test(mp2): lock the layout-span guard against regressions

[Mark2Mac]

Why

_is_layout_only_span() is the guard that keeps MP2 from reporting alignment — table rules, padded columns, box edges — as context window stuffing. It has no test.

The only layout case in the suite today is:

def test_mp2_separator_not_flagged(self) -> None:
    assert not any(f.rule_id == "MP2" for f in mp_mod.analyze("=" * 80, ...))

That one never reaches the helper: a run of a single character with no whitespace is skipped by the older guard just above it. Stub _is_layout_only_span to return False and the whole suite still passes.

How much it is carrying

Measured on 4406 files from 65 real skill/plugin units, comparing the analyzer before the helper landed with main today:

MP2 findings   279 -> 4

The four survivors are from MP2's prose patterns ("exceed context window" and friends), not from (.{2,20}?)\1{20,}. So the helper accounts for 275 of 279 findings and nothing pins its behaviour.

What this adds

Three cases that are each reported when the helper is stubbed out, which is what makes them regressions rather than restatements of the guard above them:

case	with helper	helper stubbed
"- " * 40 — dash rule	ok	MP2
"| " * 40 — padded columns	ok	MP2
"│ " * 30 — box edge + padding	ok	MP2

And one for the other side, so the exemption cannot quietly widen: "- " * 200 exceeds _MAX_LAYOUT_ONLY_SPAN and is still reported.

No source change. 255 passed in the touched file, ruff check and ruff format --check clean at 0.15.19.

Context

This replaces #322, which I closed. That PR proposed its own layout guard, and the measurement above is what retired it: on today's main it removes zero findings, because _is_layout_only_span already covers the same ground. The tests are the part that turned out to be worth keeping.