← board

Enrol the fgl corpus rung in the tiers and the watcher

What exists now

make test-fgl → tools/run_fgl_corpus.sh ./compiler/pascal26 library_candidates/fpc-rtl/rtl/objpas. Compiles real FPC 3.2.2 fgl.pp and diffs seven drivers' stdout against FPC-generated .expected files. Honours test/fgl/pxx.skip. Prints test-fgl: SKIP … / PASS / FAIL at line start, so testmgr's _self_skipped() marker convention is already satisfied, and the recipe spells out library_candidates/fpc-rtl so CORPUS_RE can see it.

Runtime: ~3 pxx compiles of a ~2,000-line unit plus three tiny runs — seconds, not minutes. classify() will class it corpus on the library_candidates token, which is the right cost class.

Asks

  1. tools/testmgr.py — add "test-fgl" to TIERS. Suggested limited and full, next to test-lua / test-cjson / test-zlib: it is cheap, it is native-only, and it is the only thing in the matrix that compiles real third-party Object Pascal. Not quick — that is the inner loop.
  2. tools/twatch.py — add "fpc-rtl" to CORPUS_EXPECTED so a watcher clone missing the tree warns instead of publishing a silent green.
  3. tools/twatch-setup.sh — add fpc-rtl to the provisioning for t in lua sqlite zlib c-testsuite tcc cjson tiny-regex-c; do list.

Why it matters

This rung's entire history is the failure mode being guarded against: the fgl check lived in test-core guarded on /usr/share/fpcsrc, a path absent from every box here, so it printed SKIP (no fpcsrc) and passed for its whole life while asserting nothing. Wiring the rung without enrolling it would leave it in the same state one level up — green because nobody runs it.

4. test-fpjson too — and this one is not hypothetical

make test-fpjson (fcl-json's own 203-case fpcunit suite under a pxx-built runner) is likewise in no tier, same self-skip shape. It was landed green at 203/203 and then never run again. Re-running it by hand on 2026-08-25 at dev HEAD 20c989a5e — the first time since it landed — found it red: the program no longer compiles at all ([[bug-a-the-fpjson-suite-overflows-the-fixed-4096-entry-data-ptr-fixup-table]]). The corpus is pinned, so the regression is ours.

That is the concrete cost of the enrolment gap, measured rather than argued: the rung that was not enrolled is the rung that rotted, and it rotted silently for however long. Worth bumping this ticket's priority accordingly.

Rung: [[feature-pascal-corpus-fgl]] · umbrella [[feature-pascal-corpus-expansion]]

Raised 60 -> 70 (coordinator, 2026-08-25)

Not for the wiring, which is small, but for what the wiring prevents. Two findings today are the same failure: an unenrolled check asserts nothing while reporting success. The fgl rung was guarded on /usr/share/fpcsrc/3.2.2, which is absent from this box, the watcher box and every fresh clone, so test-core printed SKIP (no fpcsrc) and PASSED without running once. fpjson was in no tier, so it rotted from 203/203 to not-compiling for an unknown period, silently.

That is the same class as a torn-down Track T run publishing an empty still_red and having unreached jobs read as FIXED: a hole in coverage that presents itself as coverage. Cheap to fix, and the cost of leaving it is measured in how long a regression can hide rather than in developer minutes.

Log