fix: make each signal say only what the code actually checked

[bhyi4]

A first external user filed 16 items after using Yeoul on a real project (Windows/Codex, 2026-08-24).
Each was reproduced against the real binaries rather than taken as filed. Seven reproduce, and five
of those seven share one shape: a machine-readable signal asserts something the code never tested.

	the signal said	what the code actually checked
exit code	build-handoff exit 1 = failure	the truth value of the last line, [ -z "$SUMMARY" ] && echo …
flag name	arc-list --all = every arc	every arc except _archive
guard	tokens_used=0/200000 = measured	a --tokens the caller never passed
message	"install mirror-stack for sealing"	command -v am
instruction	the JOIN prompt's mandated first action	nothing — the path was never confirmed runnable

That is the failure mode this project exists to catch, in its own wiring. Grouping by that class rather
than by the reporter's P0/P1 labels is why these land in one change.

The two that were worse or elsewhere than filed

build-handoff fails only when it succeeds. With a closed arc found, [ -z "$SUMMARY" ] is false,
&& short-circuits, and that truth value becomes the script's exit status.

state	meaning	exit before
closed arc found	success	1
no closed arc	warning	0

Both paths create dev-plan.md + TODO.md, so both are now 0; the missing arc stays a warning. Real
failures (no project, already exists) still exit non-zero with a cause. Filed as an MCP-wrapper problem;
it is one line in the script.

loop-guard's budget can be disarmed by omission. Same rounds, one optional argument apart:

--tokens omitted:  CONTINUE ×5,  status → tokens_used=0/1000     ← STOP:budget can never fire
--tokens=400    :  CONTINUE ×2 → STOP:budget (used=1200 > 1000)  ← fires on round 3

Unmeasured ticks are now counted and printed with a denominator (unmeasured=3/3-ticks); a reported
budget still stops the loop, and measured ticks gain no noise. Filed P2; it is guard integrity.

Also here

• yeoul-new is atomic. A / in the topic aborted sed "s/<topic>/$TOPIC/g" after the project
  tree existed, so the retry hit "already exists". The topic is now substituted literally (bash pattern
  replacement — no delimiter to escape, no regex, and $(…)/backticks stay text), the tree is staged and
  moved into place in one step, and a failing arc-open rolls the project back and propagates its code.
• JOIN prompts emit resolved absolute paths, and arc-open refuses to print an attach command it
  could not confirm is executable. setup/install.sh explicitly does not promise bin/ on PATH, and an
  MCP-only install has no ./bin at all — so every role's mandated first action failed.
• A continuation contract in docs/BOOTSTRAP_PROMPT.md. The prompt named exactly one legitimate stop
  (Gate-1) and was silent on the rest, so an agent reasonably treated each scaffold step as the end of a
  request and the user had to type "continue" at every one. It now enumerates the stops and says
  everything between them continues as commentary. It governs sequencing, not judgment: sealing,
  PASS/KILL, convergence, graduation and publishing stay with the operator, and it is deliberately
  unenforced — an enforced "keep going" would be a mechanism for skipping gates.
• The publish guard was red on a clean checkout. It walked the whole working tree, so the gitignored
  KNOWLEDGE_INDEX.md that every arc close appends to was reported as a "personalization leak" — about a
  file git will never publish. A red that does not mean what it says gets ignored. All three scans now run
  over git ls-files, print their denominator (✓ clean (41 files scanned)), and fail on a zero-file scan.

Evidence

tests/test_gates.sh 48/48, and the summary carries its denominator — "all gate tests passed" is also
what a run that collected zero checks prints, so an empty run now exits 1.

Each fix was reverted individually and its own checks confirmed to go red (6/6), with the revert
verified by hash change. No regression in the sealed gate numbers: evasive 28/28 caught, genuine-answer
false rejection 0/39, identical across both copies. mcp/tests/test_run_contract.py 8/8 on this base.

The first revert-to-red run reported 3 of 6 as vacuous. That was the control keying on the success
wording, which a failing line never prints — a test built so it could not win. Checks now carry item tags
that both branches print. Recorded because the wrong verdict came first.

Three defects of this same class were introduced while fixing it, and the instruments caught all three,
not review: a backtick inside a double-quoted message ran am as a command and leaked its status into six
seal paths; if ! cmd; then rc=$? captured the negation's status so a failing arc-open reported exit 0;
and the first publish-guard rewrite passed all three scans while inspecting zero files, because
python3 - <<PY makes the heredoc stdin so the piped file list never arrived.

Not claimed

• Windows is unverified. No Windows environment here; the reporter's Windows path items (YL-07 and the
  Windows preflight) are neither confirmed nor disputed.
• YL-01 reproduced only for /. Backslash, quotes, Unicode and & all passed before the fix (5/5), so
  the reporter's completion criterion is broader than the defect was.
• Not addressed here (feature additions, not defects): a relay round-append tool with TURN ordering,
  long-topic duplication across 6 files, per-role model provenance, a common result envelope, and
  structural separation of finding/interpretation.
• The arc↔claim linkage the report asks for is largely already implemented (arc-prereg plus verbatim
  kill-condition injection); the gap there is documentation, not mechanism.
• 16 items from one user, one session. Treated as evidence of existence, not of frequency.

🤖 Generated with Claude Code

[bhyi4]

What the green actually covers (read from the run log, not from the badges — a job being green is not the same as these checks running there):

job	what ran
test (ubuntu-latest)	48/48 gate checks passed · ✓ clean (41 files scanned)
test (macos-latest)	48/48 gate checks passed · ✓ clean (41 files scanned)
mcp (ubuntu-latest)	all _run contract tests passed: 8/8
mcp (windows-latest)	all _run contract tests passed: 8/8

🔴 The bash suite does not run on Windows. The Windows job exercises the 8 _run contract checks and
nothing else, so the seven fixes in this PR — all of them in bin/ shell scripts — are not exercised on
Windows, which is the platform the field report came from. Four green jobs should not be read as
"verified on Windows."

Also not covered: no real CP949 code-page run (the hostile-locale test works in the ASCII band), and the
atomic-scaffold and JOIN-prompt fixes are verified against POSIX path semantics only.

Reproducing locally: bash tests/test_gates.sh (48/48; an empty run exits 1),
python mcp/tests/test_run_contract.py (8/8 — it is a script, not a pytest suite; pytest on it errors
with fixture 'tmp' not found), and bash setup/pre-publish-check.sh (fails on a zero-file scan).