In-language test framework: quilon test over describe / it

[assapir]

PARKED — do not merge. Opened for review only.

quilon test runs a file's top-level describe blocks. Everything else strips them, so a suite can sit beside the code it tests and cost a release build nothing.

quilon test                          # every suite under the current directory
quilon test examples/test_suite.qn   # one file
quilon test tests/                   # one directory

examples/test_suite.qn
Text
  ✓ trims both ends
  ✓ finds a part
  splitting
    ✓ splits on a separator

9 cases passed

Execution model

A suite is a .qn file with top-level describe(…) calls and no ^. The parser routes such a call into a new Program.test_blocks field rather than items, and the front end takes one of two paths (driver::TestBlocks):

• Strip — check, compile, build, run. The field is simply never read, so the blocks reach neither the type checker nor codegen.
• Run — quilon test. driver::synthesize_test_entry appends a ^ = () -> Num whose body is each block in source order followed by reportSummary(), whose Num result is the process status. Synthesized nodes get spans in a SYNTHESIZED_FILE id of their own, since spans key the type oracle.

One suite per process. Cases assert with core.test's existing assertions, which are fail-fast (__exit(101)), so a run of several suites spawns this binary once per suite with stdio inherited — a failure ends its own suite, not the whole run. Always the JIT; never a native build.

State the harness needs and the language cannot hold (a top-level := does not persist across calls) lives in four thread-local counters in quilon-rt::test_registry — __test_suite_enter/_leave, __test_case_passed, __test_passed. All are () -> Num, so codegen lowers the family through one path.

Reporter seam

The registry counts and nests; it renders nothing. All rendering is .qn, in three functions core.test exports — reportSuite, reportCase, reportSummary — and the synthesized entry calls the last one by a name the driver owns (driver::REPORTER_SUMMARY_FUNCTION). A reporter of its own defines the same three.

Stripping

describe is the marker — no cfg, no attribute. A file that is only test blocks (plus fixtures: helpers and types are fine) is not a compilation unit, so run/compile/build exit 0 in silence rather than reporting a missing ^. Verified: a compile of a file carrying both code and tests emits nothing of the harness — the test asserts that over the whole define list.

Known limitation

Cases assert with the fail-fast assertions, so the first failure reports its frame and exits 101: the failing case and everything after it go unreported, and no summary prints. The reporter claims only what it can know — "N cases passed" on a clean run, or the frame naming file:line:column where it stopped. There is no "N passed, M failed" tally across cases yet; that needs a matcher API that renders and continues (expect was cut from this PR for redesign).

Also in here

• Fixes a pre-existing --debug panic: an opaque DWARF pointee had an empty name, which LLVM rejects, so any -g build of a program holding a function value crashed. Gated by examples/higher_order.qn under debug codegen.
• docs/corelib/test.md gains the harness, the seam, and the stripping rule; docs/LANGUAGE.md, CLAUDE.md, docs/ROADMAP.md gain quilon test.
• examples/test_suite.qn is the wired-in demonstration, gated by the suite.

Review findings applied

An unparseable suite is now run (and reported) instead of silently vanishing and passing, and a path that does not exist fails rather than reporting an empty run — the worst CI failure modes. Discovery no longer follows symlinks, so a link back up the tree cannot recurse forever. A missing reporter is reported at the describe, not at a synthesized span. The tests-only decision reuses the front end's parse instead of doing its own. The argument passthrough is gone: the synthesized ^ takes no parameters and could never have read it.

Gate: cargo fmt --all -- --check clean, cargo clippy --workspace --all-targets --all-features -- -D warnings clean, cargo test --workspace 778 passed / 0 failed.

🤖 Generated with Claude Code