[bong-water-water-bong]

#1824's check asks whether every tests/*.sh is referenced by something that can run it — need is that glob and nothing else, so a tool shipping a --self-test was invisible to it. Two were unregistered when the issue was filed; the rule below finds three, and the third is the reason the rule exists rather than the hand-picked pair:

tools/batched_repeat.py           --selftest      (not in the issue's list)
tools/whisper_encoder_baseline.py --self-test

Both are registered here through #1891's add_python_test() — batched_repeat.py --selftest (its own spelling, from before the hyphenated convention) and whisper_encoder_baseline.py --self-test. Both pass on this tree, device-free.

The rule, and why it is narrow. need now also includes any tools/*.py that declares a self-test. The blanket version ("every tools/*.py referenced by nothing") found 16 of 60 on this tree — most legitimately CLI-only — and a list that noisy is one nobody reads. A tool that ships a check and is registered nowhere is the defect; a tool that only has a CLI is not.

And the sweep was green on a disabled registration — the same defect one layer in. referenced() counted a basename anywhere in a place that can run things, so # add_python_test(...), or # add_test(...) for the 73 scripts, read as a runner. Comment-only lines are dropped now, and the mutation shows it: with both of this PR's registrations commented out, the sweep names them instead of passing.

Self-test

Two cases in the file's own style (synthetic roots; the verdict must NAME the offender):

• a self-testing tool nothing registers → drift, named;
• the same tool registered → clean, and a tools/*.py with no self-test is not a row (the negative that keeps the rule narrow).

Verified

tools/ci/tests_referenced_check.py --check --root .   -> 73 script(s) and 17 self-testing tool(s) are referenced (rc 0)
tools/ci/tests_referenced_check.py --self-test        -> 0 failure(s)
python3 tools/batched_repeat.py --selftest            -> selftest: ok
python3 tools/whisper_encoder_baseline.py --self-test -> whisper_encoder_baseline: ok
ctest -R 'tests_referenced|tool_selftest|ci_targets'  -> 100% tests passed, 0 failed out of 6
mutation: comment out both registrations              -> the sweep names both tools
tools/copyright.py --check (staged)                   -> rc 0

Fixes #1901. Refs #1824, #1891.

[bong-water-water-bong]

Closing as a duplicate of #1910 (475ba3e, opened at 11:43 — 19 minutes before this one), which does the same job better and found more of it.

What #1910 has that this does not:

• five registrations, not two — I treated referenced()'s by-basename match as proof that mtp_head_ref.py runs (two sibling tools import it) and never registered logprob_agree.py or template_bom.py. Importable is not run: their --selftest has never executed in CI. Registering all five is the correct reading;
• the per-tool flag spelling (--selftest for three, --self-test for two) is called out in the registration, which is the difference between an arm and a green no-op.

What I have handed to #1910 and would like kept, because it is a real hole in the sweep this PR extends:

referenced()  counts a basename in a COMMENT as a runner

Measured on my copy: commenting out both registrations left --check at rc 0 saying "all referenced". Same defect class as #1901 itself, one layer in — an arm that exists textually and never executes. Four lines of comment-stripping fixes it and strengthens the tests/*.sh half on the same head, which is why I suggested it there rather than in a separate PR.

I also reviewed #1912 (the third PR for this issue) and asked for it to be closed in favour of #1910, with one piece to fold in: argparse-parsed detection rather than a substring match, so a docstring mentioning --self-test does not turn a CLI-only tool into a row.

Rule for me, again: gh issue list and the board did not show a claim at the moment I took #1901; a gh pr list --search "1901 in:body" before writing would have found both.