Normalize unofficial skill metadata field

`user_invocable` -> `user-invocable`

The latter form seems to be more generally supported across the harness
ecosystem, if the former has any support at all.

- [Codex mentions, but doesn't use it, in a prompt file](https://github.com/openai/codex/blob/79773e8826aeb6bca09305728297908b0d5bc6c3/codex-rs/memories/write/templates/memories/consolidation.md?plain=1#L723)
- [Claude Code](https://github.com/anthropics/claude-code/blob/2301018b1f61073c501a8e7a4813ef48c239163b/CHANGELOG.md?plain=1#L7586)
- [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness/blob/d743267388641bc76f17c45ce8b4c231aed1d32c/packages/skill/skill-filesystem/README.md#L38)
- [OpenClaw](https://github.com/openclaw/openclaw/blob/ea6022cae60c509d8c0e8029ad8030557a917617/docs/tools/creating-skills.md#L123)
- [Z.AI Synapse](https://github.com/zai-org/Synapse/blob/651d39d92ff08beff0868f991c36a27c20191726/packages/api/src/modules/skills/manifest.ts#L262)

Signed-off-by: janEbert <janpabloe@nvidia.com>

Author: janEbert

skills/pr-review/SKILL.md

@@ -2,7 +2,7 @@
 name: pr-review
 description: Repository review rubric for the formal /review command.
 disable-model-invocation: true
-user_invocable: false
+user-invocable: false
 ---
 
 # PR review

tests/unit_tests/export/test_review_policy.py

@@ -58,7 +58,7 @@ def test_formal_review_rubric_is_inert_and_uses_formal_submission() -> None:
     metadata = yaml.safe_load(rubric.split("---", 2)[1])
     assert metadata["name"] == "pr-review"
     assert metadata["disable-model-invocation"] is True
-    assert metadata["user_invocable"] is False
+    assert metadata["user-invocable"] is False
     assert "mode=light" in rubric
     assert "mode=strict" in rubric
     assert "Do not run GitHub commands or" in rubric