The stamp could rebuild itself, and every shape I tried costs make -n

Split out of [[bug-a-the-mandatory-fixedpoint-step-reports-success-from-a-stale-stamp]], whose loud-stop half landed as 01dd27dd1 ("the fixedpoint step can no longer report success for sources it never saw"). The sha originally cited here, 527837d3a, was a PRE-REBASE ghost — reachable from no remote ref, and still readable in its author's own object store, which is why it never looked wrong to them. Corrected 2026-09-02 by matching the subject on origin/master; CLAUDE.md cites the same sha for the same behaviour, which is the second source. Nothing is broken today: a stamp written for other sources is refused with a one-line recovery, and running that line rebuilds correctly. This ticket is only about removing the manual step.

Why it is not just "add a witness file"

Both obvious shapes were built and measured, not reasoned about.

1. Detect in the verify recipe and recurse. rm -f $(COMPILER_STAMP); $(MAKE) $(COMPILER_STAMP) inside the $(COMPILER) recipe. GNU make executes a recipe line containing $(MAKE) even under -n, so make -n compiler/pascal26 ran the detection, deleted the stamp and exited 2. Measured: stamp empty afterwards, rc=2.

That is not a small blast radius. make -n is load-bearing here: testmgr's make_dry_run() builds its job list from it, and .claude/hooks/ no-full-suite.sh exempts dry runs precisely because they execute nothing.

2. A .PHONY-backed witness file holding the source hash, rewritten only when the hash changes, with $(COMPILER_STAMP) depending on it. Correct in a real run — measured: the honest no-op stays 0.7s and does not rebuild, and the planted stale pair rebuilds to the true fixedpoint by itself. But under -n, make must assume a PHONY prerequisite updates its target, so make -n compiler/pascal26 always reports the loop as planned. The row an honest stamp for the CURRENT sources does not re-plan the loop in tools/selfhost_stamp_devtest.sh is what caught it.

What a fix would have to do

Give $(COMPILER_STAMP) a real (non-PHONY) prerequisite whose mtime tracks "the sources last actually differed", without a rule that make must assume updates it. Candidates not yet tried:

• a parse-time $(shell ...) that refreshes the witness. Accurate under -n, but it runs on every make invocation of any target in this repo (~0.22s each) and it writes a file during make -n, which is its own violation of "-n runs nothing".
• teaching whatever writes the stamp to also touch the witness, so the witness needs no rule at all — but then nothing refreshes it when the sources change outside a build, which is the case this is about.

The honest alternative

Decide the loud stop is enough and close this. The failure it replaced was silent; what is left is one printed command. Three agents lost time to the silent version in one day and none would to this one. That is a legitimate outcome, not a cop-out — record it here rather than leaving the ticket open forever.

Positive control for whatever lands

tools/selfhost_stamp_devtest.sh already carries both arms: a stamp written for OTHER sources is refused, not read back as success must become "rebuilds", and an honest stamp for the CURRENT sources still verifies must stay green — plus a row asserting make -n compiler/pascal26 mutates nothing and does not plan the loop when the stamp is honest.