Coverage was remeasured on 8 August 2026 from the kodama-cpp CPU test binaries with LLVM 22.1.1 instrumentation (-fprofile-instr-generate -fcoverage-mapping). CUDA and Metal sources are excluded from this report and are validated separately on accelerator hardware.

Measure	Covered
Regions	79.98%
Functions	77.50%
Lines	80.48%
Branches	69.33%

The measurement executes kodama_cpp_tests and kodama_public_api_0_1_tests, uses the substantive numerical-test profile for source coverage, and reports project sources while excluding tests, examples, benchmarks, and generated build files. Reproduce it with:

bash tools/run_cpu_coverage.sh

The full text summary, LCOV record, and HTML source report are uploaded by the CPU coverage GitHub Actions workflow. The lowest-covered substantive module in this snapshot is now the matrix orchestration layer. Cycle 17 added public contract tests for graph weighting, index bases, pruning, embedding bridging, target errors, and resident-IVF failure behavior. graph_cluster.cpp now has 94.65% line and 73.94% branch coverage. Cycle 19 added public invalid-input, visualization-option, spatial-metadata, empty-resident-index, unavailable- backend, and Core state-machine invariant contracts without changing production code. Accelerator-only Metal stubs and resident-index code remain in the conservative CPU denominator; CUDA and physical Metal kernels are validated separately on hardware.

Cycle 20 added fixed-seed replay, monotone best-trace, guarded one-class, anti-collapse, shake-recovery, PLS-LDA input, and constant-predictor majority- fallback contracts. These behavioral assertions raise branch coverage to 65.37% and region coverage to 76.33% without changing production code; line coverage is unchanged because the affected stochastic paths were already executable.

Cycle 21 added generic raw-data, graph-only, and graph-plus-data matrix dispatch; CPU cosine graph construction; prepared-graph initialization; shape-mismatch; malformed dissimilarity; non-identity constraint; and unavailable CUDA/Metal routing contracts. Total line/branch coverage rose to 77.67%/66.78%, while kodama_matrix.cpp rose from 69.98%/57.61% to 74.83%/64.07%. No production numerical source was changed.

Cycle 22 added a public binary-graph UMAP contract covering duplicate and self-edge compaction, non-default min_dist curve fitting, explicit initialization metadata, finite float32 output, fixed-seed replay, and failure when no usable non-self edge exists. Total line/branch coverage rose to 80.35%/68.79%, while visualization.cpp rose from 70.32%/62.03% to 79.38%/70.57%. Production numerical source again remained unchanged.

Cycle 26 added resident-IVF null-view, non-positive-k, unavailable-backend, move-construction, and move-assignment contracts. Total line/branch coverage is now 80.48%/69.33%; resident_ivf.cpp is 47.79%/39.58%. The physical Metal suite separately exercises native resident-index ownership and explicit failure behavior, which cannot contribute to the portable CPU denominator.

The public-API executable is still run as a compile/link test but is not merged into source coverage. It contributes no additional project-source lines and defines a different function named main, which otherwise produces an LLVM profile-hash warning when combined with the numerical test executable.