.gremlins.yaml

# Mutation testing configuration.
#
# A mutation run changes one operator in the source, runs the tests, and asks
# whether any test noticed. A mutant that lives is a line no test constrains.
#
# This exists because five tests written on 2026-08-28 were decoration: each
# ran the code it claimed to test, constrained nothing, and was found only by
# reverting the repair by hand and watching the test stay green. That is this,
# one mutant at a time.

unleash:
  # The keys MUST be nested under the command. A top-level `timeout-coefficient`
  # is silently ignored, and the run then reports a score computed from the
  # handful of mutants that beat the default timeout. Measured: with the key at
  # the top level, 88 of 88 mutants timed out for pkg/vfs/vfstest and gremlins
  # reported "Test efficacy: 100.00%" from the one that did not.

  # The default timeout is derived from the baseline test run and is far too
  # short here. A timed-out mutant is not a killed mutant, and a score computed
  # from 2 mutants out of 374 is not a score. Do not lower this without
  # re-reading the paragraph above.
  timeout-coefficient: 20

  # Two, not four.
  #
  # Each worker compiles and runs the package under test. Four workers on
  # pkg/lsm reached 669% CPU and exhausted this machine's memory into swap.
  # Mutation testing is the most resource-hungry thing in this repository by a
  # wide margin, and it is not worth an unusable machine to save 90 seconds.
  workers: 2

# Thresholds are deliberately absent.
#
# A floor set before the number is understood is how the coverage floor was got
# wrong in #469: it came from a developer machine and failed every pull request
# until it was re-measured from CI. Baselines live in MUTATION_BASELINE.md and
# become gates only once measured where they will be enforced.