Coverage evidence

Coverage answers which instrumented source ran. Parity answers whether observed behavior matched the pinned oracle. Neither a declared API nor a selected case is automatically covered.

Release 0.1.3

At released commit fd78eb80402a0d6d99e6b696d0a1ebf9ba11d5fb:

Observation	Result	Boundary
Python 3.10 parity	11,348 / 11,348	Selected public inputs on CI's runner
Python 3.12 parity	11,348 / 11,348	Selected public inputs on CI's runner
Node and browser WASM parity	11,348 / 11,348 each	Selected runtimes and corpus
Rust coverage plans	24 complete; zero failed executions	11,328 target coverage executions
Measured changed Rust lines	25 / 25 hit	Native macOS LCOV against base below
Added lines without LCOV records	7	Comments, attributes, or blank lines; no credit
WGSL source lines	Unmeasured	Rust coverage does not instrument shaders

Main CI and release CI identify that source. The changed-line base is fcfb956d172d744238a86f30129233eb00ca8b8e. These observations do not imply 100% repository coverage, every platform combination, or complete Pillow support.

The measured LCOV SHA-256 is 0622dde0a4ceba6ad0a5987e9249490742bb6c58d8d44654dbbeab2fa3d1a995. Its retained collection uses target/coverage/release-0.1.3-fd78eb804-rust.lcov with its .context.json, and the matching coverage/changed-line JSON under build/migration-parity/release-0.1.3-fd78eb804-*. Later documentation commits do not change those receipts.

Collect evidence for a change

make migration-parity-inputs-check
make migration-parity-coverage-rust
make migration-parity-changed-line-coverage MIGRATION_COVERAGE_DIFF_BASE=<base>
make migration-parity-coverage-receipt-test

Use distinct report paths for independent or incremental campaigns. The source receipt binds revision, source hashes, actual assets, instrumented binary, selected inputs, and execution status. The collector rejects source/input changes during collection. Never attach an old context to a regenerated report.

Combined reports preserve each backend's selected IDs and execution results. Unique plans form the plan denominator; test totals count executions across requested backends. A failed or incomplete backend fails collection.

Reverse coverage and missing paths

make migration-parity-pillow-coverage
make migration-parity-pillow-missing-manifest

Reverse coverage measures which pinned Pillow source the corpus reaches. It helps find missing application paths; it does not measure Rust coverage. Generated missing-feature reports are dated execution artifacts under build/migration-parity/.

Reduce repetitive parity inputs

Use make migration-parity-reduction MIGRATION_REDUCTION_ARGS='--candidates <reviewed-pairs.json> --output-dir <new-directory>' after reviewing inputs and live observations for equivalent behavior. The candidate file is an array of case_id / replacement_case_id objects; each replacement must remain selected with the same target profiles, and every requirement/profile mapping must remain represented in the corpus. Merge the requirement labels when identical workflows are consolidated. Measurement artifacts belong in ignored output directories, never in active fixture inputs.

The command measures the same canonical parity cases separately on CPU, SIMD and GPU, then tries removals in batches of 100. A changed coverage set restores half the batch and recursively checks both halves against the initial baseline. Rust executable lines, functions, regions and branch outcomes and Python lines and branches must match exactly, including their instrumented denominators. Unchanged percentages alone do not pass. External dependency source and WGSL shader internals are outside these measured locations. The GPU report must reach an actual hardware dispatch; selecting GPU while its adapter is unavailable and all work falls back to CPU does not pass.

MIGRATION_COVERAGE_PARITY_ONLY=1 omits native-only supplements from both baseline and trials; the regular coverage command continues to run them. Reports retain source, input and instrumented-build identities. The reduction command never changes the corpus. Apply accepted removals in the generator, regenerate parity/coverage/benchmark inputs, then verify parity and coverage again. Matching coverage is necessary here but is not proof that two tests protect the same behavior. Keep distinct contracts and numerical regressions.

make migration-parity-reduction-test checks the restoration and comparison guards. Reusing an output directory resumes matching completed measurements; changed sources, inputs, candidates or execution failures stop the comparison.

The manifest and maturity guide define the selected scope. Missing records, unavailable adapters, and unexecuted paths stay unmeasured.

Native font coverage observations

make migration-parity-font-native-coverage runs the supplemental font inputs against the checkout's Python binding and a test-only Rust driver. The driver calls text_bbox, getbbox_binary, and render_text_binary directly; these methods are not exposed through Python. The command writes individual returned values or API exceptions to build/migration-parity/font-native-observations.json. These are coverage probes, not parity passes: an API exception records what ran without asserting that the exception is correct. Harness failures, unknown operations, duplicate case IDs and unavailable drivers fail the command.

The full Rust coverage lane builds and instruments that driver, includes font wrapper execution in Python coverage, and binds each backend's observation report to its coverage receipt. make migration-parity-font-native-test checks the adapter and failure-reporting guards. Keep native input removals separate from generator-owned parity removals and compare fresh full-lane measurements before and after any native removal.