The two pieces of logic the previous commit added had no tests, and both are
the kind that fail silently in opposite directions.
`blend_key_of` is now ONE definition shared by the regen test and the blend
itself, rather than the same tuple built at two call sites. That drift is the
actual bug shape: a key that reports "changed" while the drawn frame is
identical re-encodes for nothing, and a key that reports "unchanged" while the
pointer moved freezes it on screen. Both directions are asserted — a hidden
pointer keys identically wherever it moves, and every visible change (position,
shape serial, and the visible→hidden transition that must strip the pointer
from the frame) moves the key.
`next_blend_backoff` is extracted for the same reason `mono_planes_to_rgba`
was: the arithmetic a bug hides in does not need a live D3D11 device around it
to be checked. The test walks the escalation well past its ceiling and asserts
it PARKS there — an unbounded doubling would mean a device that comes back
after a long stall never gets picked up.
Verified: `scripts/xcheck.sh windows` green; `cargo fmt --all` clean. The tests
themselves compile and run only on Windows (the module is `cfg(windows)`), so
they are pending a run on a box.
enricobuehler
7f6d1622ee