build-and-test

The project's single verification gate. tests/all.c runs every test twice — once natively (build/cx <file>) and once through self-hosted cx (build/cx cx.c <file>). A change is not done until both paths report SUMMARY: 45/45 tests passed.

The command

cc -Wall -Wpedantic -o build/cx cx.c && build/cx cx.c tests/all.c

That's it. No flags, no selection — run the whole suite. Selecting a subset hides regressions in other tests the change might have broken by accident.

Reading the output

The suite prints per-stage headers and per-test OK/FAIL lines. The line to grep for at the end is:

SUMMARY: N/45 tests passed.

• 45/45: done. Report success and stop.
• <45/45: regression. Find the first FAIL: line (or the first stage that shows lower counts on its second pass), that's where the self-host path diverged.

When it fails under self-host only

A test that passes natively but fails the second time through is almost always hitting one of the patterns in .claude/rules/cx-constraints.md:

• stacked case labels
• break inside braced switch case
• local shadowing a non-zero global
• bytecode overflow of code_pool
• a stubbed #else __cx__ intrinsic branch

Narrow the failing test to the smallest repro, then check which pattern fits.

What not to do

• Don't --no-verify or -w your way around warnings. Fix them.
• Don't skip tests because they're "flaky." The suite is deterministic; flakiness means a real bug.
• Don't claim success on partial runs.