Should a failed compiler build delete its target?

• Track U. Files if it goes ahead: Makefile (one line). The alternative costs nothing in the tree and one line of prose in CLAUDE.md, which is the owner's file.
• Found by frankS while landing af9c92a6f, from inside the failure: the positive control for a new guard passed, on output produced by the previous build, because the compile had errored and the binary was never rewritten.

The mechanism

GNU make deletes a target on recipe failure only under .DELETE_ON_ERROR. This Makefile has none, so a failed $(COMPILER_STAMP) recipe leaves compiler/pascal26 exactly as it was. Nothing is corrupt — that would be loud. What survives is a working binary built from the previous source: it runs, it exits 0, and it answers about code you have just changed.

compiler/pascal26 is untracked, so git status says nothing about it, and the converged / verified verbs — the tell CLAUDE.md already documents — are simply absent rather than wrong when the compile failed.

Why it is worse than the four routes already documented

CLAUDE.md lists a seeded tree, a reverted experiment, a sync that pulled someone's compiler/**, and the positive-control revert cycle. All four are accidents of housekeeping and are uncorrelated with your change. This one fires precisely when your change does not compile, which is exactly the moment you are about to look at output — so the two failures are perfectly correlated: the worse your change, the more certain you are to measure the old compiler. A change that compiles cannot hit it.

The fork

A — add .DELETE_ON_ERROR. The wrong measurement becomes impossible: a failed build leaves no binary and the probe cannot run at all. This is the strongest form of the guard, and it converts a silent failure into a loud one, which is what this repo's whole instrument doctrine asks for.

B — leave it, and carry a one-line rule. Neither converged after N round(s) nor self-host fixedpoint: verified prints when the compile failed, so the check is: did I see either verb since my last edit — plus make … && ./probe rather than two statements separated by ;, which is exactly the shape that runs the old binary.

The cost is what makes this a decision. Under A, every failed edit costs a pin-seeded rebuild rather than a 12-second one, and in this loop a failed edit is the common case, not the rare one. That is a tax the whole fleet pays on its most frequent action, to close a hazard that a one-line habit also closes. frankS's own position, and the reason they filed rather than did it: "that is a trade the whole fleet pays and I do not think it is mine to make."

THE TWO RATES BEING COMPARED ARE NOT COMMENSURABLE — read this before the recommendation

Added 2026-09-06 after frankuser corrected the first version of this ticket, and it is the paragraph most likely to decide the fork, so it goes above the recommendation rather than below it.

The cost of A is measurable and certain: every failed edit pays a pin-seeded rebuild, and a failed edit is the common case.

The benefit of A is measured by self-report, and self-report is blind precisely to the cases where the hazard did its damage. Both known instances share one property: the seat caught it themselves and then said so. That is the only path by which an instance becomes countable. A seat that does not catch it does not file a correction — it files a CONCLUSION, and the conclusion is specific, plausible and wrong. frankB's is the illustration: the old terminus warning fired, "which is precisely what a non-firing arm would look like". Undetected instances leave a ticket, a resolution or a closed row, not a confession.

So "twice in one day" is not a rate. It is a floor over the subset that failed visibly enough to notice.

Options weighed on observed frequency will systematically favour the status quo whenever one side's frequency is countable and the other's is not. That is the asymmetry, and it decides forks quietly and by accident. Whoever settles this should decide which error they would rather make, not which number is larger — because only one of the two numbers exists.

Recommendation

B, held weakly, and the weakness is the point. The verb guard is already in CLAUDE.md for a different reason, so B is a short addition to an existing paragraph and costs the common case nothing. A is the better guard in isolation and the worse trade at this loop's frequency.

But that comparison is exactly the one the paragraph above says cannot be made on frequency, so read this as "the cheap half first, and A stays live" rather than as a settled call. A concrete thing that would move it: any instance where the hazard was NOT self-caught — a wrong conclusion traced back to a stale binary after the fact, found in a ticket rather than in a confession. One of those is worth more than ten more self-reports, because it is the only evidence that can come from the invisible half.

The prose half is the owner's call regardless: CLAUDE.md is not edited by agents, so the sentence goes up as an ask rather than as a commit.