gate-selftests names `# NO-SELF-TEST:` as the remedy on a branch that returns before reading it

[jobordu]

The gate tells you to add a declaration that it then ignores on the path that prints the message

scripts/gate-selftests.sh classifies a subject that takes required arguments as
UNESTABLISHED, and prints this remedy:

⚠ check-freshness.py CANNOT BE INVOKED BARE — `--self-test` alone is rejected for MISSING
   REQUIRED ARGUMENTS, so no control is reachable this way. ⛔ That is not
      'has no self-test'; it is a limit of the invocation, and it needs a
      `# NO-SELF-TEST:` line or an invocation this gate can derive.

Adding that line changes nothing. The branch returns at :272; DECLARES_NONE is first
consulted at :340, sixty-eight lines later.

★ Two-pole proof, not inference — same tree, one edit

POLE A   tools/check-freshness.py carries `# NO-SELF-TEST: controlled by …`
         ⚠ check-freshness.py CANNOT BE INVOKED BARE …
         ⚠ UNESTABLISHED: check-freshness.py job-log.py reference-check.py

POLE B   the same file, declaration renamed to `# xx-SELF-TEST:`
         ⚠ check-freshness.py CANNOT BE INVOKED BARE …
         ⚠ UNESTABLISHED: check-freshness.py job-log.py reference-check.py

⇒ IDENTICAL BYTES.

⚠ CONTROL — the declaration is honoured on the other path: gh-complete.py carries the same
line, hits the flag-REJECTED branch at :340, and moved from UNESTABLISHED to
declared NO self-test in the same run. So this is not "declarations don't work" — it is one
branch of two.

grep -n DECLARES_NONE scripts/gate-selftests.sh
  73:DECLARES_NONE="^# NO-SELF-TEST:"
 340:  if grep -q "$DECLARES_NONE" "$f"; then      ← flag-REJECTED path: honoured
 360:  2) if grep -q "$DECLARES_NONE" "$f"; then
                                                   ← :272 bare-args path: returns first

⇒ Why this is a doctrine defect and not a cosmetic one

This repo filed #73 for exactly this shape: an absence report that does not name its remedy
converts a gap into a wall. The remedy there was that every VOID must state its disposition class.
⛔ This is the next failure along: the report names a remedy, and the remedy is inert. That is
worse than naming none, because the reader does the work and gets no signal that it did nothing —
which is how I spent a cycle adding declarations to two files and measuring no change.

⚠ And the branch has a real reason to exist, quoted from its own comment, which is why this is
a question and not a patch:

"A SUBJECT THAT CANNOT BE INVOKED BARE IS ITS OWN STATE, not 'declared none'. … Folding that
into 'has no self-test' would be a claim about the world made from a limit of the invocation."

That reasoning is sound. If a declaration should not satisfy this branch, then the branch is
right and its message is wrong. Both readings are defensible and the choice belongs to whoever
owns the gate.

Two ways to close it — they are not equivalent

1. Honour the declaration in the bare-args branch. Consult DECLARES_NONE before returning at
   :272. ⚠ Cost: it weakens the distinction the branch's comment is defending — a declared
   subject that also cannot be invoked bare becomes indistinguishable from one that simply has no
   --self-test.
2. Fix the message instead. Stop naming # NO-SELF-TEST: on this path and name what actually
   works: "an invocation this gate can derive" — i.e. a # SELF-TEST-ARGV: line, or whatever the
   gate can consume. ⚠ Cost: nothing exists to consume yet, so the remedy would be unbuildable
   until it does.

⇒ Option 2 is the honest minimum and needs no semantic decision; option 1 is the one that lets
these two instruments leave UNESTABLISHED at all.

Close condition

• Adding # NO-SELF-TEST: to a subject that takes required arguments either changes its
  verdict, or the message stops naming it as the remedy.
• A two-sided control asserting the difference: one subject where the declaration changes the
  verdict, one where it provably does not, both named in the output.

Caller: add # NO-SELF-TEST: to tools/check-freshness.py, run SUBJ_DIR=tools ./scripts/gate-selftests.sh, and diff the output against the same run with the line removed. Today
those two runs are byte-identical while the message says they should differ.

⚠ Bounds

1. ON local only. The gate's own header warns the reading is a property of the tools on that
   machine; nothing here was measured on a runner.
2. I did not establish whether reference-check.py — the third remaining UNESTABLISHED — reaches
   the same branch. Its cause is separate and recorded on tools/test_reference_check.py fails on main — 3 FAILED, and nothing reports it #280.
3. Which of the two options is correct is not established here and is deliberately left open.

---

Measured by TEAMLEAD, session 15b69750, at origin/main. Surfaced by #602,
which adds those declarations and reports honestly that two of the three buy nothing today.

[jobordu]

⛔ Reopened. I auto-closed this myself, with a qualifier the parser cannot read.

#604's body opened with:

Closes #603's decision-free half.

GitHub parsed Closes #603 and closed this issue as completed at 03:04:17Z. The words
"'s decision-free half" are invisible to the linker.

⇒ A natural-language qualifier attached to a machine-parsed keyword does not qualify anything.
The keyword is consumed by a parser that stops at the issue number; everything after it is read
only by humans, who arrive to find the issue already closed and reasonably assume it is settled.

⚠ And this is the exact defect I filed #601 about, one hour earlier

#601: "Closing an issue COMPLETED marks the measurement done. Nothing marks the consequences
done."

I filed that after finding #327 closed completed with five open conditions still depending on
the field it refuted. Then I closed this issue completed while its central question was
explicitly unresolved — and I did it with a sentence I wrote specifically to avoid that.

⇒ The failure is not carelessness about the risk. I was thinking about exactly this risk and
mitigated it in prose, in a field that is not prose.

What is actually settled, and what is not

Settled by #604 — the decision-free half:

• the message no longer names # NO-SELF-TEST: as a remedy on a path that ignores it
• it now states that the declaration does not change the verdict, why, and what would
• the branch has its first two controls, both poles (17 → 19 checks in the gate's own self-test)

⛔ NOT settled — the whole reason this issue exists:

Should a # NO-SELF-TEST: declaration satisfy the cannot-invoke-bare branch?

Both readings remain defensible and neither is mine to pick:

1. Honour the declaration at :272. ⚠ Weakens the distinction the branch's own comment
   defends — "folding that into 'has no self-test' would be a claim about the world made from a
   limit of the invocation."
2. Keep the branch strict. ⚠ Then check-freshness and job-log stay UNESTABLISHED
   permanently, while their controls exist, pass, and are gated — the gate simply cannot see them.

Today's consequence of leaving it open: 2 of the 3 remaining UNESTABLISHED instruments are in
that state for a reason unrelated to whether they are controlled.

Close condition — unchanged, and #604 does not meet it

• Adding # NO-SELF-TEST: to a subject that takes required arguments either changes its
  verdict, or the message stops naming it — ✅ the second was done in #603: stop naming an inert remedy, and give the branch its first controls #604.
• ⛔ The semantic choice above is made by whoever owns the gate. Not done, not mine.

⚠ Note the first bullet was satisfiable by the half that needs no decision — which means the close
condition I wrote was too weak, and would have let this close green on its own merits even
without the accidental link. That is the same defect 94 caught in my #598 close condition three
hours ago: a close condition satisfiable by a partial fix converts an open hazard into a closed
issue. Twice in one session, on two issues I wrote.

---

Reopened by TEAMLEAD, session 15b69750. ⇒ Practical note for anyone linking a PR here: put
Closes #N only when you mean the whole issue. To reference without closing, use a bare #N or
Refs #N — a qualifier in the sentence will not save you.