fix(render): hold the shadow camera across a skipped map

[avpbynf]

What changes

WAIT. Not tried on a Mac from this machine. Do not merge until a Mac reporter signs it.

When the shadow stage opens and then gives up without drawing, the published matrices used to treat the miss as a fresh map and walk the light with the player. They now keep the camera of the frame that last drew the map. Written for the Mac reports on #161 where lighting followed the player. This does not close #161.

How it differs from the reference, and for what obstacle

Iris draws the shadow map in the same frame it is sampled. This engine draws the map at the end of the frame for the next one, so the published pair is usually the previous frame's. A skipped map made that pair advance with the player while the texels stayed put, which reads as lighting that follows you. Holding the last drawn camera is the workaround for that skip. What it costs the image on a frame that did draw is nothing; the hold only runs on a skip.

What proves it

• gradlew build green.
• The out-of-game harness, and which corpus it ran over.
• In the game: not tried on a Mac from this machine. No Windows-verified claim.

What it leaves owing

WAIT on a Mac sign-off. #161 stays open. This is the hold across a skipped map, not a proof that Apple Silicon is fast.

---

• Rebased onto dev, so the merge is a fast-forward.
• gradlew build green locally, after the rebase and not before it.
• CHANGELOG.md carries an entry under Unreleased, or this changes nothing a player sees.
• Every place that states a fact this branch changed now states the new one. A fact is cited
  in more places than it lives in: javadoc, docs/, README.md, and the lines the engine
  prints at load.