fix(scripts): correct the silent-revert incident corpus attribution

.github/workflows/silent-revert-canary.yml

@@ -4,11 +4,11 @@ name: silent-revert-canary
 # another recently-merged commit had just added.
 #
 # On 2026-08-15 three squash merges each landed a tree that dropped work a
-# sibling PR had merged minutes earlier (#2633 dropped #2632, #2639 dropped
-# #2635, #2641 dropped #2639). Every check stayed green through all three,
-# because each reverting squash removed the code AND its tests in the same
-# commit -- no suite can fail for a behavior whose tests are gone. Two issues
-# sat CLOSED as COMPLETED with their fixes absent from main.
+# sibling PR had merged minutes earlier (#2633 dropped both #2644 and #2642,
+# #2639 dropped #2635, #2641 dropped #2639). Every check stayed green through
+# all three, because each reverting squash removed the code AND its tests in
+# the same commit -- no suite can fail for a behavior whose tests are gone.
+# Two issues sat CLOSED as COMPLETED with their fixes absent from main.
 #
 # DETECTION, NOT PREVENTION -- and deliberately so.
 #

scripts/check-silent-revert.sh

@@ -13,9 +13,19 @@
 # WHY THIS EXISTS (#2691)
 # ----------------------
 # On 2026-08-15, three squash merges each landed a tree that dropped work a
-# sibling PR had merged minutes earlier. #2633 dropped #2632's rollups (853
-# lines); #2639 dropped #2635's report-ordering fix (346); #2641 dropped
-# #2639's guard work (451) -- destructive_guard.py went 1643 -> 1424 lines.
+# sibling PR had merged minutes earlier. #2633 dropped #2644's finding rollups
+# (853 lines) AND #2642's aliased GraphQL merge evidence (301) in one squash --
+# plugins/repo-fleet-hygiene/skills/audit/scripts/audit-fleet.sh went 2178 ->
+# 1700 lines with every `rollup` and `graphql` marker at zero; #2639 dropped
+# #2635's report-ordering fix (346); #2641 dropped #2639's guard work (451) --
+# destructive_guard.py went 1643 -> 1424 lines.
+#
+# Every PARENTHESIZED line count above is a blame attribution: the number of
+# lines the reverting squash deleted that `git blame` credits to that one
+# culprit commit. It is not the squash's diffstat (cc58cbc53 removed 1165
+# lines in total) and not the number of lines the culprit added. The bare
+# `2178 -> 1700` and `1643 -> 1424` figures are a different measurement --
+# whole-file line counts before and after.
 #
 # Every check stayed green through all three, and that is the point: each
 # reverting squash removed the code AND the tests covering it in the same
@@ -49,10 +59,10 @@
 # chosen.
 #
 #   Curated marker strings (#2691's own suggestion 3). Catches only what
-#   somebody pre-registered. Nobody had registered #2632, #2635 or #2639 --
-#   registration happens after you already know a fix matters, which is exactly
-#   the knowledge the incident destroys. It also decays: the list is only as
-#   fresh as the last person who remembered to append to it.
+#   somebody pre-registered. Nobody had registered #2644, #2642, #2635 or
+#   #2639 -- registration happens after you already know a fix matters, which
+#   is exactly the knowledge the incident destroys. It also decays: the list is
+#   only as fresh as the last person who remembered to append to it.
 #
 #   Merge-base staleness (the PR's branch point vs. what landed since).
 #   Tested and REJECTED on evidence: it exonerates all three real incidents,
@@ -65,10 +75,12 @@
 #   open, so it fires on almost everything.
 #
 # What is left is content: blame the lines a merge deleted and see who had just
-# added them. Measured over the last 500 first-parent commits of main, the three
-# known incidents score 853 / 451 / 346 lines against a single recent commit,
-# and the highest verified-legitimate commit scores well below the threshold
-# below. The separation is what makes the canary livable.
+# added them. Measured over the last 500 first-parent commits of main, the
+# three known incidents score 853 / 451 / 346 lines against a single recent
+# commit -- plus a fourth attribution of 301 lines on the SAME #2633 squash,
+# whose deletions trace to two different culprits and are reported separately.
+# The highest verified-legitimate commit scores well below the threshold below.
+# The separation is what makes the canary livable.
 #
 # FALSE-POSITIVE STRATEGY (the whole design rests on this)
 # -------------------------------------------------------
@@ -101,8 +113,10 @@
 # THE RESIDUAL FALSE POSITIVE, MEASURED RATHER THAN ASSUMED
 # ---------------------------------------------------------
 # At these settings, over the last 500 first-parent commits of main, the canary
-# fires 5 times -- 1%. Three are the confirmed incidents. The other two are both
-# real, and neither is a bug in the detector:
+# fires on 5 commits -- 1%. (Six findings, not five: #2633's squash deleted
+# content from two different culprits and each attribution is reported on its
+# own.) Three of the five commits are the confirmed incidents. The other two
+# are both real, and neither is a bug in the detector:
 #
 #   6f0a31109 (#2640, 390 lines) is the manual RESTORE of #2633's revert. To
 #   put back what #2633 dropped it had to delete what #2633 had added, so a

scripts/silent-revert-acknowledged.txt

@@ -26,9 +26,10 @@
 # the canary's first fire is a NEW event rather than known backlog.
 
 # #2640 is the manual restore of the content #2633's squash dropped. Putting
-# #2632's rollups back required deleting what #2633 had put in their place, so
-# a large removal of very recent content is precisely what this commit is.
-6f0a3110942375fb292112f5df5e2c39bbf56cb0  #2640 restores #2632's rollups; the removal undoes #2633's revert
+# #2644's finding rollups and #2642's aliased GraphQL merge evidence back
+# required deleting what #2633 had put in their place, so a large removal of
+# very recent content is precisely what this commit is.
+6f0a3110942375fb292112f5df5e2c39bbf56cb0  #2640 restores #2644's rollups and #2642's GraphQL merge evidence; the removal undoes #2633's revert
 
 # #2135's author noticed main move under the PR, merged origin/main into the
 # branch rather than force-pushing, and argued the disposition explicitly in the

scripts/silent-revert-incidents.txt

@@ -30,11 +30,30 @@ fires f603880dad4003477ec7cac27654fce92c75eec8 #2639 dropped #2635 (346 lines),
 # went to zero occurrences. #2618 was left CLOSED as COMPLETED.
 fires 9239f1541b1b15a2a183ad6b4067e577f931e3f1 #2641 dropped #2639 (451 lines), 10 minutes later
 
-# Found by this canary's own calibration run, NOT by the #2691 audit: the same
-# night, #2633's squash dropped #2632's finding rollups (853 lines). #2640 had
-# to restore them by hand. Nobody had filed it, which is the clearest argument
-# for automating the detection.
-fires cc58cbc53fbbb2cca68e071507b82d3e3ff896ff #2633 dropped #2632's rollups (853 lines) -- unfiled until now
+# The third merge from the same night, surfaced by this canary's own
+# calibration run rather than by the #2691 audit. #2633's squash dropped work
+# from TWO siblings at once, and the canary reports each culprit separately:
+#
+#   853 lines traced to bfb66beb8 (#2644, finding rollups and handoff plans)
+#   301 lines traced to eda5ae5ed (#2642, aliased GraphQL merge evidence)
+#
+# Both numbers are blame attributions -- the count of lines cc58cbc53 deleted
+# that `git blame` credits to that one culprit. Neither is the squash's
+# diffstat (it removed 1165 lines in total) and neither is what the culprit
+# itself added. plugins/repo-fleet-hygiene/skills/audit/scripts/audit-fleet.sh
+# went 2178 -> 1700 lines with every `rollup` and `graphql` marker at zero;
+# #2640 restored both surfaces by hand.
+#
+# This event was NOT unfiled. Issue #2656 recorded it, pinned to the same
+# commits, roughly 23 hours before this canary merged -- from the test-coverage
+# angle (the hand restore left `bare-repo-with-working-tree` uncovered) rather
+# than as a silent revert. What the canary adds is detection speed and
+# automation, not discovery.
+#
+# An earlier version of this row named #2632 as the victim. That was wrong:
+# #2632 never merged (closed unmerged, no merge commit, not an ancestor of
+# main) and its diff contains no rollup work at all.
+fires cc58cbc53fbbb2cca68e071507b82d3e3ff896ff #2633 dropped #2644's rollups (853 blamed lines) and #2642's GraphQL merge evidence (301); already recorded in #2656
 
 # --- Verified-legitimate commits that must stay quiet -------------------------
 # The closest sub-threshold miss in 500 commits: 129 lines of a doc section that