fix(compress): match fences nested in list items

FENCE_OPEN_REGEX bounded indent at 0-3, correct only for top-level
fences. Inside a list item indent counts from the item's content
column, so a fence under "- " sits at 4 absolute and was missed.

Two effects. Its body backticks leaked into inline-code pairing, so
whole paragraphs came back as lost inline code and validation failed
unfixably. And extract_code_blocks never saw the block, so
validate_code_blocks could not tell it had been rewritten - the
copied-EXACTLY guarantee did not hold for any fence under a bullet.

Refs #619

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: rammicz

skills/caveman-compress/scripts/validate.py

@@ -4,7 +4,8 @@
 from pathlib import Path
 
 URL_REGEX = re.compile(r"https?://[^\s)]+")
-FENCE_OPEN_REGEX = re.compile(r"^(\s{0,3})(`{3,}|~{3,})(.*)$")
+# A list item's fence indents from its content column, so one under "- " is at 4.
+FENCE_OPEN_REGEX = re.compile(r"^([ \t]*)(`{3,}|~{3,})(.*)$")
 HEADING_REGEX = re.compile(r"^(#{1,6})\s+(.*)", re.MULTILINE)
 BULLET_REGEX = re.compile(r"^\s*[-*+]\s+", re.MULTILINE)
 

tests/test_validate_code_blocks.py

@@ -0,0 +1,56 @@
+import sys
+import unittest
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(REPO_ROOT / "skills" / "caveman-compress"))
+
+from scripts.validate import (  # noqa: E402
+    ValidationResult,
+    extract_code_blocks,
+    validate_code_blocks,
+)
+
+MANGLED = "- step:\n{i}```\n{i}rm -rf /  # rewritten\n{i}```\n"
+INTACT = "- step:\n{i}```\n{i}rm -rf /important\n{i}```\n"
+
+
+class TestExtractCodeBlocks(unittest.TestCase):
+    def test_top_level_fence(self):
+        self.assertEqual(len(extract_code_blocks("```\ncode\n```\n")), 1)
+
+    def test_fence_indented_within_list_item(self):
+        self.assertEqual(len(extract_code_blocks(INTACT.format(i="    "))), 1)
+
+    def test_fence_indented_within_nested_list_item(self):
+        text = "- a:\n  - b:\n        ```\n        code\n        ```\n"
+        self.assertEqual(len(extract_code_blocks(text)), 1)
+
+    def test_tilde_fence_indented_within_list_item(self):
+        text = "- step:\n    ~~~\n    code\n    ~~~\n"
+        self.assertEqual(len(extract_code_blocks(text)), 1)
+
+
+class TestValidateCodeBlocks(unittest.TestCase):
+    def test_rewritten_block_is_rejected_at_every_indent(self):
+        # A block the extractor cannot see is one the compressor rewrites
+        # unchecked, breaking SKILL.md's copied-EXACTLY guarantee.
+        for indent in ("", "  ", "    ", "        "):
+            with self.subTest(indent=len(indent)):
+                result = ValidationResult()
+                validate_code_blocks(
+                    INTACT.format(i=indent), MANGLED.format(i=indent), result
+                )
+                self.assertFalse(result.is_valid)
+
+    def test_untouched_block_passes(self):
+        for indent in ("", "  ", "    ", "        "):
+            with self.subTest(indent=len(indent)):
+                result = ValidationResult()
+                text = INTACT.format(i=indent)
+                validate_code_blocks(text, text, result)
+                self.assertTrue(result.is_valid)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_validate_inline.py

@@ -52,6 +52,17 @@ def test_indented_fence_backtick_not_leaked_as_inline(self):
         result = extract_inline_codes(text)
         self.assertEqual(result, ["inline"])
 
+    def test_list_nested_fence_backtick_not_leaked_as_inline(self):
+        # Fence under "- " sits at 4, past the old 0-3 bound.
+        text = "- step:\n    ```\n    `weird`\n    ```\nReal `inline` span here."
+        result = extract_inline_codes(text)
+        self.assertEqual(result, ["inline"])
+
+    def test_deeply_nested_fence_backtick_not_leaked_as_inline(self):
+        text = "- a:\n  - b:\n        ```\n        `weird`\n        ```\nReal `inline` span."
+        result = extract_inline_codes(text)
+        self.assertEqual(result, ["inline"])
+
 
 class TestValidateInlineCodes(unittest.TestCase):
     def test_match(self):