How we test a tiny language¶

2026-07-14

wick is small. The test suite is allowed to be small only if it is mean. We do not measure coverage theater. We measure the bug classes the language claims to kill.

Three layers¶

1. Semantics (must run)¶

A temporary main.wick full of check(condition, "label") assertions: arithmetic, short-circuit bools, optionals, lists, maps, rand/srand determinism, and now records (construct, field get/set, list of records).

Run headless:

LANTERN_FIXED_DT=1 LANTERN_SHOT=... LANTERN_SHOT_FRAME=1 ./build/lantern game/

2. Must-fail-to-compile (must not run)¶

The whole point of wick is programs that should not exist. Each case is a tiny game that must exit non-zero and print a known substring:

Case	Substring
unchecked `str?` into `lt.print`	`str?`
`if 1`	`must be bool`
assignment to undeclared name	`no implicit globals`
`"score " + 5`	`str(x)`
wrong native type / arity	`argument` / `at least`
`let x = nil`	`annotate`
unknown record field	`unknown field`
record field type mismatch	`field`

If a must-fail case starts compiling, the suite fails. That is the regression test for the constitution.

3. Real games (must paint)¶

Lantern Night (wick) self-plays under SHOWCASE_AUTO for N frames.
Package round-trip: pack showcase_wick, run folder vs .lant, cmp framebuffers; corrupt one byte, require CRC refusal.
Path sandbox: wick fixtures that attempt ../ and /etc/....
KORA kitchen: KORA_SCENE=3 frame must differ from the title screen control shot.

What we do not do¶

Re-implement the compiler in the test harness and assert against that.
Golden-master huge ASTs.
Benchmark theater before a game is slow.

Run it¶

cd lantern
cmake --build build -j
bash tests/wick_test.sh ./build/lantern
bash tests/path_sandbox_test.sh ./build/lantern
bash tests/package_test.sh ./build/lantern ./build/lantern_pack .

If you change the language, add a must-fail or a semantics check in the same commit. Same rule as the bug journal.