feat: read wave from the language, and accept the harmonics form

[spacedevin]

Draft — blocked on the deckfile 1.5.0 release from spacedevin/deck#16. The pin is bumped ahead of it, so CI will not go green until that publishes.

wave <name> <32 hex nibbles> started life here, as a fork of the .deck parser. When the fork was undone it survived as a registered dialect extension — but it was never actually GBA-specific: deck-player registered its own independent copy, and conformance/profiles.json still listed it under gba.extensions.topLevel. A statement every host implements is a language feature, so it has been adopted into the language proper.

What changed here

The hex validation and nibble unpacking are gone from this crate. The parser hands over 32 resolved levels whichever spelling wrote them, so the bake only packs two levels per byte:

for w in &program.waves {
    let mut packed = [0u8; 16];
    for i in 0..16 {
        packed[i] = ((w.levels[i * 2] as u8) << 4) | (w.levels[i * 2 + 1] as u8);
    }
    named_waves.insert(w.name.clone(), packed);
}

Songs can now write a table as harmonic amplitudes instead of digits:

wave organ 8beffecbbbbaa9888776554444310014
wave organ harmonics 1 0.5 0.33 0.2

Nothing about the ROM changes

The additive sum happens in the parser at bake time — the only place it can, since the device is no_std on ARM7TDMI with no FPU and no libm. DeckSong.waves: &'static [[u8; 16]] is unchanged, psg::wave_table still copies 16 bytes verbatim into WAVE_RAM, and the size delta is zero. A harmonics table costs exactly what a hex one did.

Verification

• Corpus digest identical across all 54 decks. DECK_CORPUS_DIGEST=1 cargo test -- --nocapture corpus before and after diffs clean — not a single note moved.
• New test wave_harmonics_and_hex_pack_identically pins the thing that would otherwise be invisible: a harmonics table and its hex literal pack to byte-identical wave RAM. This is the hardware end of the language's equality guarantee.
• cargo test — 11 pass.
• profile_test::gba_profile_is_honest passes against the new 013-wave conformance case. Worth noting it earned its keep: it caught that my first draft of that case used two type wave tracks, which the LR35902's single wave channel cannot do. The case was fixed upstream rather than the profile loosened.

No .deck file in this repo currently uses a top-level wave statement — all 44 wave tracks use built-in wave_shape — so there is no existing content to regress.

[spacedevin]

Unblocked and out of draft — pinned to the released deckfile 1.5.3.

• builds against the published crate (no path override)
• cargo test — 11 pass
• corpus digest across all 54 decks byte-identical to the pre-change baseline
• wave_harmonics_and_hex_pack_identically passes against the real crate: a harmonics table and its hex literal produce the same 16 bytes of wave RAM