the recorded readings are the ones the top-level table actually produces

The verification script carried a reading taken while the absence table was
still nested inside `[c-abi]`: three failures against 2026.9.17.1, one of them
`[c-abi] has no member 'absent'`. That failure no longer exists, and recording
it would have had the sandbox run confirm a shape the engine no longer has.

Re-measured against the genuine published 2026.9.18.3 archive: fails=2, and
fails=0 on this release. Each CHANGE section now has one leg that passes on
BOTH, and the header says why that is the evidence rather than a hole -- both
new tables are top-level, an older engine ignores them, so a graph that
satisfies its requirements builds either way. What distinguishes the releases
is the refusal in each, because an engine cannot refuse what it never read.

The self-review's corresponding row moves with it.

Author: speak-agent

.agents/docs/2026-09-20-cenv-interfaces-verify.sh

@@ -16,21 +16,27 @@
 # its own directory. A section that cannot run says so and is listed again at
 # the end, because a run that reports only failures cannot be told from one that
 # examined nothing.
-# TWO RUNS, AND THE READING FROM EACH (host dry run, 2026-09-20):
+# TWO RUNS, AND THE READING FROM EACH (host dry run, 2026-09-20, against the
+# genuine published archive of the older release rather than a local build):
 #
-#   mcpp 2026.9.17.1 (published)   fails=3
+#   mcpp 2026.9.18.3 (published)   fails=2
 #     B  a requirement the implementation does not provide must be refused
-#     C  a declared absence with a known shape must be accepted
-#     C  the refusal does not name the shapes -- "[c-abi] has no member 'absent'"
+#     C  an absence with an unknown shape must be refused
 #   mcpp 2026.9.20.1               fails=0
 #
-# B's FIRST leg passes on both, and that is the documented behaviour rather
-# than a hole: `[kernel-abi]` is an unknown top-level table to an older engine
-# and is ignored, so a graph that satisfies its requirements builds either way.
-# The leg that distinguishes the releases is the refusal. C fails outright on
-# the older engine because `[c-abi-absent]` is a new key inside a table it
-# knows, where an unrecognised key is a parse error -- the asymmetry docs/22
-# records.
+# ONE PASSING LEG IN EACH CHANGE SECTION IS THE EVIDENCE, NOT A HOLE. Both new
+# tables are top-level, and an older engine IGNORES an unknown top-level table:
+# so B's first leg (a graph that satisfies its requirements) and C's first leg
+# (an absence with a known shape) build on both releases, and that is exactly
+# the backward compatibility this wave claims. What distinguishes the releases
+# is the REFUSAL in each: an older engine cannot refuse what it never read.
+#
+# An earlier revision of this file recorded three failures against 2026.9.17.1,
+# one of them `[c-abi] has no member 'absent'`. That reading was taken while
+# the absence table was nested inside `[c-abi]`, where an unrecognised member
+# is a parse error and the whole manifest was refused. Moving the table to the
+# top level is what turned that failure into the passing first leg above --
+# see docs/22 and the design's §5.7.5.
 set -u
 
 VER="${MCPP_VERIFY_VERSION:?set MCPP_VERIFY_VERSION}"

.agents/docs/2026-09-20-wave-self-review.md

@@ -59,7 +59,7 @@ status: active
 | `[c-abi-absent]` 与产物一致 | 是 | openkal-musl CI 翻转一行并要求脚本拒绝 |
 | `provides-interfaces` 与产物一致 | 是 | openkal-linux / -windows CI 重新生成并 diff |
 | 接口清单不是"打印 SURFACE.txt 的每一组" | 是 | openkal CI 的半组腿：`openkal.fs` 只导出两个名字，必须不被列出 |
-| 沙箱 CHANGE 段确实区分两个版本 | 是 | 对 2026.9.17.1 跑出 fails=3，对 2026.9.20.1 跑出 fails=0 |
+| 沙箱 CHANGE 段确实区分两个版本 | 是 | 对**真正发布的 2026.9.18.3 归档**跑出 fails=2，对 2026.9.20.1 跑出 fails=0。每个 CHANGE 段里各有一条腿两版都过——那不是洞，正是本轮主张的向后兼容：旧引擎忽略未知顶层表，所以**读不到就拒不了**，区分两版的是各自的那条拒绝腿 |
 
 **一条判据曾经不可证伪并已更正**：`CenvProbe` 的两个端到端探针测试在 gcc 宿主上 skip。
 跳过的判据等于没有判据，所以不变量被提升为 `assemble_argv` 这个纯函数，五个新测试不需要