fix(cli): re-exec the hash seed pin as a module, not by replaying argv[0]

[ayushcodes10]

Fixes #3779 — a regression in my own #3743 (issue #3641), reported within a day of the 0.9.66 release shipping. Sorry about that.

_pin_hash_seed_if_needed re-execs like this:

os.execvpe(sys.executable, [sys.executable, *sys.argv], {**os.environ, "PYTHONHASHSEED": "0"})

This replays sys.argv[0] as the script path for the interpreter to run. That works for a POSIX console-script wrapper or a python -m graphify invocation — both real, runnable Python content — but not for a uv/pip/pipx console-script launcher on Windows, which installs as a native .exe stub with no such content. The interpreter has nothing to open, and every command this function touches (update/extract/cluster-only/label) failed outright with can't open file the moment it tried to pin the seed.

Fix (the issue's own suggestion, which is correct and minimal): re-exec via -m graphify instead, which never depends on argv[0] being anything runnable at all:

os.execvpe(sys.executable, [sys.executable, "-m", "graphify", *sys.argv[1:]], {**os.environ, "PYTHONHASHSEED": "0"})

This also now matches the exact invocation shape the existing end-to-end test for this function (test_update_still_runs_end_to_end_with_hashseed_unset) was already exercising.

Verified the failure mode empirically before touching anything — simulated a non-Python launcher stub path as argv[0] and confirmed it broke pre-fix (can't open file) and is fixed post-fix. Updated the existing probe assertions to the new argv shape and added a dedicated regression test that drives the probe with a launcher stub path that isn't even a real file, confirming the fix no longer depends on it being runnable — the exact shape of the Windows uv-launcher scenario. Full suite is green aside from 16 pre-existing failures in the newly-added Erlang/R/Solidity/VB.NET extractor tests, caused by their optional grammar packages not being installed in my dev environment — unrelated to this change (confirmed via python -c "import tree_sitter_r" etc. all failing with ModuleNotFoundError, and confirmed these packages weren't needed by any branch cut before the 0.9.66 release landed).

🤖 Generated with Claude Code

https://claude.ai/code/session_017qfdzgbA5KedGEjD1AayNh

[safishamsi]

Shipped in v0.9.67 (on PyPI). Cherry-picked with authorship preserved. Thanks @ayushcodes10! This fixes the Windows regression the 0.9.66 hash-seed pin introduced — the -m graphify re-exec is exactly right.