Release · 2026-08-26T09:03:45+02:00 · By Threadnote
Threadnote 4.3.7
Threadnote 4.3.7 bounds the remaining large-repository registration work in native code-graph indexing.
What's new
Threadnote 4.3.7 bounds the remaining large-repository registration work in native code-graph indexing.
Code-graph registration
- Replaces a repository-wide
git ls-files --stage -v -zsemantic fallback with a bounded, checksummed Git-index fingerprint for ordinary index formats 2, 3, and 4. - Ignores only documented stat-cache fields and optional cache extensions while preserving mode, object ID, merge stage, relevant index flags, and raw path bytes.
- Fails closed to Git's canonical observation for malformed input, split or sparse indexes, unknown required extensions, source-index races, or unsupported evidence.
- Keeps Threadnote's Bun runtime free of Node-library imports.
Evidence discipline
The preceding exact v4.3.6 IntelliJ observation passed cold indexing, total one-file indexing, post-scan work, proportionality, parity, transaction bounds, polyglot controls, and all failure controls, but registration measured 5.178 seconds against the strict 5.000-second gate. That artifact remains retained; it was not rerun unchanged.
On the same pinned 45.9 MB / 278,115-entry Git index, the direct semantic fingerprint took 153.720 milliseconds. Under a controlled metadata-only source-index rewrite, the complete v4.3.6 observation took 1,430.342 milliseconds and the corrected candidate took 440.104 milliseconds.
The tree-identical optimization PR merge ref passed the governed reduced production ratchet and the 10k and 100k lexical benchmark checks. The ratchet keeps independent timing, storage, correctness, graph-shape, proportional-work, and failure controls so a win in one metric cannot hide a regression in another.
The exact v4.3.7 IntelliJ release-scale observation will be retained separately after publication.
Existing standalone installations upgrade with:
threadnote update