hook-guard: repo-level staleness keeps emitting the MANDATORY nudge after recent commits

[Luoxin]

Problem

_run_hook_guard (cli.py) decides staleness per target file: os.stat(fp).st_mtime > gmtime. When the graph predates a batch of recent commits but the file being read hasn't changed since the build, the guard still emits the strict wording:

MANDATORY: graphify-out/graph.json exists. You MUST run `graphify query` ...

Real session: graph built ≤ 2026-09-08; ~21 commits landed in browser-extension/viewer/ after that date (new modules, deleted modules). Every Read in that area got the MANDATORY nudge, so the agent paid graphify query calls on a snapshot that had never seen those modules — queries returned nothing useful and the agent had to fall back to Read/Grep anyway. The stale wording exists (_READ_NUDGE_STALE) but only fires on per-file mtime.

Proposal

When the project is a git repo, also compare repo-level freshness:

head = int(subprocess.run(["git", "log", "-1", "--format=%ct"], ...).stdout or 0)
repo_stale = head > int(gmtime)

If repo_stale, emit the softened wording — e.g.:

graphify-out/graph.json may predate recent commits (last build , latest commit ). For recently changed areas prefer Read/Grep directly; graphify update refreshes the graph.

Fails open as today if git is missing or errors.

Why

The current design already accepts "stale → soften" (#1840b); it just can't see staleness that lives between files, at the commit layer. New/deleted modules are exactly where the graph is most confidently wrong.