# Test runners

Normally, run `npm test`. Use `npm test -- --offline` to skip live upstream fetching
(this also skips the WG17 syntax check, which is discovered live and has no
offline snapshot).

For development, run a focused check directly from the repository root:

```sh
node test/run-regression.mjs        # all regression sections
node test/run-regression.mjs docs   # documentation checks only
node test/run-conformance-all.mjs   # conformance layers; accepts --offline
node test/run-iso-strict.mjs        # strict ISO core
node test/run-iso-part2-amendment.mjs
node test/run-neumerkel.mjs         # live upstream; --cached reproduces last fetch
node test/run-neumerkel-tests.mjs   # upstream-fetch harness
node test/run-examples.mjs
node test/run-playground.mjs
node test/run-architecture.mjs
node test/run-properties.mjs        # seeded random-term invariant checks (see below)
node test/run-openrulebench.mjs
node test/run-http-json.mjs
node test/run-interop.mjs           # requires the comparison engines
```

These runners retain their existing options; there is no separate npm alias for
each one. Focused checks do not replace the full release gate.

`run-properties.mjs` is different in kind from the rest of the suite: instead
of hand-picked inputs, it generates many random ground terms from a seeded
PRNG and checks invariants that must hold for any input. It covers term/list/
arithmetic/string library predicates (write/read round-tripping, `sort/2`
ordering and idempotence, `append/3`/`length/2` agreement, `compare/3`
symmetry, `keysort/2` stability, `=..` round-tripping, `copy_term/2` sharing,
`reverse/2` involution, `atom_codes/2`/`atom_chars/2`/`char_code/2`/
`number_codes/2` round-tripping, `succ/2` and `abs/1`/`sign/1` arithmetic
identities, `nth0/3`/`nth1/3` agreement, `last/2`, `min_list/2`/`max_list/2`
bounds, `sum_list/2` additivity, `atom_concat/3`/`string_concat/3` regrouping,
`atom_string/2`/`number_string/2` round-tripping, `list_to_set/2` idempotence,
`permutation/2`, `between/3` range generation, `split/3`/`join/3` inversion)
and, more broadly, control constructs, exceptions, the database, and grammar
rules (`catch/3` ball delivery, `findall/3` order/duplicate/no-match
behavior, double negation, the exact `existence_error` indicator for an
undefined predicate, `assertz/1`+`retract/1` leaving no trace, a DCG list
rule accepting exactly the list it was built from, `bagof/3` agreeing with
`findall/3`, and `numbervars/3` counting distinct variables) -- around 35
properties, each reported as its own trial (currently ~530 individual test
lines) so a failure names exactly which trial of which property broke rather
than leaving a reader to dig through one aggregate error. The whole file
still runs in about two seconds. The seed (`SEED` in that file) is fixed so a
failure is exactly reproducible by rerunning it -- change it only
deliberately, and say why, never to make a transient failure disappear.

There is no separate wall-clock benchmark harness: `npm test`'s own elapsed
time is the project's performance indicator.

See the [conformance guide](conformance/README.md) for report maintenance.