Workspace CI validation
The workspace failure onmain came from process-global equality read counters:
parallel unit tests could increment or reset another test’s measurement.
Unit-test counters now belong to the test thread and its current-thread Tokio
runtime. Production-coverage builds retain their global atomics.
The barrier regression fails with the original counters and passes with isolated
counters. Workspace tests, doctests, strict workspace and DB test-target Clippy,
formatting, and the multithreaded production equality contract passed locally.
Focused LLVM coverage reached 100% of lines, regions, and functions in the new
counter module.
Production coverage fingerprint
The final coverage gate also failed before this PR. Compare the complete JSON summaries emitted by these two independent Linux ARM64 CI jobs: The DB sources and coverage scripts atc834d51e match its merge into main
at 1ffadba6. The two parsed CI summaries are identical: all 187 integration
tests passed, every scope threshold passed, and vector coverage was 97.3473%
functions, 97.1564% source lines, and 94.5472% regions.
Both runs report 10,497 uncovered non-vector source lines with SHA-256
6d29b946f78eab3f28136bca75e6d0e9fa6bc3b23fc925e8fe0f4329425a5143.
The stored fingerprint still described 10,424 lines from an earlier revision.
Refreshing only its count and digest records the existing main backlog;
those lines retain their test-required classification. Scope thresholds,
vector thresholds, exclusions, dispositions, and exact fingerprint enforcement
remain unchanged. This refresh does not claim those uncovered lines are tested.