fix(PS-140): scope the cross-package gate's skip to the ROOT so a rename can fail it

[ywatanabe1989]

The defect: a gate that cannot fail

tests/integration/test_cross_package_imports.py promised, in its own docstring:

Renames or moves in peer standalones (or in the umbrella's deep-import
surface) surface as a hard failure here — not as a silent
ModuleNotFoundError when an end user runs the CLI.

It did the opposite. Its helper caught ModuleNotFoundError raised by importing
the full dotted path and skipped on it:

try:
    return importlib.import_module(module_name)
except ModuleNotFoundError as exc:
    pytest.skip(f"{module_name}: not importable in this environment ...")

A renamed or moved submodule — scitex_io._load_cache →
scitex_io._loading._load_cache, the incident this file was written for —
raises exactly that exception. So the one failure the gate exists to catch was
swallowed and the run reported green. An absent peer and a rename inside an
installed peer raise the same exception type, and the old code could not tell
them apart, so it treated every rename as an absence.

The fix: skip on the ROOT, hard-import the FULL path

The two cases are separable by scope, not by exception type:

• the ROOT distribution is not installed → a legitimately absent peer (optional
  extra, marker-gated dependency, lean install). Skip.
• the root IS installed and a SUBMODULE of it is missing → a rename or a move.
  Fail. That is the gate.

root = module_name.split(".")[0]
pytest.importorskip(root)              # skip on the ROOT
return importlib.import_module(module_name)   # hard-import the FULL path

Simply dropping the skip would be the opposite pole of the same bug: a lean
install where a peer is legitimately absent would fail — a gate that cannot
PASS in place of one that cannot FAIL.

The umbrella needs one further absence recognised, and it does not widen the
hole: scitex.<x> forwards to a separate scitex_<x> distribution, so
import scitex.io fails naming scitex_io while the root scitex is present.
That is skipped only when the missing name is a different top-level
distribution with no importable spec at all. A missing submodule of an
installed distribution never matches it, which is what keeps a rename loud.

Proof the gate is alive — both control arms, one run

A bogus entry "scitex_io._this_module_does_not_exist" was injected into the
old helper and the new one and both were run together:

tests/integration/test_ps140_control_old.py s      [ 50%]
tests/integration/test_ps140_control_new.py F      [100%]

SKIPPED [1] test_ps140_control_old.py:247:
  scitex_io._this_module_does_not_exist: not importable in this environment
  (auto-gen snapshot drift / optional peer absent):
  No module named 'scitex_io._this_module_does_not_exist'

FAILED test_ps140_control_new.py::test_cross_package_import[scitex_io._this_module_does_not_exist]
E   ModuleNotFoundError: No module named 'scitex_io._this_module_does_not_exist'

1 failed, 1 skipped

Old code: SKIPPED (green). New code: FAILED. The control files were then
deleted and the gate re-run:

110 passed, 49 skipped, 4 xfailed

Baseline on the same tree before the change: 110 passed, 84 skipped. Same 110
passing — the 35 skips that disappeared were the gate's blind spot.

Two consequences of the gate coming alive

31 stale snapshot rows, pruned. None appears anywhere in src/:
src/scitex/ has no bridge, _dev/ holds only __init__.py, and gen is a
module (gen.py) so scitex.gen._x can never be a submodule. They were rows
that could only ever skip. (For scale: the SSoT collector
_collect_cross_package_imports finds 18 cross-package imports in the
current source against 194 declared — the umbrella's snapshot has never been
regenerated, because PS-140 exempts the umbrella and so nothing ever complained.)

4 entries are a real, live defect — recorded as strict xfails, not skipped
away. src/scitex imports scitex_dev.{docs,search,skills,types} from
cli/docs.py, cli/skills.py, cli/mcp.py, _mcp/__init__.py and
_mcp/_umbrella_tools.py, and the pinned scitex-dev==0.28.0 ships none of
them:

$ git -C ~/proj/scitex-dev ls-tree --name-only v0.28.0 src/scitex_dev/
... src/scitex_dev/_docs   src/scitex_dev/_skills ...      (no docs/skills/search/types)

So this is not a local version skew — it is broken against the version this
package pins. _show_list in cli/docs.py imports unguarded, so it raises at
runtime. Where the symbols moved to is not unambiguous — scitex_dev._docs
exports none of get_docs / build_docs / search_docs — so the import sites
are left to a separate change rather than guessed at. strict=True means this
gate fails the day they are corrected, until the record is deleted; it cannot
rot into a permanent exemption.

This is the defect class PS-140 was written for, and it was invisible for as
long as the gate could not fail.

Audit note

scitex-dev ecosystem audit-project scitex reports no PS-140 both before
and after this change: check_ps140_integration_gate returns early on
_is_umbrella(repo), so the umbrella's gate is never checked by the rule. The
defect was real regardless, which is why it is fixed here. No PS-224 finding is
reported for this repo either.