Two findings + three suggestions from the latest Rust PR Reviewer pass:

1. Stage-3 typed-deserialize errors after codemods reused the same
   "Failed to parse YAML front matter" context as the Stage-1 raw
   parse, so a codemod that produced an invalid shape was
   indistinguishable from a user typo. Switch to with_context that
   surfaces the applied codemod ids when codemods ran:
   "Failed to parse YAML front matter after applying codemods (id1, id2);
    a codemod likely produced an invalid shape".

2. rename_key with ConflictPolicy::PreferNew returns Ok(true) when
   it drops a stale `old` key even though `new` is unchanged. That
   triggers a source-rewrite warning saying "we migrated" when
   really we "cleaned up a remnant". The behavior is correct (the
   mapping did mutate); the doc was misleading. Rewrote the
   rename_key doc comment to enumerate exactly when Ok(true) fires
   and call out the PreferNew/cleanup case.

3. Added `// TODO(codemods): remove when the first real codemod is
   registered.` markers next to every #[allow(dead_code)] /
   #[allow(unused_imports)] in src/compile/codemods/{mod,helpers}.rs
   so search-based cleanup is trivial when the registry stops being
   empty.

4. tests/codemod_tests.rs leaked temp dirs on assertion failure
   because `let _ = fs::remove_dir_all(&dir)` only ran on the happy
   path. Refactored to return tempfile::TempDir from the helpers
   and rely on RAII cleanup, matching the pattern already used by
   src/compile/codemod_integration_test.rs.

Skipped: the empty-mapping concern in reconstruct_source (a
mapping reduced to empty wouldn't deserialize as FrontMatter
anyway because of required name/description fields, so the
existing typed-deserialize error path catches it; #1 makes that
error informative).

All 1349 tests pass.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>