Release · 2026-08-26T13:02:04+02:00 · By Threadnote

Threadnote 4.3.8

Threadnote 4.3.8 reduces the remaining large-repository registration cost for native code-graph indexing while strengthening the permanent proportional-work ratchet.

What's new

Threadnote 4.3.8 reduces the remaining large-repository registration cost for native code-graph indexing while strengthening the permanent proportional-work ratchet.

Faster sparse admission

  • Shrinks reusable inventory receipts to the package.json and tsconfig.json content actually consumed by repository attribution, while retaining the complete derived workspace needed for cross-language indexing.
  • Reuses the parsed attribution plan across candidate discovery, exact path-backed attribution, and bounded project closure instead of rebuilding it within one admission attempt.
  • Bumps the reuse contract so older, broader receipts fail safely to normal inventory once and are replaced by the compact format.

On the pinned IntelliJ fixture, the reusable receipt model fell from 3,417 resolution-context files and 10.1 MB of encoded JSON to 87 consumed attribution manifests and 251 KB. Deep receipt validation fell from a 272.4-millisecond median to 6.7 milliseconds in the same-machine component probe.

Regression protection

  • Treats bounded-work counters as non-increasing ceilings: doing less work is accepted as an improvement, while any increase still fails the governed production ratchet.
  • Keeps exact graph and primary-query parity, one-file changed/inspected/fanout counts, timing, storage, transaction, graph-shape, and failure controls independently enforced.

The exact v4.3.7 IntelliJ observation remains retained as a failed product gate: cold indexing, total one-file time, post-scan work, parity, proportionality, transaction bounds, polyglot controls, and all failure controls passed, but registration measured 5.255 seconds against the strict 5.000-second goal. It was not rerun unchanged.

The exact clean development HEAD passed the reduced governed shape locally with four workers: cold indexing 71.075 seconds, one-file indexing 1.281 seconds, registration 300 milliseconds, post-scan 60 milliseconds, one changed and inspected file, zero dependency probes, and exact structural parity. The authoritative Linux ratchet, 10k and 100k lexical benchmarks, full test matrix, and release builds passed in pull-request CI.

Existing standalone installations upgrade with:

threadnote update