nescript/docs/future-work.md

Future Work

This document tracks the gaps between what NEScript currently compiles and what the spec describes. Items are grouped by area. Anything implemented and tested is omitted — git log is the authoritative record of what shipped.

---

PNG-sourced assets

What ships today. palette Name { colors: [...] } and background Name { tiles: [...], attributes: [...] } declarations with inline byte arrays, plus palette Name @palette("file.png") and background Name @nametable("file.png") for PNG-sourced variants. The palette path maps each pixel to its nearest NES master-palette index (via nearest_nes_color() in src/assets/palette.rs), deduplicates, and emits the 32-byte blob; the nametable path slices a 256×240 PNG into the 32×30 tile grid, deduplicates (max 256 unique tiles), and emits the 960+64 byte nametable/attribute blobs. The nametable path now also auto-generates the per-tile CHR data via png_to_nametable_with_chr and slots it into CHR ROM after the user's sprite tile range — see examples/auto_chr_background.ne for the end-to-end flow. --memory-map reports per-blob PRG ROM addresses and a running total alongside the variable layout.

Still TODO.

• Per-state background rendering control — programs currently load a single nametable at reset. Per-state swaps work but are limited by the NMI-time write budget (~2273 cycles, enough for a palette but not a full 1024-byte nametable).
• Per-quadrant palette selection from PNG sources — the png_to_nametable_with_chr attribute path picks sub-palettes based on brightness buckets, which is fine for grayscale demos but doesn't let the user say "this 32×32 tile uses sub-palette 2". A separate palette_map: shortcut exists for inline backgrounds; the PNG path could grow a sibling @palette_map("hint.png") that overrides the brightness buckets.

---

User code distribution across switchable banks

What ships today. bank Foo { fun bar() { ... } } nesting places user functions into a specific switchable bank. The codegen emits per-bank instruction streams; the linker runs a two-pass assembly (discover labels per-bank, then resolve with the merged label table) so banked code can still reference fixed-bank symbols. Cross-bank calls — both fixed → banked and banked → banked — are rewritten to JSR __tramp_<name>, where each trampoline is a per-function stub in the fixed bank that reads the caller's current bank from ZP_BANK_CURRENT, pushes it on the hardware stack, switches to the target, JSRs the entry, then pulls and restores the caller's bank. gen_mapper_init seeds ZP_BANK_CURRENT with the fixed bank index at reset so the first cross-bank call from the fixed bank still leaves the fixed bank mapped at $8000. See examples/uxrom_user_banked.ne (fixed → banked) and examples/uxrom_banked_to_banked.ne (banked → banked).

Still TODO.

• Greedy size-packing. Placement is explicit-only today — there is no pass that takes a program with too much fixed-bank code and automatically spills the biggest leaf functions to declared empty banks.
• MMC3 per-state-handler split — the mmc3_per_state_split.ne example still uses the legacy fixed-bank placement for its handlers. Extending the banked-fun syntax to state handlers (plus trampoline emission on handler dispatch) would unify the two paths. The blocker isn't the trampoline — those work for any caller now — but the state-handler dispatcher in the IR codegen needs to learn that state handlers can live in a switchable bank, and to JSR through a trampoline whose entry is the handler label.

---

Language feature gaps (post-v0.1)

From the spec's "Reserved for Future Versions" section:

Feature	Description
Fixed-point	fixed8.8 type for sub-pixel movement with operator support.
Text / HUD	Font sheet declarations + layout system for scores, health, menus.
Tilemaps	Declarative level data with built-in collision queries.
SRAM / saves	Persistent storage declarations for battery-backed save data.

NES 2.0 headers are now supported via game Foo { header: nes2 } — see src/rom/mod.rs.

Metasprites are now supported via metasprite Name { sprite: ..., dx: [...], dy: [...], frame: [...] } — see examples/metasprite_demo.ne. The IR lowering expands draw Hero at: (x, y) into one DrawSprite op per tile, with each tile's frame index offset by the underlying sprite's base tile so the codegen sees a stream of regular draws and the OAM cursor allocator picks them up unchanged. Negative offsets and runtime-varying tile selection are still TODO — the current form takes literal u8 offsets.

Struct / array field widths

Nested struct fields (hero.pos.x) and array struct fields (hero.inv[i]) now compile end-to-end. The analyzer recursively flattens the struct layout into per-leaf synthetic variables (with intermediate Struct(...) symbols for the dotted prefixes), and the parser loops the dotted chain in parse_primary and parse_assign_or_call so the existing format!("{name}.{field}") synthetic-name model still works without IR changes. Array-of-structs is still rejected with E0201 — the synthetic-variable model can't index per-element struct layouts without further codegen work, see src/analyzer/mod.rs::register_struct.

---

Audio pipeline

What ships today. Frame-walking pulse driver with sfx Name { duty, pitch, volume } and music Name { duty, volume, repeat, notes } blocks; builtin effects and tracks; a 60-entry period table; __audio_used marker that elides the whole subsystem when no program statement references it. Plus channel: triangle and channel: noise on sfx blocks, which splice in per-channel slots that write to $4008-$400B (triangle) or $400C-$400F (noise) when a program declares them. Plus per-frame pitch envelopes on Pulse-1 sfx — a pitch: array with more than one distinct value opts into a separate __sfx_pitch_<name> blob that the audio tick walks in lockstep with the volume envelope, writing $4002 on every NMI for a real frequency-sweeping pulse channel. Pulse-only programs without varying-pitch sfx still produce byte-identical driver code. See examples/noise_triangle_sfx.ne and examples/sfx_pitch_envelope.ne.

Still TODO for richer audio.

• DMC channel — delta-modulation sample playback is not wired yet.
• Multi-channel tracker playback — one notes list per channel on music blocks (the triangle/noise SFX are one-shot envelopes, not a tracker).
• @sfx("file.nsf") / @music("file.ftm") — neither the NSF nor the FamiTracker format is parsed yet.
• Per-frame pitch envelopes on triangle / noise sfx — the data shape (a parallel pitch array on the sfx block) is the same as for Pulse-1, but the runtime triangle/noise tick blocks currently only write their volume registers ($4008 / $400C). Extending them to also walk a per-channel pitch envelope and write $400A / $400E is the natural next step now that the pulse path is proven.

---

Debug instrumentation

What ships today. debug.log(...) and debug.assert(...) lower to $4800 writes when --debug is passed, and are stripped entirely in release builds. --symbols <path> writes a Mesen-compatible .mlb file listing function, state-handler, and variable addresses (with PRG ROM offsets for code and CPU addresses for RAM). --source-map <path> consumes the SourceLoc IR op and writes a plain-text map of <rom_offset> <file_id> <line> <col> entries for every lowered statement. --dbg <path> writes a ca65-compatible .dbg debug-info file that Mesen / Mesen2 / fceuX pick up automatically for source-level stepping, labelled variable inspection, and symbol-based breakpoints. The file stitches together the linker's label table, the __src_<N> IR markers, and the analyzer's variable allocations into the file/mod/seg/scope/span/line/sym records documented at https://cc65.github.io/doc/debugfile.html. ooffs on the segment record tracks the fixed bank's PRG-relative start, so banked ROMs (MMC1/UxROM/MMC3) also map cleanly inside the debugger. Debug builds emit array bounds checks (CMP against size, BCC past a JMP __debug_halt wedge) and bump an overrun counter at $07FF in the NMI handler when the main loop didn't reach wait_frame before the next vblank.

Plus four query expressions that mirror the counter/sticky pattern: debug.frame_overrun_count() / debug.frame_overran() return the cumulative overrun counter and a per-frame sticky bit so user code can write debug.assert(not debug.frame_overran()) guards, and debug.sprite_overflow_count() / debug.sprite_overflow() do the same for the NES PPU's sprite-per-scanline flag ($2002 bit 5), which the NMI handler samples once per frame in debug mode. All four sticky bits clear on the next wait_frame.

Still TODO.

• debug.overlay(x, y, text) — needs the text/HUD subsystem (see Language feature gaps).

---

Code quality / tooling

Register allocator

IR temps still spill to a recycled zero-page slot ($80-$FF). Most of the obvious waste is now handled:

• The remove_dead_loads peephole pass steps past opcodes that touch neither A nor the flags A cares about — INC/DEC/STX/STY/LDX/LDY in memory mode, INX/INY/DEX/DEY, and the flag ops (CLC/SEC/CLI/SEI/CLD/SED/CLV). That lets a later LDA overwrite kill a redundant earlier LDA even when an index-register load or counter bump sits between them. The canonical LDA #imm / LDY oam_cursor / LDA #imm / STA $0200,Y shape emitted by every single-tile draw collapses to just LDY oam_cursor / LDA #imm / STA $0200,Y, saving 2 bytes per draw. Across the committed examples this shaved 4-9 % of the LDA count in the larger programs (platformer, war, pong).
• Copy propagation + dead-store elimination in the same pass handle the STA $80 / … / LDA $80 spill/reload pattern whenever the source and destination aren't separated by a JSR or a label.

Constraint worth recording. Any cycle-saving optimization shifts audio timing: the SFX/music driver reads a frame counter on every NMI, and shortening the main-loop pass between two wait_frame calls changes how many poll iterations complete before the next NMI fires. The emulator harness captures a sample-exact audio hash at frame 180, so any such change flips the audio goldens for every audio-emitting example (audio_demo, friendly_assets, noise_triangle_sfx, platformer, pong, war). The visual PNG goldens stay stable — PPU writes still land in the same vblank — but the audio churn is real and has to be accepted on every pass that touches the codegen. The initial LDX/LDY extension on this branch went through that re-baseline once; future passes will too.

Remaining wins to chase if someone wants to push density further:

• Cross-block A-tracking. remove_redundant_loads clears its A-equivalence tracker on LDX/LDY because LDX/LDY clobber the N/Z flags that an immediately-following branch might rely on. Splitting the tracker into "value still valid" vs "flags still valid" lets it keep the value-side equivalence across LDX/LDY for downstream STA rewrites.
• CFG-aware allocation into X / Y. The codegen never picks X or Y as a temp slot today; they're used only in the specific shapes the existing IR ops emit. A pass that takes a temp with ≤2 uses, lives in a straight-line block, and picks X or Y would remove both the STA and the matching LDA.
• Skipping the spill for temps consumed by the very next op. The IR codegen always emits STA slot after producing a temp; when the next op's first source is that temp, the value could stay in A and the consumer skips its LDA. The peephole's remove_sta_then_lda
  • remove_dead_temp_stores already pick up most cases post-hoc, but producing tighter code at codegen avoids the overhead and handles cases where an intervening label blocks the peephole.

State-local memory overlay follow-ups

State-local variables are now overlaid across mutually-exclusive states (see the analyzer's per-state allocation cursor rewind and the IR lowerer's on_enter initializer prologue), but a few pieces are still missing:

• Same-named locals across different states. register_var stores state-locals under their bare name, so two states each declaring var timer: u8 collide with E0501. A per-state symbol-table scope prefix would let each state carve its own namespace while keeping the overlay.
• Struct-literal and array-literal initializers on state-locals. The on-enter prologue lowers scalar initializers cleanly, and struct-literal initializers fall back to per-field stores, but array-literal initializers (var xs: u8[4] = [1,2,3,4]) are skipped. A runtime memcpy from a ROM blob into the overlay slot (mirroring the reset-time global path) is the natural lowering.
• Handler-local overlay. Handler-local vars declared inside on_frame { ... } are already per-handler scoped via current_scope_prefix, but they get a dedicated RAM slot for the program's lifetime. Overlaying them inside each handler's stack frame — using a per-handler bump allocator that resets on each call — would shave a few bytes more on programs with many deep handlers.

Cross-block temp live-range analysis

The slot recycler is function-local per-block. Temps that flow across block boundaries get a dedicated slot for the entire function, even if a later block could reuse the slot.

WASM build target

To build a browser IDE we would need to route file I/O through a trait so the core pipeline works on &str → Vec<u8> without touching std::fs. Today the parser's preprocess pass and the asset resolver read files directly.

---

Error message polish

Unused error codes

ErrorCode only defines codes that are actually emitted. Previously there were placeholder variants (E0202 invalid cast, E0403 unreachable state) marked #[allow(dead_code)]; those were removed during cleanup. If those semantics come back, add the codes at that point.

---

cc65/nesdoug parity gaps

The nesdoug tutorial series + neslib expose a broad surface that NEScript can't currently express. This section enumerates the gaps in the order they should probably be tackled (cheapest/highest- leverage first). Anything we finish moves out of this section and into the top of the file with a "ships today" note.

A. Numeric types beyond u8 / i8 / u16 / i16 / bool

i16 ships today; see examples/i16_demo.ne and examples/signed_compare.ne. Signed ordering comparisons and narrow-to-wide sign-extension both ship too — Cmp{Lt,Gt,LtEq,GtEq} and their 16-bit siblings carry a Signedness field that drives the canonical CMP / SBC / BVC / EOR #$80 overflow-correction idiom in gen_cmp_signed_set_n, and widen() emits IrOp::SignExtend for signed narrow values instead of the zero-extended LoadImm 0 it used to. Casts to u8/u16 strip the signed flag so explicit as opt-outs stay unsigned.

Lower-priority numeric follow-up:

• u32 / i32. Realistically needed only for score totals and frame counters. A synthesizable pair of 16-bit halves is usually enough.

B. Pointers & function pointers

NEScript has no pointer type. This blocks indirect-dispatch tables (jmp (vec,x)), variable-size buffer manipulation, and passing "which thing" to a helper. cc65's __fastcall__ function pointers via wrapped-call + bank IDs are load-bearing for every game over 32 KB. Design sketch:

• Introduce *T / fn(T) -> U type grammar.
• Spell a new IR op CallIndirect that takes an address in a 16-bit temp, plus a BankHint so cross-bank pointers trampoline automatically.
• For fixed-bank-only code we can lower to a raw JSR ($vec) equivalent (JMP ($vec) + return stub).

C. Bitfields and unions

OAM attribute bytes, controller masks, collision-flag words, and MMC3 register bits all want bitfield syntax (struct OamAttr { pal: u2, priority: u1, flip_h: u1, flip_v: u1 }). Unions show up less often but are useful for reading/writing the same bytes as two different shapes (e.g. a 16-bit counter viewed as lo: u8, hi: u8).

D. Full inline-assembly escape hatch

.label: ships today — the codegen mangles dot-labels into per-block unique names (__ilab_<N>_label) so two inline-asm blocks in the same function can both use .loop: without colliding. See tests/integration_test.rs::inline_asm_dot_labels_are_per_block_unique.

Still TODO:

• Accept cc65's asm() format specifiers — %b (byte), %w (word), %l (long), %v (var), %o (offset), %g (global), %s (string) — so users can splat a compiler-allocated symbol into a hot-loop fragment.
• Extend the directive allow-list: .byte, .word, .res, .repeat / .endrep, .macro / .endmacro. The assembler can already encode these.

E. Dense-match jump tables

match u8 { 0 => ..., 1 => ..., ... } desugars to an if/else chain at parse time today, which is O(n) compares. For dense (<= 256 entries with <= 4× spread) integer matches, lower to:

ASL A           ; index *= 2
TAX
LDA table_lo,X
STA $00
LDA table_hi,X
STA $01
JMP ($0000)

…with a per-branch .word table emitted in the function prologue. This matters most for state dispatch and attack/weapon tables.

F. Recursion stance (design constraint, not a bug)

The analyzer rejects recursive calls with E0402. That's the right call for a compiler targeting a 6502 hardware stack, but it's not documented as a design choice anywhere. Add a paragraph to the language guide explaining why, plus a pointer to the hand-rolled explicit-stack pattern (small u8[N] stack + u8 top).

G. VRAM update buffer follow-ups

nt_set(x, y, tile), nt_attr(x, y, value), and nt_fill_h(x, y, len, tile) ship today — see examples/vram_buffer_demo.ne (minimal drain test) and examples/hud_demo.ne (realistic score/lives HUD with shadow-compare writes and one-shot attribute paint). The runtime ring lives at $0400-$04FF (gated on the __vram_buf_used marker; the analyzer bumps the user-RAM bump pointer to $0500 when the buffer is in use). Each append lays down [len][addr_hi][addr_lo][data…] and writes a fresh 0 sentinel; the NMI drains the buffer at vblank via LDA $0400,X / STA $2007 indexed-absolute (4 cycles per data byte, no ZP cost), then resets $2006/$2005 to zero so the PPU's scroll latch is clean for the next frame (otherwise trailing $2006 writes shift the rendered image by however many cells we wrote past $2000).

Still TODO:

• Vertical (column) writes — nt_write_v(x, y, ...) would set $2000 bit 2 (auto-increment 32) before the data writes and clear it after. Useful for tilemap-driven scrolling.
• Variable-length writes from a u8 array global — today nt_fill_h repeats one tile; a nt_copy_h(x, y, src_var, len) variant that copies from a declared u8[N] global removes the fill-only restriction.
• Buffer-overflow detection — the runtime drain assumes the 256-byte buffer never overflows. A debug-mode check that traps when head would advance past $04FF would catch the worst failure mode (writes wrapping into adjacent RAM).

H. Metatiles + collision as a first-class construct

metatileset Name { metatiles: [...] } and room Name { metatileset: ..., layout: [...] } ship today — see examples/metatiles_demo.ne. Each metatile bundles four CHR tile indices (TL / TR / BL / BR) plus a collide: true|false flag; rooms lay them out as a 16×15 grid that the compiler expands at compile time into a 32×30 nametable, a 64-byte attribute table, and a 240-byte collision bitmap. All three blobs land as PRG ROM data blocks with per-room labels (__room_tiles_N, __room_attrs_N, __room_col_N).

paint_room Name reuses the existing load_background vblank-safe update machinery against the room's tile/attribute blobs, 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 where Y = (y & 0xF0) | (x >> 4) and returns the 0/1 bit directly. The helper is gated on a __collides_at_used marker, so a program that declares a room but never queries it pays zero bytes for the subroutine.

Still TODO.

• NEXXT file import (@room("file.nxt")). Today only inline layout: [...] (or [value; count] for repeated fills) is supported. The NEXXT exporter's .nxt format would let users author rooms in a dedicated editor.
• Per-quadrant palette hints. The generated attribute table is currently all zeros (sub-palette 0 for every quadrant). A palette: 0..3 field on each metatile entry could drive the attribute table expansion.
• collides_at against u16 coordinates. The intrinsic takes u8 x/y today, which covers the full 256×240 visible area but can't address scrolling playfields bigger than one nametable.
• Multi-room streaming. The ZP pointer model already supports multiple rooms (each paint_room rewrites the collision pointer), but the nametable blit still uses the nametable-0-only load_background path. Streaming into a VRAM update buffer would let rooms swap without the vblank-budget risk.

I. RLE + LZ4 nametable decompression

vram_unrle and vram_unlz4 — for scrolling/multi-room games, packing rooms is mandatory. cc65 ships both in neslib with concrete timing (0.5f RLE vs 2.8f LZ4). The per-state background swapping item in "What ships today" is exactly this problem: without a decompressor that can stream into the VRAM buffer, the NMI-time write budget (~2273 cycles) is too tight for a full nametable. RLE is the smaller first step — emit a nametable that can declare compression: rle and decompress at swap time.

J. Palette-fade brightness LUT follow-up

set_palette_brightness(level: u8) and the blocking fade_out(step_frames) / fade_in(step_frames) builtins all ship today — see examples/palette_brightness_demo.ne and examples/fade_demo.ne. One follow-up still worth doing: a brightness-LUT path that recolours the active palette in addition to the emphasis bits, for non-NTSC-assumption fades. The current implementation only manipulates $2001 emphasis bits, so the "dimmed" end of the fade still shows colour tint rather than a true colour-space darken.

M. Sprite cycling follow-ups

Auto sprite cycling ships today via game { sprite_flicker: true } — the IR lowerer injects an IrOp::CycleSprites at the top of every on frame handler when the flag is set. See examples/auto_sprite_flicker.ne. A companion draw ... priority: pinned modifier for HUD sprites that must stay at low OAM slots is still missing — today pinning has to be manual (draw the HUD sprites first).

O. DPCM / DMC sample playback

Already listed under Audio Pipeline. FamiStudio's DMC support (including bankswitched DMC) is the reference API shape — import @dpcm("file.dmc") into a named sample slot and expose play_dpcm(Slot, pitch: u8, loop: bool).

P. Expansion audio (VRC6, MMC5, FDS, N163, S5B, VRC7)

FamiStudio has a single export path with FAMISTUDIO_CFG_EXTERNAL and per-chip feature flags. If/when we import a FamiStudio-export format (see Q), the expansion chips come along almost for free — the runtime just has to wire up the extra write ports and the mapper has to expose them (MMC5 for the extra pulse channels, VRC6/VRC7 via their own mappers).

Q. FamiStudio text-export import

@music("file.famistudio.txt"). FamiStudio's text export is the pragmatic ingestion path; parsing it gives full tracker semantics (volume/pitch slides, arpeggios, vibrato, release notes) without reinventing the engine. FamiTracker's binary .ftm is a worse target — undocumented, version-skewed.

R. NEXXT metatile/collision import

NEXXT is the dominant asset editor in the nesdoug workflow; it emits metatile tables + collision maps as ca65-compatible assembler source. An @metatiles("room.nxt") loader (and @room("level1.nxt") for layouts — see §H) removes a whole class of hand-typed tile arrays.

S. SRAM / battery-backed saves follow-ups

save { var ... } ships today — see examples/sram_demo.ne. The analyzer allocates save vars from a separate $6000+ bump pointer, the linker flips iNES byte-6 bit-1, and W0111 warns when a save var carries an initializer (SRAM is preserved across power cycles so initializers would either silently not run or clobber the player's data on every boot). Two follow-ups are still worth doing:

• First-power-on detection — an on first_boot { ... } handler that runs once when a magic-byte sentinel is missing. Today users have to roll the sentinel check by hand.
• Struct fields in save blocks — currently rejected because the field-flattening path uses the main-RAM allocator. Routing it through the SRAM allocator instead is a few lines of analyzer refactor.

T. PAL/NTSC region abstraction

Neslib exposes ppu_wait_frame as a virtual-50Hz wait on PAL. Add a region: ntsc | pal | dual field on the game { } block. For dual, the runtime probes $2002 bit 7 timing at reset and sets a ZP flag; the audio engine's frame tick and any frame-counted timing respects the flag.

U. Additional controller types

Expose Zapper (light-gun) and Power Pad via typed inputs:

input gun: zapper on port: 1
input mat: power_pad on port: 2

gun.trigger, gun.light_detected, mat.button(i: u8) are the three reads every program needs.

V. Additional mappers

AxROM (mapper 7), CNROM (mapper 3), and GNROM (mapper 66) all ship today; see examples/axrom_simple.ne, examples/cnrom_simple.ne, and examples/gnrom_simple.ne. CNROM and GNROM both have CHR bankswitching but user-visible CHR swaps aren't reachable from user source yet — the reset-time init writes bank 0 and the __bank_select routine exists but has no user-exposed API. The next set:

1. MMC2 (mapper 9, Punch-Out only realistically). Medium.
2. UNROM-512 (mapper 30). The modern homebrew sweet spot — 512 KB PRG + CHR-RAM + self-flashing. Mapping is UxROM-like plus a one-screen bit.
3. MMC5 (mapper 5). Big. Driven by FamiStudio's expansion audio more than by the extra PRG/CHR modes. Probably last.

Each new mapper needs a Mapper::X variant, a reset-time gen_xrom_init() in the runtime, bank-select support in gen_bank_select(), and an iNES mapper number in rom::mapper_number. The PR checklist ("example + behaviour test + negative test") is still the bar for each of these.

W. NSF output target

The audio engine is already a standalone subsystem. An NSF-output target (--target nsf) would wrap the existing music/sfx blocks in the NSF header and expose init/play entry points. Nearly free, gets the chiptune audience for ~a day of work.

X. Mesen trace-log documentation follow-up

game { debug_port: fceux | mesen | 0xXXXX } ships today — see the integration test debug_log_targets_configured_port. Mesen's trace-log tool invocation still isn't documented anywhere in the NEScript docs; a short section under docs/nes-reference.md walking through how to hook a Mesen trace-log against a ROM built with debug_port: mesen would close the gap.

Y. FCEUX .ld line-info follow-up

--fceux-labels <prefix> ships today and emits <prefix>.<bank-index>.nl + <prefix>.ram.nl. FCEUX also reads a .ld line-info file for source-level stepping; wiring that up against the existing source-map data would close the loop without much code.

Z. Explicit bank-placement hints on functions and data

bank Foo { fun bar() } already exists; extend the sugar to attributes on individual items so users don't have to restructure their source:

@bank(3) fun slow_helper() { ... }
@bank(3) const LEVEL_DATA: u8[1024] = [...]

This is particularly useful for const data, which today lands wherever the analyzer decides; users sometimes need to pin data to a specific bank to avoid bank-switch cost on a hot path.

Priority ranking

Remaining gap items in order of user value:

1. Register allocator (existing section) — compounding size win, but with the audio-goldens caveat spelled out in that section.
2. NEXXT @room(...) import + per-quadrant palette hints (§H follow-ups) — most impactful extensions on top of the shipped metatiles base.
3. Inline-asm format specifiers + directive list (§D follow-ups).
4. VRAM buffer follow-ups (§G) — vertical writes, array copy, overflow detection.
5. Arrays-of-structs + bitfields (§C) + fn pointers (§B) — turns NEScript into a general-purpose NES language.
6. UNROM-512 + MMC5 (§V) — ecosystem fit.
7. FamiStudio import (§Q) + DPCM (§O) + expansion audio (§P).

---

Open design questions

1. Inline asm label syntax. .label: (ca65 style) vs label: (generic)? Today the inline-asm parser accepts label: but not .label; migrating would be cheap but would invalidate any copy-pasted ca65 fragments.
2. Debug port address. $4800 is conventional but not universal. Should we support multiple debug output methods?
3. Pinned OAM slots for HUD sprites. Automatic sprite cycling is decided — both the manual cycle_sprites keyword and the game { sprite_flicker: true } attribute ship today (see §M). Open question: should we add a draw ... priority: pinned modifier that opts a specific sprite out of the rotation so HUD elements stay at low OAM slots?
4. Error recovery granularity. How aggressively should the parser recover? More recovery means more errors per compile but also risks cascading false errors.