What is true, and how the first version of this ticket got it wrong

Filed 2026-09-03 as "the C conformance battery is skipping entirely because its corpus is not installed", escalated as a coverage hole needing a network fetch, and the premise was false. Corrected within the hour by frankuser and franka-29 (2f58b23be, addffd260), and the correction is the useful part.

seven publishes all 30 test-c-conformance jobs as PASS, job_last_pass 0975f200bd17, its most recent FULL tier (2026-09-03 12:57Z). A job cannot pass without its corpus, so the battery runs at the current tip on the host that produces the fleet's breadth.

The discriminator cost one grep and was available BEFORE the escalation. The error was reading per-checkout state as a property of the harness: the corpus is gitignored, so ls in one checkout answers about that checkout and nothing else. Two independent sessions made the identical inference from the identical skip, which is what makes it worth writing down rather than quietly deleting — the observation was correct and only the scope of the conclusion was invented.

What is actually left

Only this: make test-c in frankA or frankB delivers its test-core half and silently drops its conformance half. That is a real cost to a Track C worker following the documented gate in those trees, and it is one command to remove — tools/install_lib_candidates.sh c-testsuite, a network fetch, so it belongs to whoever owns the box rather than to an agent. p25 because the coverage it would restore locally is already being produced on seven.

The defect that WAS real is fixed and is not this ticket

test-c-conformance-cross printed test-c-conformance-cross: all targets green unconditionally, over four rows that each printed SKIP, and test-c printed base gate + c-conformance green the same way. A summary line that cannot say no, and it is the reason a local skip read as a coverage hole to two sessions in a row rather than to none. Both now branch on the suite directory (72c431bd9). Same family as the test-c-abi-glibc-oracle rows twenty lines above, which count what ran and go RED on a partial run; the difference is that a missing gitignored corpus must not be a RED for everyone, so this one reports honestly instead of failing.

THE PASCAL SIBLING, MEASURED 2026-09-06 — same corpus root, same gitignore, wider spread

Found by looking for this ticket's sibling rather than by a census, which is the cheaper search once a name exists.

tools/run_pascal_conformance.sh:34 reads its corpus from $ROOT/library_candidates/fpc-testsuite/tests/test — the same gitignored root (.gitignore:36) as the C suite. Line 95:

if [ ! -d "$SUITE" ]; then
  echo "$LABEL: SKIP — no suite at $SUITE (run tools/install_lib_candidates.sh fpc-testsuite)"
  exit 0
fi

The message is honest and names its own remedy — this is the good version of the RUNNER-ABSENT shape, not the silent one. The residual is exit 0, which is only safe while something counts skips and something expects the count.

The spread across 28 trees under /home/neo

The middle group is the trap. Those five have the parent directory, so a presence check written against library_candidates/ passes and the corpus still is not there — the same one-level-too-shallow error as checking a mount point instead of a file.

And the reassurance, stated deliberately so this is not read as an alarm

frankZ and trackt-watch both have it. So the tier and the three test-pascal-conformance job lines worked on 2026-09-05/06 ran against a real 1447-program population. Nothing green on that date is green by vacuum. As this ticket's own first version proves, the expensive error here is claiming a coverage hole that is not there.

One more thing that reads as an absence and is not

test/pascal-conformance/ in the repo contains exactly one file, pxx.skip (168 lines), and zero programs — by design. It is the skip list; the corpus is the gitignored root above. test/c-conformance/ likewise holds five pxx.skip* files and no .c. A curated 168-line skip list sitting alone in a directory reads as a populated suite that is mostly passing, and one session did read the empty directory as its missing population on 2026-09-06 while holding 1447 programs elsewhere on disk. Worth a line in each directory saying where the corpus actually lives.