Skip to content

Latest commit

 

History

History
36 lines (31 loc) · 2.38 KB

README.md

File metadata and controls

36 lines (31 loc) · 2.38 KB

Performance and selected optimizations

Tools/Benchmarks measures twelve public workloads in release mode with nine samples, three warmups and three iterations per sample. Compare equal inputs, profiles and configuration on an idle supported host. A 10% median or retention change is a review trigger, not an automatic veto.

The reviewed 100000-record stream median is 1.0749626803 seconds versus 0.968493986 seconds previously: +10.9932%. Three current runs measured 1.118063139, 1.0952811667 and 1.0749626803 seconds. The unchanged 10% trigger was exceeded and reviewed. The current cost was accepted for the local candidate; no threshold or default cap was relaxed and no causal performance improvement was claimed. A controlled A/B attribution was not performed.

These observations used MacBookPro18,2, Apple M1 Max, 32 GiB, macOS 27, Swift 6.4 (swiftlang-6.4.0.34.1), Xcode 27.2 and SDK 27.2, with configuration SHA-256 114a8f1279da751fcb1b9af9d5e0cda3e27b33757f960f114ea9262b5c70c728. Library source is unchanged by public repository preparation. Raw historical receipts remain in the maintainer's preserved local candidate envelopes.

Scalar scanning retains no reusable record scratch. SIMD/SWAR and literal/start prefilters remain unselected. Immutable plan-local regex sharing reduced logical compiler retention from 113986 to 103074 bytes (9.57%) in the preserved experiment; timing remained within observed variance. Searches retain invocation-owned state. The stream's logical peak remains 52 bytes. Repeated patterns still consume their pattern and aggregate-byte quotas; sharing does not bypass admission or work.

Logical retained bytes differ from process RSS. Darwin ru_maxrss is cumulative; allocator end observations are not isolated workload allocations. These results make no cross-machine throughput or OS-enforced memory guarantee. Source-attached structural views in unreleased Inspection carry no performance claim (G4 HOLD). Consolidated 1.7 source includes set/matrix and directed graph capabilities in JSONInspection, plus package-only store construction/preparation. Their source admission does not close G4 or establish a speed, allocation or RSS claim. Public store/follow, wider retrieval/persistence, NAV13 batches and JSON text-sequence framing remain separately gated, unadmitted work. See the combined scope and holds.