Grade the overlay water skip and the mark painters in Rust (#227)

[wormeyman]

Second of the cheap-tier delete-with-loss ports, and it closes the same claim in
four deleted specs at once. renderCliffs, renderRocks, renderEnemies and
renderResources each carried a "never paints over water" test. Rust has the
predicate - is_nauvis_water, derived from the palette so it cannot drift - and
every overlay passes it, but the only test of it was the tree one.

paint_mark and paint_mark_skipping had no test at all, though every
placement overlay on both planets goes through them.

Two plants shaped the water test, and both were mine

The obvious form of this test is worth writing down as a dead end, because it
looks convincing:

attempt	why it does not grade anything
render each overlay, assert no water pixel moved	Nearly vacuous. Rocks are excluded from water at PLACEMENT time, so in an ordinary window no mark ever gets the chance to spill onto water. Deleting the rock overlay's skip entirely left it green.
build an all-water base from a colour sampled out of a render	Vacuous the other way. The window handed back deepwater, so a predicate that stopped recognising SHALLOW water also passed.

Both were planted and observed, not reasoned about. The shipped test names both
tiles from the palette rather than sampling one, and under the shallow-water
plant it now fails while all 34 other tests pass.

It also paints the same mark on a land base and requires all 25 pixels, so a
painter that does nothing cannot satisfy it.

The rest

• every_nauvis_overlay_paints_and_leaves_water_alone is kept, and is
  deliberately described as the weaker check it is. What it does pin is that
  each overlay paints SOMETHING in a real window. The enemy arm runs at a
  different origin because enemy bases are excluded from the starting area by
  design - the anti-vacuity guard caught that rather than passing silently, and
  that is how the window was chosen.
• Three paint_mark tests: the (2r+1) square, clipping at the edge instead of
  wrapping, and consulting the skip predicate per pixel rather than per mark.

Claude-Session: https://claude.ai/code/session_01UVcbv1pAhPUoCC6aBwZUtg

Co-authored-by: Claude Opus 5 (1M context) noreply@anthropic.com