perf(vapor): store the GBA font as 1bpp and expand it at upload

[lfkdsk]

What

The GBA font table vp_font_tiles was emitted as 4bpp tiles (3040 B), which was 32.5% of the Todo ROM and 47.4% of the six-button ROM — larger than every code symbol combined. This change emits the font as 1bpp (760 B) and expands it into the 4bpp charblock inside upload_font at boot.

Measured

ROM	before	after	delta
todo.gba	9356 B	7160 B	−2196 B (−23.5%)
playdate-six-button.gba	6420 B	4224 B	−2196 B (−34.2%)

Stable over 5 builds per side. The delta reconciles in the ELF symbol table: −2280 B of table, +84 B of expansion code. State RAM, overlay and reactive figures are unchanged.

Correctness

• All 3040 bytes that reach the GBA font charblock are byte-identical to the old table, verified by booting the ROM in headless libmgba and dumping charblock 0 (SHA-256 955a983b…), not by re-implementing the expansion.
• The existing GBA parity rig cannot see glyph pixels (it probes the screenblock at 0x6004000; glyphs live at 0x6000020), so this adds vapor/tests/gba-font.test.ts, which reads VRAM back from a booted ROM. Three runtime mutations (halfword order, ink/paper swap, row overrun) each make it fail while parity stays green.
• GB, NES, ESP32 and Playdate generated C is byte-identical; GB/NES final ROMs are byte-identical. ESP32/Playdate final packages were not built locally (no SDK) and their embedded build IDs hash the changed shared vapor.h, so only generated-C identity is claimed for those two.
• bun test vapor/tests/: 74 pass / 0 fail (was 71), including GBA, GB, NES and oracle parity.

Provenance

Produced by an autonomous fleet run: a Scout task measured the size baseline and proposed this candidate with a hand-patched prototype; a Builder implemented it; three independent cross-family reviews (different model families) rebuilt the ROMs, re-dumped VRAM and re-ran the mutations. The Scout report lives on the fork branch lfkdsk/pocketjs:fleet/scout-artifact-size; this PR carries only the implementation and its tests.

🤖 Generated with Claude Code

[lfkdsk]

Rebased onto current main (061c6be) and dropped the fork-only base commits, so the PR now contains only its own commit(s). bun test vapor/tests/ on the rebased branch matches main in this environment: everything passes except the NES parity build, which fails identically on unmodified main here (local ld65/cc65 library mismatch).