[jeany55]

Texts with numbers in them, in a map of our own (no eudplib, no Python): step 2 of the way off eudplib in the plan.

Not a table of digits after all. The plan proposed a table of 1,000 three-digit cells. Working it out, the table only
saves the last step — a number still has to be cut into groups of three, a division by 1,000 — and the harder part is
where the digits go: a number's width changes from print to print, so everything after it moves a byte, and the game
writes whole dwords. What was built instead (compiler/backend/emit.ts, layout.ts, graph.ts, scripts/build-own.mts):

• A template for each print, laid out when the map is built: its text, and for each number a field of 12 bytes on
  a dword's start, the room before the number 0x0D — a carriage return, drawn as nothing. eudplib pads its strings with
  it, and its own prints carry it around every number, so it is proven in Remastered already. Nothing in a text ever
  moves. A byte from a value (a colour from JSX) is one byte of the template.
• One routine, shared by every print, writes a number into a field: how many digits by comparisons with powers of
  ten (four at most), then four triggers a digit — 8, 4, 2 and 1 times its power of ten, taken from the number and
  added to the digit's byte. eudplib's build divides by ten for each digit, a bit at a time.
• A print copies the field's three dwords into its template (routed) and shows it with Display Text. The template
  is a string of the map's — a placeholder it is written into, its offset moved to the first dword's start — among the
  map's own strings, since 1.16.1's table names only their first 64 KB.

probes/own-print.ts, lifted game, Remastered's weights:

	Own	eudplib
One text, one number	0.001 ms	0.013 ms
A hundred texts of two numbers in a frame	0.133 ms	2.411 ms

Checked: the random programs print now (words, Korean, numbers signed and u32, bytes from values): about 9,600
texts compared byte for byte with what JavaScript makes of them, as graphs and from the laid-out bytes, in the programs
with functions that call themselves too; both builds of the probe print the same twenty lines at the same frames once
the 0x0D are taken out of both; game-check: none of 178 accesses ends the game; the calls, recursion and own-loader
probes still all ok. Played in Remastered: every line word for word as the lifted game's, the Korean drawn, nothing
where the padding is — in Remastered's graphics and, after F5, Classic's.

The suite: 1194 pass; tests/sound.test.ts could not load in this worktree — its shared node_modules lacks
wasm-media-encoders, which the music work added — nothing of this change.

Not yet: a fixed number printed (3.25), a player's name or colour, a text of the program's, position: "center".

🤖 Generated with Claude Code

https://claude.ai/code/session_01T1CM9pLuTuAn3mgTEkjgVg