ctxloom v1.7.13
This release repairs and rebaselines the external-oracle benchmark introduced in v1.7.0.
- Validates all 15 merged-PR pins before benchmarking.
- Requires merged PRs on each repository's current default branch and indexes the exact merge commit.
- Audits exact independently resolved local import edges for JavaScript/TypeScript, Python, and Go.
- Reports comparable corpus token totals using the production
ceil(characters / 4)estimator. - Adds an explicit benchmark-only snapshot version-mismatch opt-in while keeping normal product snapshot loading strict.
The clean post-merge corpus run reports F1 0.43, source recall 0.63, graph reachability 0.93, and exact symbol/import coverage 1.00. Modeled code context falls from 1,367,753 estimated tokens without ctxloom to 71,046 with ctxloom, a 94.8% saving (19.3x weighted; 27.0x mean per PR).
Validated with 1,452 passing tests, Node 20/22 CI, CodeQL, security checks, production builds, benchmark CI, and a fresh-install publish smoke test.