c694896f

build: an empty diff is not an invitation to mutate the tree

a73x 2026-08-22 16:18

Commit message

build: an empty diff is not an invitation to mutate the tree

Gremlins reads a diff with no entries as "every line changed", so the
check that runs on each push became a 22-minute whole-tree sweep the
moment main was pushed and origin/main caught up with HEAD. That sweep
reaches the one mutant that SIGKILLs its own process group, so the
build did not fail so much as vanish.

Skip when no Go file differs from the base, and put the run in its own
session so a mutant cannot signal the build that started it.

Makefile

Old		New
@@ -201,8 +201,20 @@ mutate-diff: gremlins-tool
201	# first: the tree has equivalent mutants that CANNOT be killed (internal/	201	# first: the tree has equivalent mutants that CANNOT be killed (internal/
202	# cloudinit's `i+1 < len(root.Content)` is the same loop for every even-length	202	# cloudinit's `i+1 < len(root.Content)` is the same loop for every even-length
203	# YAML mapping either way), so a 100% floor would fail honest work.	203	# YAML mapping either way), so a 100% floor would fail honest work.
		204	# The empty-diff guard is not an optimisation. Gremlins treats an EMPTY parsed
		205	# diff as "everything changed" (internal/diff: IsChanged returns true when the
		206	# diff has no entries), so -D against a base that equals HEAD silently becomes a
		207	# whole-tree run — 22 minutes, and it reaches internal/agent/cloudhv, whose
		208	# `if pid != 0` mutant turns into kill(0) and SIGKILLs this make process. Which
		209	# is exactly what a push to main does: it moves origin/main to HEAD, and the
		210	# NEXT ci run has nothing to diff.
		211	#
		212	# setsid for the same reason: a mutant that signals its own process group must
		213	# not be able to reach the build that spawned it.
204	mutate-report: gremlins-tool	214	mutate-report: gremlins-tool
205	-@$(GREMLINS) unleash --workers $(GREMLINS_WORKERS) \	215	@git diff --quiet --merge-base $(MUTATE_BASE) -- '*.go' 2>/dev/null && \
		216	{ echo "mutate-report: no Go changes against $(MUTATE_BASE) — skipped"; exit 0; } \|\| \
		217	setsid $(GREMLINS) unleash --workers $(GREMLINS_WORKERS) \
206	--timeout-coefficient $(GREMLINS_TIMEOUT) \	218	--timeout-coefficient $(GREMLINS_TIMEOUT) \
207	-D $(MUTATE_BASE) . 2>&1 \| tail -6 \|\| true	219	-D $(MUTATE_BASE) . 2>&1 \| tail -6 \|\| true
208		220

RETRO.md

Old		New
@@ -9,3 +9,7 @@ One line per push to `main`: what slowed the work down. Enforced by
9	filesystem, not the tool, was producing the numbers — ext4 reported 207	9	filesystem, not the tool, was producing the numbers — ext4 reported 207
10	timed-out mutants and a flattering 100% efficacy where tmpfs reported 199	10	timed-out mutants and a flattering 100% efficacy where tmpfs reported 199
11	killed and an honest 93%.	11	killed and an honest 93%.
		12	- The diff-scoped mutation check was wired without asking what an empty diff
		13	means. Pushing main answered it: origin/main became HEAD, gremlins read a
		14	diff with no entries as "everything changed", and the whole-tree run it
		15	started was killed by the one mutant that signals its own process group.