diff --git a/.cursor/rules/learnings-index.mdc b/.cursor/rules/learnings-index.mdc
index 6ff67a0..d54fa7d 100644
--- a/.cursor/rules/learnings-index.mdc
+++ b/.cursor/rules/learnings-index.mdc
@@ -24,6 +24,10 @@ propose a retirement, a consolidation, or a glob-scoped sub-index split.
   class they silently assert nothing.
   → .agents/skills/gem-rspock--rspock/ (ships in the rspock gem; linked by
   `dev up` / `dev install-deps`)
+- [testing/patch-coverage-boundary-wrappers] codecov/patch targets 100% of
+  added lines — real boundary-wrapper bodies (tests inject fakes) and
+  `--help`-only builtin blocks need one executing test each.
+  → .cursor/skills/learnings/patch-coverage-boundary-wrappers/
 
 ## process
 
diff --git a/.cursor/skills/learnings/patch-coverage-boundary-wrappers/SKILL.md b/.cursor/skills/learnings/patch-coverage-boundary-wrappers/SKILL.md
new file mode 100644
index 0000000..3f2e608
--- /dev/null
+++ b/.cursor/skills/learnings/patch-coverage-boundary-wrappers/SKILL.md
@@ -0,0 +1,37 @@
+---
+name: patch-coverage-boundary-wrappers
+description: >-
+  MUST be used when adding an injectable boundary wrapper (an
+  Executor-style shell-out, a --help-only builtin registration) or
+  diagnosing a failing codecov/patch check: the patch gate targets 100%
+  of added lines.
+---
+
+# Patch coverage includes the real boundary wrapper
+
+The codecov/patch gate targets 100% of added lines, and the lines most
+often missed are exactly the ones tests deliberately route around: the
+real body of an injectable executor (every unit test injects a fake) and
+the body of a builtin that exists mainly to surface in `dev --help`.
+Give the real wrapper one test that runs a real subprocess, and execute
+the builtin body once with a fake collaborator injected through its
+owning class's constructor (never `any_instance` — a body reachable only
+that way is missing its seam; d3mlabs/plans#37 restructures Runner's
+builtins so every one has its own).
+
+Wrong: ship `GhCloner` with tests that only ever inject
+`RecordingCloneExecutor` — the real `Executor#system` line is the
+patch's only miss and codecov/patch fails at 97.x% while every named
+test passes.
+
+Right: one test runs `Executor.new.system("echo", …)` with a real file
+standing in for `$stderr` (plus a `system("false")` false-return case),
+and one dispatch test drives the builtin body end to end with a fake
+injected at the constructor (`fake.expects(:run).with([…])`), asserting
+the argv reaches the collaborator.
+
+learned-from: dev#107 (codecov/patch reported 97.61% vs the 100% target;
+the two misses were the real gh executor body and the clone builtin's
+registration block); dev#110 (an interim constructor-seam sweep,
+superseded by the d3mlabs/plans#37 Runner layering).
+date: 2026-08-15