Enrol test-uforth in the tiers — it is currently gated by nobody

Measured 2026-08-08, while clearing uforth's blocker chain:

$ grep -c uforth tools/testmgr.py
0

test-uforth appears in no entry of TIERS. The watcher has never run it. Every uforth fix — five compiler bugs cleared today, taking the driver suite from 2/11 to 10/11 byte-identical with CPython — is protected only by somebody remembering to type make test-uforth.

This is the same hole test-nilpy was in until 2026-08-01, and the comment recording that fix is right above where this one belongs:

test-nilpy: MAINLINE and gated ... but it was in no tier at all — so 238 of the 309 .npy files the Makefile compiles were invisible to the watcher and make test-nilpy could be RED while the full tier reported GREEN

uforth is the Track N forcing corpus — ~4300 lines of unmodified Python plus a layered .UFO stdlib, exercising closures, bound methods, Callable fields, variant receivers, the pyeval bridge and dataclasses at once. It is the single densest NilPy regression signal in the tree, and nothing runs it.

Which tier — with the numbers

limited + full, NOT native. The TIERS comment is explicit that native is the tier dev boxes gate their pushes on and the one number T must not inflate — enrolling nilpy there took the fast verdict from ~104 s to ~235 s. 46 s is the same order and does not belong there. Placement matches test-nilpy exactly.

If a fast uforth signal is wanted at native later, the pattern already exists: a quick-tier canary (test/quick_canary_nilpy.npy, broad-not-deep, ~1 s) rather than the whole corpus.

Note: it SKIPS cleanly when the clone is absent

test-uforth exits 0 with SKIP — no uforth tree at $(UFORTH_SRC) when ~/projects/uforth is missing, and the corpus half also skips when there is no python3 to be the oracle. So enrolling it cannot turn a box red for lacking the checkout — but it also means the watcher box needs the clone for this to be worth anything. Check that before closing:

git clone [email protected]:yoctobyte/uforth ~/projects/uforth

A SKIP that nobody notices is the failure mode to avoid here; consider having the tier report a skipped uforth distinctly from a passing one.

Already done (not part of this ticket)

The corpus half landed with the uforth work: make test-uforth now runs uforth's own testje*.for and tests/_drv_*.fth DIFFERENTIALLY — the same uforth.py under CPython is the oracle, so there is nothing recorded to go stale when uforth moves. 10/10 byte-identical today.

tests/_drv_file.fth is deliberately excluded by name (the ANS FILE word set, [[bug-nilpy-uforth-file-word-set-include-redefinition]]); adding it back to UFORTH_CORPUS is that ticket's gate.

Gate

tools/testmgr.py --tier limited and --tier full both run and report the uforth job; --tier native unchanged in composition and wall time.

Log

• 2026-08-08 — resolved, commit 44ba3b1d8.

Resolution (Track T, 2026-08-08) — commit d6f83cebe

Enrolled in limited + full, native untouched. Verified with --list: native 1200 jobs / 0 uforth (unchanged), limited 1656 / 1, full 2178 / 1. test-uforth PASSes in-tier in ~33 s.

Cloned yoctobyte/uforth to ~/projects/uforth on plexus — the ticket flagged this as the thing to check before closing, and it was absent, so the enrolment would have been a silent SKIP.

The ticket's own warning was the real work

"A SKIP that nobody notices is the failure mode to avoid here" — there were two such holes, and enrolling without closing them buys a green that tested nothing.

1. A self-skipping target read as a pass. test-uforth exits 0 when the tree is absent, so testmgr scored it PASS. It now detects a target that guarded its own precondition (anchored ^<target>: SKIP in its log) and marks it skip, reported as SKIP.

   The first attempt used skipped and turned a box lacking the checkout RED — the exact false red this ticket forbids, because the run loop treats skipped as a dependency failure. skip is the pre-existing pass-equivalent did-not-run status and is now pass-equivalent for the gate too. Both paths verified: absent → SKIP + GREEN + exit 0; present → PASS.

2. The corpus denominator concealed itself. The Makefile loop continued past absent corpora and reported $ok/$ok — always complete-looking.

UFORTH_CORPUS cannot be satisfied by a clean clone — for the user

The ticket records 10/10 byte-identical. A fresh clone of yoctobyte/uforth yields 4. The six tests/_drv_{c,file0,locals,string,t,x}.fth files have never existed in uforth's git history (git log --all --diff-filter=A finds no _drv_), and are not in pxx either — they are local scratch on whichever box did that work. Nothing here invented or regenerated them.

make test-uforth now says so out loud:

test-uforth: INCOMPLETE — 6 of 10 corpora absent from /home/neo/projects/uforth: ...
test-uforth: PASS — smoke + 4/10 corpora byte-identical to CPython (6 ABSENT)

Committing those six to uforth upstream would take the enrolled signal from 4 corpora to 10 with no further pxx change. That is the highest-value follow-up here and it needs the box that has them.

Known limitation, deliberately not widened

twatch maps skip → pass into tstate, so an absent-uforth box still reads green at the tstate level. That is [[bug-t-tstate-launders-skip-into-pass]], already filed; this ticket makes the skip visible in the testmgr report, which is where its own gate lives.

tests/_drv_file.fth remains excluded by name pending [[bug-nilpy-uforth-file-word-set-include-redefinition]] (the ANS FILE word set) — unchanged by this ticket, and its own gate.