signed compares, metatiles + collision, register-allocator pass

[imjasonh]

Works through the top three items in the priority ranking at the bottom of docs/future-work.md.

1. Signedness on Cmp16/Cmp ops (§A follow-up)

Ordering compares (<, <=, >, >=) on signed integers now use the canonical 6502 CMP / SBC / BVC / EOR #$80 overflow-correction idiom so the N flag reflects the true sign of the difference. The lowerer tracks signedness on each IR temp (analogous to the existing wide_hi map) and threads it onto a new Signedness field on CmpLt/CmpGt/CmpLtEq/CmpGtEq and their 16-bit variants.

widen() also sign-extends when the source temp is signed, via a new IrOp::SignExtend, so var w: i16 = some_i8_neg round-trips negative values instead of zero-extending to $00F6.

Casts to u8/u16 strip the signed flag so expr as u16 < other_u16 stays on the unsigned path.

See examples/signed_compare.ne — four pip sprites gate on signed comparisons; three light (signed-correct) and one stays dark (would only light if the lowering regressed to unsigned).

2. Metatiles + collision (§H)

metatileset Name { metatiles: [{ id, tiles, collide }, ...] } and room Name { metatileset: M, layout: [...] } ship as a cohesive feature. Each metatile bundles 4 CHR tile indices (TL/TR/BL/BR) plus a collide flag; rooms lay them out as a 16×15 grid that the compiler expands at compile time into three PRG blobs:

• __room_tiles_<name> — 960-byte 32×30 nametable
• __room_attrs_<name> — 64-byte attribute table
• __room_col_<name> — 240-byte collision bitmap (one byte per metatile)

paint_room Name reuses the existing load_background vblank-safe update machinery for the nametable blit and additionally installs the room's collision bitmap pointer into ZP_ROOM_COL_LO/ZP_ROOM_COL_HI (ZP $18/$19).

collides_at(x: u8, y: u8) -> bool JSRs into a small runtime helper that reads (room_col),Y with Y = (y & 0xF0) | (x >> 4) and returns the 0/1 byte directly. Gated on a __collides_at_used marker — programs that declare a room but never query it pay zero bytes for the subroutine.

parse_byte_array grows a [value; count] shortcut so 240-entry layouts stay readable.

See examples/metatiles_demo.ne: a probe walks right, bounces off the right wall when collides_at fires, and lands on the left side of the playfield by frame 180.

3. Register-allocator follow-up (Code quality)

remove_dead_loads now steps past opcodes that touch neither A nor the flags an LDA sets — LDX/LDY/INX/INY/DEX/DEY and the flag ops (CLC/SEC/CLI/SEI/CLD/SED/CLV) on top of the INC/DEC/STX/STY opcodes the pass already stepped past.

The highest-leverage case is every single-tile draw. Copy propagation and dead-store elimination together leave:

LDA #<y>          ; stray producer, value never consumed
LDY oam_cursor
LDA #<y>          ; real load before STA
STA $0200,Y

The first LDA was surviving because the pass bailed on the LDY. With the step-past, it drops — one LDA gone per draw, 2 bytes each.

LDA-count reductions on the committed examples:

example	before	after	Δ
platformer	242	221	-8.7 %
war	785	754	-4.0 %
pong	843	827	-1.9 %

Audio-goldens churn

The cycle savings shift the main-loop ↔ NMI boundary in audio-emitting programs, which re-times which frame each SFX trigger lands in. Six audio hashes re-baselined as a result: audio_demo, friendly_assets, noise_triangle_sfx, platformer, pong, war.

All 50 PNG goldens, the platformer/war/pong demo gifs, and every non-audio program stay byte-identical. The re-baselined audio is still sample-accurate; what changed is the first-SFX offset within the captured 132 084-sample window. This tradeoff is spelled out in docs/future-work.md's register-allocator section along with the remaining wins worth chasing next (cross-block A-tracking, X/Y allocation, spill skipping at codegen time).

Test plan

• cargo test --all-targets — 779 pass
• cargo fmt --check, cargo clippy --all-targets -- -D warnings clean
• tests/emulator/run_examples.mjs — 50/50 ROMs match their (re-baselined) goldens
• Every examples/*.nes re-committed to match its .ne source
• docs/{platformer,war,pong}.gif regenerated (byte-identical to the pre-change gifs)

What was added / removed in docs/future-work.md

Removed from §A and §H — both features ship in this PR.
Rewrote the register-allocator section to describe what's shipped (the peephole step-past), the remaining wins, and the audio-shift constraint.
Updated priority ranking to reflect the three top items being done.