Add the AKLABS megademo example

A six-scene demoscene production that pushes the interpreter to its
edges on purpose: an LCG where there is no RND, redrawn strokes where
the palette cannot be written, the sprite bank as the scrolltext, a
supernova and black hole out of SOLID/COLLISION/RSPPOS, wireframe 3D
from two rotations and a perspective divide, and a single-queue
beeper-style soundtrack with a noise drum break.

The same file runs in the standalone frontend and in the clockless
screenshot host, which gets each scene as a still. Three interpreter
lessons learned the hard way are recorded in the example's README and
TODO.md section 4: a scalar first assigned in a GOSUB dies with its
scope, predeclaring scratch runs the 128-slot variable pool dry, and
a TI# busy-wait starves the PLAY queue of its one release per frame.

Co-Authored-By: Claude Code (Fable 5) <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EwxGB6TdoVvZ11KQQME9cL
Co-Authored-By: Andrew Kesterson <andrew@aklabs.net>

TODO.md

@@ -727,6 +727,73 @@ Also on the queue, from the reference's own defect list:
   so moving one operand out of the way only shifted the collision along. Covered by
   `tests/language/arrays_in_parameter_lists.bas` and `tests/runtime_evaluate.c`.
 
+### What a demoscene program wanted and could not have
+
+`examples/megademo/megademo.bas` was written to push the interpreter to its edges, and these
+are the edges it hit. None of them blocked the demo — every one has a workaround, and the
+program carries all six — but each workaround is a routine every future game will carry too,
+and that is the argument for the verb. Ordered by how much BASIC each one would delete.
+
+- **`RND` is still missing.** Chapter 17 teaches the LCG workaround and the demo's `RANDOM`
+  label carries it: nine tokens, two convention globals (`RMAX#` in, `RND#` out), and a seed
+  the caller must remember to take from `TI#`. That is the right *teaching* example and the
+  wrong thing to make every program transcribe. Closing it is the standard function shape —
+  dispatch-table row, exec handler, tests, a Chapter 12 row — and Chapter 17's section becomes
+  a history lesson instead of a requirement. `ASC` is in the same boat with a worse workaround
+  (`INSTR` into a 41-character alphabet string); `INT` and `SQR` are not, because integer
+  assignment and `^ 0.5` already are those functions.
+- **The palette cannot be written.** `PALETTE[17]` in `src/graphics_tables.c` is `static
+  const`, and `COLOR` can only choose among its sixteen entries. Palette rotation was the
+  cheapest animation of the whole era — one register write cycled every pixel of a colour —
+  and here the same effect is a full redraw: the demo's title "colour cycle" re-strokes the
+  logo every frame (`DRAWTEXT` with `CM# = 1`), and its raster bars erase and redraw five
+  ramp lines apiece per frame. Closing it means palette state moves from a static table to
+  the graphics backend, plus a verb form to set an index (`COLOR` grew nothing new on a C128,
+  so this would be ours — Chapter 13 material), plus the docs figures that assume Pepto's
+  values stay put.
+- **`GSHAPE` stamps in one mode.** BASIC 7.0's takes a mode argument — replace, OR, AND,
+  XOR — and XOR is the one that mattered: stamp to draw, stamp again to erase, no damage to
+  what was underneath. That is software sprites beyond the eight real ones, and it is the
+  difference between the demo's raster bars chewing holes in the starfield (they erase in
+  black) and leaving it intact. Needs a blend argument through `SSHAPE`/`GSHAPE`'s render
+  path in `src/graphics_akgl.c`, and libakgl may need to offer the blend mode first — check
+  before filing there.
+- **Nothing reports whether `PLAY` has drained.** The queue depth exists in
+  `akbasic_AudioState` (`count - head`); no function reads it. The demo's `TUNE` label
+  requeues each batch on a wall-clock guess — it stamps `MT#` with the batch's computed
+  length plus a breath, and the sums drift against `TEMPO` and had better stay under the
+  queue's 128-note cap when a requeue lands early. An `RPLAY(v)` returning notes pending —
+  same shape as `RSPPOS` — deletes the arithmetic. It also fixes the quiet failure Chapter 7
+  warns about, where a program `QUIT`s before its music finishes and never knows.
+- **`PLAY` is monophonic: one queue, one clock, one instrument per batch.** An earlier
+  version of this entry claimed there was no noise waveform, and that was wrong —
+  `ENVELOPE`'s sixth argument selects triangle, sawtooth, square or noise
+  (`src/runtime_audio.c`), which Chapter 7 does not document and should. The real gap is in
+  `akbasic_play_service()` (`src/play.c`): every voice's notes go into **one** queue with one
+  release clock, and the envelope applied when a note sounds is whatever the last `T`
+  *parsed* selected — so a program that writes three `V` tracks gets them end to end, on one
+  instrument, and `M` is a no-op precisely because there is nothing to synchronise. The
+  backend already has three independent voices and `tone()` takes a voice number; what is
+  wanted is a queue head and clock per voice, and the envelope resolved at parse time into
+  the note rather than read from shared state at release. The demo works around it the way
+  1985 did on one channel: an interleaved kick–bass–arpeggio line (the `TUNE` label), which
+  is authentic and should not be mandatory.
+- **The only vsync is spinning on `TI#`, and the spin is worse than it looks.** Chapter 13
+  documents the polling; the demo's first cut had the loop — read `TI#`, branch to yourself
+  until it moves — and it did not just waste CPU, it **starved the `PLAY` queue dry**. The
+  chain is measurable: the host calls `settime()` once per `runtime_run()` batch, and
+  `akbasic_play_service()` can therefore release at most one note per batch (after a release,
+  `nextms` sits ahead of the frozen clock). A batch of 256 *held* `SLEEP` steps is nearly
+  free, but a batch of 256 *executed* spin statements is slow enough to drop the release rate
+  below a sixteenth-note tune's fifteen notes a second — and the queue never reclaims slots
+  until it drains completely, so the backlog compounded across scenes and overflowed the
+  128-note queue with `PLAY queue is full`. Reproduced both ways in isolation: the same tune
+  requeued eight times over a `SLEEP` loop survives, over a `TI#` spin it dies on the third
+  batch. The demo's `VSYNC` label now sleeps a frame instead. A verb that parks the program
+  until the next batch boundary (`SLEEP 0` is an available spelling) is the same
+  hold-without-blocking shape `SLEEP` and `GETKEY` already have in `src/runtime_console.c`,
+  and it would make the honest spelling also the safe one.
+
 ---
 
 ## 5. Deliberate deviations from the reference