Keep what a program draws, instead of making it a sprite

A drawing lasted exactly one frame. The verbs are immediate, they went to the
back buffer, SDL double-buffers and the frontend never clears -- so the only way
to keep a picture was to capture it with `SSHAPE` and install it as a sprite,
which is what `examples/breakout/sprites/breakout.bas` spends two of its eight
sprites doing. That was TODO.md section 9 item 9.

The drawing verbs now render into a layer texture the frame composites under the
text and the sprites. Draw once; it is there on every frame after.

**Bracketed around the step phase, not around each verb.** One pair of
`SDL_SetRenderTarget` calls a frame instead of one per `DRAW`, and it is also
what makes `SSHAPE` read back what the program has just drawn rather than
whatever the last frame left.

**The layer is transparent where nothing was drawn.** It covers the whole window
and composites underneath, so an opaque one would black out the frame the moment
a program issued a single `DRAW`. And a fresh SDL target texture's contents are
undefined, so it is cleared on creation -- skipping that puts uninitialised
memory under the first frame's text and looks like a driver bug rather than a
missing memset.

**The line editor forced a wrinkle worth naming.** `akbasic_frontend_akgl_pump()`
is called from two places with different answers to "is a render target current":
the frame loop calls it between steps, and the sink's editor calls it from
*inside* a step, borrowing a frame while it waits for a typed line. SDL refuses
to present while a target is current, so the pump ends the layer, presents, and
puts it back only if it was the one that ended it. `akgl_frontend` caught this --
it drives a REPL session, and it failed with "You can't present on a render
target" the first time the brackets went in.

This does not make a drawing *visible* on its own. The text layer still repaints
every row it owns, opaque, every frame, and by default it owns the whole window;
`WINDOW` shrinks it and that half was already fixed. The two together are what a
picture needed, and the tests assert both -- a pixel still there a frame later
with nothing redrawn, and a pixel below a shrunk text area surviving the text
repaint. The second assertion wipes to a non-black colour first, because against
black it could not tell a transparent layer from an opaque one.

The tests found two of their own bugs on the way: `stop_runtime()` was not
tearing the graphics backend down, so re-initialising it dropped a live texture
on the floor; and a first draft called `begin()` before `start_runtime()`, which
re-inits the backend, so the assertion read back off an orphaned render target
and passed while proving nothing.

Chapters 6 and 13 stop saying a drawing has to be redrawn every frame, because it
does not. The batch-boundary tear stays documented -- it bites an `SSHAPE`
capture, which matters much less now that capturing is not the only way to keep a
picture.

Both games still run clean. 111 with akgl, 110 without.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EwxGB6TdoVvZ11KQQME9cL

docs/13-differences.md

@@ -142,19 +142,21 @@ interpreter's error code, which bears no relation to a Commodore error number. P
 - **`SSHAPE` puts a handle in the string, not the pixels.** You can pass it to `GSHAPE`
   and `SPRSAV`; you cannot save it or take its `LEN`.
 - **`WIDTH` is emulated** by drawing parallel passes.
-- **Drawing does not persist across frames.** The verbs draw straight to the renderer,
-  so anything drawn is overwritten on the next frame unless the program redraws it.
-  This is a defect, not a design.
-- **And a redraw has to *fit*.** The host runs a fixed number of source lines and then
-  presents the frame, and presenting discards the drawing buffer — so a sequence of
-  drawing verbs longer than one batch comes back **half drawn**, with an `SSHAPE` at the
-  end capturing only what was issued since the present. "Redraw it every frame" is
-  therefore not sufficient advice on its own. Measured against the standalone frontend's
-  256 lines a batch: after synchronising to a jiffy edge, 220 lines of drawing survive a
-  capture and 250 do not. The only way a program can see the boundary is to watch `TI#`,
-  which is refreshed once per batch — so the step on which it changes is the first step
-  of one. [Chapter 18](18-tutorial-breakout-artwork.md#step-5-find-the-frame-boundary)
-  builds a pacing routine out of exactly that.
+- **Drawing persists, and the text layer covers it.** A drawing goes into a layer that
+  survives the frame, so a program draws its picture once and it stays — no redrawing, no
+  capturing it into a sprite. What is still true is that the text layer repaints every row
+  it owns, opaque, every frame, and by default it owns the whole window. `WINDOW` shrinks
+  it and hands the rest over. The two together are what makes a picture usable:
+  `WINDOW 0, 0, 39, 1` keeps a one-row status line and gives the drawing verbs everything
+  below it.
+- **A drawing still has to fit in one batch.** The host runs a fixed number of source lines
+  and then presents, and an `SSHAPE` capture spanning that boundary comes back **half
+  drawn** — the part issued since the present, over whatever was there before. Measured
+  against the standalone frontend's 256 lines a batch: after synchronising to a jiffy edge,
+  220 lines of drawing survive a capture and 250 do not. This bites a *capture*, not the
+  drawing itself, so it matters far less than it did when capturing was the only way to
+  keep a picture. The only way a program can see the boundary is to watch `TI#`, which is
+  refreshed once per batch.
 - **`CHAR` ignores its colour argument** and needs a text device with a cursor.
 
 ## Sound