Native coverage format

Karva stores durable coverage data as deterministic, versioned JSON. This is the internal interchange format used by collection, combination, and reporting. It is not coverage.py's SQLite format and it is not the exported coverage.json report.

Version 1 records the producing Karva version, collection mode, original project root, project-relative source roots, an optional run context, and files keyed by normalized project-relative path. The original root is provenance, not portable project identity; combination uses source roots, paths, and content fingerprints so artifacts can move between CI machines. Each file records a stable source-content fingerprint, executable, excluded, and executed lines, line contexts, and optional possible and executed branch arcs with their contexts. The fingerprint is lowercase hexadecimal fixed-key SipHash-128/1-3 over the source bytes.

Objects use lexicographically sorted keys. Sets serialize as sorted arrays, so equivalent coverage data produces identical bytes. Writers replace artifacts atomically. Readers accept version 1 only and report the artifact path, found version, and supported version when the schema is incompatible.

Before reporting, Karva compares each stored source fingerprint with the current source bytes. Changed or missing sources fail reporting instead of silently producing results from stale code.