[Bug] Generated CLAUDE.md stamps live HEAD instead of the commit the index was built against

## Summary The `Last indexed` stamp in a generated `CLAUDE.md` reports whatever `git HEAD` was when the file was *written*, not the commit the index it describes was built from. The two routinely disagree, nothing self-heals the drift, and every fact under the stamp inherits it. ## Mechanism `EditorFileDataFetcher.fetch` shells out at `generation/editor_files/fetcher.py:73`: ```python indexed_commit=_get_head_short_sha(self._repo_path), ``` ```python def _get_head_short_sha(repo_path: Path) -> str: """Return the short SHA of HEAD, or empty string if not a git repo.""" try: result = subprocess.run( ["git", "rev-parse", "--short", "HEAD"], ``` The value it wants already exists on the row it just loaded. `repo` is in scope at line 65, and `Repository.head_commit` is documented for exactly this (`persistence/crud/repository.py:52-58`): > `head_commit` records the git commit the index was built against - the value > the MCP `_meta` freshness check compares to the live HEAD. ## Why the drift is routine, not a corner case A rebase fires the post-commit hook once per replayed commit. Each firing rewrites `CLAUDE.md` with the then-current HEAD while the index underneath is unchanged, so the stamp walks forward on its own. There is no path that pulls it back. This is also why the mismatch is easy to misread: the same commit before and after a rebase has two SHAs, so the stamp and the index can look divergent when only the stamp moved. ## Blast radius Everything below the stamp is a static snapshot taken at the same moment and carries the same drift, including the code-health "worst performer" line. The stamp is the only thing claiming to date them, so if it is wrong they are all silently misdated. ## Fix Read the stored commit, falling back to the shell call only when the column is empty (pre-backfill indexes): ```python indexed_commit=(repo.head_commit or "")[:7] or _get_head_short_sha(self._repo_path), ``` Worth confirming the short-SHA length matches what `git rev-parse --short` produces on this repo, so the rendered stamp does not change width. ## Tests The existing assertion only checks the `YYYY-MM-DD` shape of `indexed_at`, so it survives untouched. A new test should set `head_commit` to a known value, leave the working tree on a different commit, and assert the stamp reports the stored one.