf[offset++] = node[0] | 0x01;
  } else {
    buf[offset++] = node[0];
  }

  // `node`
  for (let n = 1; n < 6; ++n) {
    buf[offset++] = node[n];
  }

[sergioperezcheco]

Changed in d28a601: the generated node now gets its multicast bit only when writing the first output byte, with no node-byte copy or input mutation. Explicit node bytes stay unchanged. The regression, full Node test suite, Node examples, lint, strict publint and an independent built-export replay pass on Node 22; I have not rerun the performance benchmark for this revision.

[Copilot]

🟡 Changes recommended

Ordinary-array inputs now throw, and overlapping random/output views corrupt v1 node bytes.

2 open issues

• Overlapping random and output views corrupt generated node bytes · New
• Using subarray breaks ordinary-array random inputs · New

What changed in this PR

This PR prevents v1/v6 node generation from modifying caller-owned random buffers while preserving multicast-bit behavior.

Changes:

• Set the multicast bit only when writing the generated node to output.
• Add regression coverage for random sources, buffer views, deterministic reuse, and explicit nodes.

File	Description
src/​v1.ts	Moves multicast-bit handling to output writes.
src/​test/​v1.test.ts	Tests input preservation across v1/v6 generation paths.

_{Review effort: Balanced}

[sergioperezcheco]

Both findings are reproducible on a53ded9. The fix captures the six node bytes in scalar locals before any output writes and uses indexed reads instead of subarray(), so ordinary-array random/rng inputs work and overlapping output cannot overwrite unread node bytes. This keeps the no-node-array-allocation approach from LinusU's review and preserves explicit node multicast bits. The full Node suite changes from 13 failures to 129 passing tests; rebuilt public-export checks, Node examples, lint and strict publint pass on Node 22/macOS. The existing format:check script rejects its --check option identically before and after the fix; direct changed-file formatting and the full Biome check pass, with the existing schema-version informational warning retained.