Repository navigation
v2.8.0
Regression corpus. PROPERTY_DB graduates from "remember the last failing seed"
to a small per-property corpus of past failures, stored as data wherever
possible.
- A falsified property now records its minimised input, not just a seed.
When every shrunk argument is representable as data —null, scalars, arrays,
enum cases, byte strings — the input itself is stored and replayed as a single
run, reported through the new@apiRegressionViolationException
(getArguments(),getSeed()). Such an entry costs one run instead of a whole
random phase and keeps reproducing the bug even when the generation sequence
shifts. Counterexamples that cannot be represented (objects, closures, in-body
Gen::draw()values) still fall back to storing the seed and replay the random
phase as before. - Several failures per property. The corpus keeps up to 8 values entries and
2 seed entries (oldest evicted, duplicates collapsed) and replays them all
before the random phase, so fixing the newest regression no longer forgets the
older ones. An entry that no longer fails — or that the property now discards
viaAssume::that()— is pruned. - Stale entries are dropped instead of misreported. A values entry whose
argument names no longer match the property's signature (renamed, reordered or
added parameter) and a seed entry from a superseded generation-sequence epoch
are both discarded rather than replayed as a different input. The on-disk
format carries a version and is ignored wholesale when it is not the expected
one. - Storage moved from
<sha1(id)>.seedto<sha1(id)>.json. A corpus directory
written by an earlier version is not read; it is an opt-in local cache and
regenerates on the next failure.