test(gate-3): pin the #434 exemption against the #422 mask — the two reads must stay split

[rubenvdlinde]

test(gate-3): pin the #434 exemption against the #422 mask — the two reads must stay split

#445 (#422) and #434 (#339) both changed the SAME arm of gate-3, three lines
apart, so they merged cleanly and the interaction never appeared in a diff.
Both are now on main and the behaviour there is CORRECT — but nothing asserts
it, and this is a shape that breaks silently.

The arm reads two texts in one coordinate system, and only php_mask being
line- and offset-preserving makes that legal:

the body questions  -> the MASK      _sig / _sig_region / _body / _count
                                     ("is $uid referenced in CODE?")
the exemption       -> the ORIGINAL  --file "$f" --line "${_line_no}"
                                     (#434's `@param … (unused)` marker,
                                      which lives in a COMMENT BY DESIGN)

Repoint the exemption at the mask — a one-word edit, and the obvious
"consistency" cleanup for anyone who notices the arm reading two different
files — and the marker vanishes into blanks. The exemption silently stops
working and procest's three GuardEvaluatorInterface implementors go red WITH NO
CLOSING ACTION AVAILABLE: deleting the parameter breaks the interface and every
call site, referencing it pointlessly is dead code written to satisfy a stub
detector. #434's own note says neither is a fix, which is why it added the
exemption in the first place.

That is gate-17's @spec exclude trap, which #444 hit for real in the sibling
gate of the same sweep: masking the body slid the reported line into the
blanked docblock, _method_docblock walked from the wrong place, and a
reason-bearing exclusion silently stopped exempting. Caught there by an
anti-widening arm. This is the same arm for gate-3.

CONTROL 6b — a supertype-declared authorize(string $uid, …) whose docblock
reads @param string $uid Current user UID (unused by this implementor).
must still be EXEMPT with the mask in place.

MUTATION-CHECKED, so it is a coupling and not a comment: with --file pointed
at ${_stub_code} the arm reports FAIL — 1 (9 passed, 1 failed); restored,
10 passed, 0 failed. Measured against main at 43ecb0c.

No production code changes — one arm, in a suite tests/run-helper-suites.sh
already discovers.

Refs #422, #339, #445, #434.