Filed by the domain:devx execution seat as the successor for an out_of_scope_findings entry returned by #17602's dev (report on #17602; PR #18165). ⛔ Unlabelled for domain:* and ungraded — triage's to route and grade. ⛔ The seat did not dedupe; dedupe words are at the bottom.
The defect class, proven live once
PR #18165 fixes a real instance in scripts/git-merge-regen.mjs: the --self-test dispatch printed its failure verdict —
✗ merge driver wiring is inconsistent — N failure(s)
— and exited 0, because the exit code was set only inside fail(), while the verdict LINE is printed from a count of returned booleans. Every callee happened to call fail(), so the split was invisible until a new callee reported a failure by returning false alone. The repair sets the code from the same count the verdict is printed from:
if (failures > 0) process.exitCode = 1;
⚠️ Why it matters beyond cosmetics: pnpm check:merge-driver chains these files with &&, so a zero from any one of them is the gate passing. A callee in that state makes the whole battery advisory without saying so — the #13799 shape ("a battery that never ran is indistinguishable from one that passed"), one level in.
What is unaudited
The sibling self-tests chained beside it in check:merge-driver were not audited for the same split:
scripts/git-env.mjsscripts/check-regen-pending.mjs
⛔ Neither is known to have the defect — this card is an audit, not a bug report against them. ⛔ Do not write a PR body claiming they are broken before measuring.
Suggested shape of the answer
Per-file, the question is mechanical: is there any path on which the dispatch prints a failing verdict (or counts a failure) while process.exitCode stays 0? The decisive reading is an ablation per file — regress one callee to report failure by return false alone, run the dispatch, read the exit code, not the output.
⭐ If the answer is "the split is possible in more than one of them", the better fix is one shared verdict/exit helper rather than three copies of if (failures > 0) — but ⛔ that is a judgement for whoever takes it, on measurements, not a prescription from this card.
Acceptance
- Each chained file in
check:merge-driver is either shown immune (with a firing control — an ablation that does produce a non-zero exit, so the reading discriminates) or repaired.
- ⛔ No gate weakened, no exemption added: every change here can only turn a previously-green-but-failing run red.
- State the population — which files
check:merge-driver actually chains, re-derived from the script, ⛔ not copied from this card. This card's list is a hint, not a reading.
Dedupe words
check:merge-driver · verdict exit code · self-test exits 0 · assertion floor · advisory battery
⚠️ Related but ⛔ not asserted as duplicates: #13799 (the assertion-floor population, domain:devx, p2) is the parent class and may already cover this; #16717 records check:merge-driver staying red on macOS for unrelated reasons. Check both before grading.
Generated by Claude Code
Filed by the
domain:devxexecution seat as the successor for anout_of_scope_findingsentry returned by #17602's dev (report on #17602; PR #18165). ⛔ Unlabelled fordomain:*and ungraded — triage's to route and grade. ⛔ The seat did not dedupe; dedupe words are at the bottom.The defect class, proven live once
PR #18165 fixes a real instance in
scripts/git-merge-regen.mjs: the--self-testdispatch printed its failure verdict —— and exited 0, because the exit code was set only inside
fail(), while the verdict LINE is printed from a count of returned booleans. Every callee happened to callfail(), so the split was invisible until a new callee reported a failure by returningfalsealone. The repair sets the code from the same count the verdict is printed from:pnpm check:merge-driverchains these files with&&, so a zero from any one of them is the gate passing. A callee in that state makes the whole battery advisory without saying so — the #13799 shape ("a battery that never ran is indistinguishable from one that passed"), one level in.What is unaudited
The sibling self-tests chained beside it in
check:merge-driverwere not audited for the same split:scripts/git-env.mjsscripts/check-regen-pending.mjs⛔ Neither is known to have the defect — this card is an audit, not a bug report against them. ⛔ Do not write a PR body claiming they are broken before measuring.
Suggested shape of the answer
Per-file, the question is mechanical: is there any path on which the dispatch prints a failing verdict (or counts a failure) while
process.exitCodestays 0? The decisive reading is an ablation per file — regress one callee to report failure byreturn falsealone, run the dispatch, read the exit code, not the output.⭐ If the answer is "the split is possible in more than one of them", the better fix is one shared verdict/exit helper rather than three copies of
if (failures > 0)— but ⛔ that is a judgement for whoever takes it, on measurements, not a prescription from this card.Acceptance
check:merge-driveris either shown immune (with a firing control — an ablation that does produce a non-zero exit, so the reading discriminates) or repaired.check:merge-driveractually chains, re-derived from the script, ⛔ not copied from this card. This card's list is a hint, not a reading.Dedupe words
check:merge-driver·verdict exit code·self-test exits 0·assertion floor·advisory batterydomain:devx, p2) is the parent class and may already cover this; #16717 recordscheck:merge-driverstaying red on macOS for unrelated reasons. Check both before grading.Generated by Claude Code