C and Python coverage baseline — 22 August 2026

This is the first measured coverage baseline for the current native and Python cores. It records evidence, not a quality score: no percentage is a required gate, and the build contains no fail-under option. A future ratchet should be considered only after the uncovered paths have been classified and exercised where a representative test is possible.

Reproduction

The measurement started from commit dd3a65a and includes the coverage targets introduced with this report. Run it from the repository root on the supported Linux/POSIX environment:

uv sync --frozen
make coverage

The target runs every C test with GCC -O0 -g --coverage, then runs all Python tests through coverage.py with branch measurement enabled. Generated data is disposable and remains below build/coverage/:

build/coverage/c/coverage.txt
build/coverage/c/coverage.xml
build/coverage/c/index.html
build/coverage/python/coverage.json
build/coverage/python/html/index.html

The C object tree is isolated at build/coverage/c-build. The parser path test still writes its temporary input below its established fixed build/tests/c location, so the target creates that directory before running the isolated binaries.

Measured environment:

Debian GNU/Linux 13
Linux 6.12.101+deb13-amd64 x86_64
GCC and gcov 14.2.0
Python 3.13.5
coverage.py 7.15.4 with C extension
gcovr 8.6

The exhaustive native witness test produces legitimately large execution counters. Therefore the target disables gcovr’s suspicious-hit magnitude heuristic with --gcov-suspicious-hits-threshold 0; it does not discard parse errors or negative counters.

Baseline results

The complete C suite passed. Coverage over executable statements in src/ was:

The lowest line percentages among measured native modules were 88% in src/solver/solver_serial.c and src/builder/yang_zhang.c, 90% in the failed-leaf trace writer, and 92% in the formula parser. All measured native functions were entered at least once.

All 64 Python tests passed in 51.6 seconds. Coverage over python/ was:

The immutable region model, tileset, native library loader, reduction and region adapters, tiling checker, and witness checker reached 100% of their measured opportunities. This does not establish correctness by itself; their independent semantic and differential tests remain the relevant evidence.

Significant uncovered paths

The baseline exposes several useful groups rather than one undifferentiated percentage:

• Native builder, parser, trace, and solver misses are concentrated in integer overflow guards, allocation failures, short reads or writes, mmap and finalization failures, and cleanup after partially initialized state. These need controlled fault injection rather than oversized ordinary fixtures.
• Formula parser misses include out-of-memory, stream I/O, and failed-close paths. Its normal syntax and domain rejection cases are already broadly exercised; parser fuzzing is a separate kind of evidence and is not claimed by this baseline.
• Solver misses include defensive failures while allocating arrays, metrics, the byte-support table, deduplication storage, snapshots, and result copies, plus verifier and trace-finalization rejection. They align with the need for focused ownership and cleanup tests before concurrent execution is added.
• Python misses are concentrated in cross-check assertions for internally inconsistent solver/oracle results and in adapter rejection of malformed native ABI output or non-sequence inputs. The main immutable models and independent tiling validation paths are substantially better covered.

The native percentage includes the serial sources linked into libwang.a. The empty JSON placeholder and the placeholder OpenMP implementation have no meaningful behavior to measure here. Python source discovery similarly omits empty square-to-hex placeholders. The baseline must be regenerated when those modules acquire executable behavior.

Coverage instrumentation changes optimization and runtime behavior, so these figures are not performance measurements. HTML, XML, JSON, .gcda, .gcno, and .coverage files are build artifacts and are intentionally not versioned.