feat: adopt wave into the language, with a generative harmonics form - #16

…#17)

The Rust crate job has been posting four `##[error]` annotations on green runs:

  ENOENT: opendir '…/crate/target/package/deckfile-0.1.0/tests/trybuild'
  ENOENT: opendir '…/crate/target/package/deckfile-0.1.0/tests/target'

They come from `Post Cache cargo`, not from any build step. `cargo publish
--dry-run` leaves an extracted copy of the crate under `target/package/`, and
that copy has a `tests/` directory (holding `conformance.rs`). rust-cache prunes
`target/` before saving it, sees `tests/`, and goes looking for the trybuild
layout — `tests/target` and `tests/trybuild` — which this crate has never had.

It only surfaces on a cache MISS, because rust-cache skips the prune entirely on
a hit. That is why it looks sporadic: it appeared on #14, not on #15, and again
on #16, each time a change to the emitted crate moved the cache key.

Nothing should be caching the output of a dry-run publish, so remove it once the
check has served its purpose. The check itself is unchanged.

Worth fixing rather than muting: a green run that reports four errors teaches
everyone to skim past annotations, which is when a real one gets missed.

* feat: read, play and emit named wavetables

Deckard could not read a `wave` line. The statement originated in tish-gba and
deck-player added its own copy, but the reference host never had one — so a
`.deck` with a named table loaded with "unexpected top-level: wave", and the DMG
wave channel offered three hardcoded shapes and nothing else.

`wave` is core grammar now (spacedevin/deck#16), which resolves both the hex and
`harmonics` spellings to 32 levels before Deckard sees them. What was missing
here is the rest of the path:

- the project model holds the tables, keyed by name
- Apply collects them and points each channel at the table its `wave_shape`
  names; a UI knob turn arrives as a one-track `.deck` fragment through the same
  path, so the binding stays current when the shape is changed by hand
- the DMG wave channel plays a named table in preference to a built-in shape,
  which is the order the GBA bake uses
- the Wave selector lists the song's tables instead of only Saw/Square/Sine

Emit writes the tables back, and writes a `harmonics` line AS harmonics rather
than flattening it to 32 digits. That is the whole reason the parser keeps the
source spelling: a save should not throw away what the author wrote. Only tables
a channel actually names are emitted, or a project would grow a dead `wave` line
every time it loaded a file.

* chore(deps): @spacedevin/deck 1.5.3

The released language package, where `wave` is core grammar — which is what
makes this branch work against a real install rather than a linked build.

All 12 wavetable tests pass against the published package, and the existing
suites are unchanged.

…20)

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

`wave <name> <32 hex nibbles>` originated here, as a fork of the parser, and
survived as a registered dialect extension after the fork was undone. It was
never GBA-specific though — deck-player registered its own copy too — so it has
been adopted into the language proper (spacedevin/deck#16).

That deletes the hex validation and nibble unpacking from this crate. The parser
resolves both spellings to 32 levels, and all that is left here is packing two
levels per byte for wave RAM.

The additive sum lands at bake time, which is the only place it can: the device
is no_std on ARM7TDMI with no FPU and no libm, so there is no sin() on the far
side. Nothing about the ROM changes — still 16 bytes per table, copied verbatim
into WAVE_RAM, so a harmonics table costs exactly what a hex one did.

The corpus digest is unchanged across all 54 decks, and a new test pins the part
that would be invisible otherwise: a harmonics table and the hex literal it
resolves to pack to byte-identical wave RAM.

Blocked on the deckfile 1.5.0 release; the pin is bumped ahead of it, so CI will
not go green until that publishes.

* chore(deps): pin deckfile 1.5.3

The released crate, in place of the version pinned ahead of it. `wave` is core
grammar there, so this is what actually makes the branch build.

All 11 tests pass against the published crate, and the corpus digest across all
54 decks is byte-identical to the pre-change baseline.

* style: satisfy rustfmt and clippy in the wave bake

CI runs `cargo fmt --check` on every crate and `cargo clippy --all-targets -D
warnings` on this one; the wave changes passed neither.

Formatting was mechanical. Clippy found two needless range loops and one real
problem: `assert!(b >> 4 <= 15 && (b & 0x0f) <= 15)` is always true, because a
nibble of a u8 cannot exceed 15. The assertion proved nothing.

Replaced with the invariant actually worth pinning — that packing is lossless
and high-nibble-first. Getting that order wrong still yields 16 plausible bytes,
so nothing downstream would catch it; the ROM would just play a scrambled
waveform.

Tests still 11/11 and the corpus digest is unchanged across all 54 decks.

* feat: read, play and emit named wavetables

Deckard could not read a `wave` line. The statement originated in tish-gba and
deck-player added its own copy, but the reference host never had one — so a
`.deck` with a named table loaded with "unexpected top-level: wave", and the DMG
wave channel offered three hardcoded shapes and nothing else.

`wave` is core grammar now (spacedevin/deck#16), which resolves both the hex and
`harmonics` spellings to 32 levels before Deckard sees them. What was missing
here is the rest of the path:

- the project model holds the tables, keyed by name
- Apply collects them and points each channel at the table its `wave_shape`
  names; a UI knob turn arrives as a one-track `.deck` fragment through the same
  path, so the binding stays current when the shape is changed by hand
- the DMG wave channel plays a named table in preference to a built-in shape,
  which is the order the GBA bake uses
- the Wave selector lists the song's tables instead of only Saw/Square/Sine

Emit writes the tables back, and writes a `harmonics` line AS harmonics rather
than flattening it to 32 digits. That is the whole reason the parser keeps the
source spelling: a save should not throw away what the author wrote. Only tables
a channel actually names are emitted, or a project would grow a dead `wave` line
every time it loaded a file.

* chore(deps): @spacedevin/deck 1.5.3

The released language package, where `wave` is core grammar — which is what
makes this branch work against a real install rather than a linked build.

All 12 wavetable tests pass against the published package, and the existing
suites are unchanged.

… form (#27)

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

`wave <name> <32 hex nibbles>` originated here, as a fork of the parser, and
survived as a registered dialect extension after the fork was undone. It was
never GBA-specific though — deck-player registered its own copy too — so it has
been adopted into the language proper (spacedevin/deck#16).

That deletes the hex validation and nibble unpacking from this crate. The parser
resolves both spellings to 32 levels, and all that is left here is packing two
levels per byte for wave RAM.

The additive sum lands at bake time, which is the only place it can: the device
is no_std on ARM7TDMI with no FPU and no libm, so there is no sin() on the far
side. Nothing about the ROM changes — still 16 bytes per table, copied verbatim
into WAVE_RAM, so a harmonics table costs exactly what a hex one did.

The corpus digest is unchanged across all 54 decks, and a new test pins the part
that would be invisible otherwise: a harmonics table and the hex literal it
resolves to pack to byte-identical wave RAM.

Blocked on the deckfile 1.5.0 release; the pin is bumped ahead of it, so CI will
not go green until that publishes.

* chore(deps): pin deckfile 1.5.3

The released crate, in place of the version pinned ahead of it. `wave` is core
grammar there, so this is what actually makes the branch build.

All 11 tests pass against the published crate, and the corpus digest across all
54 decks is byte-identical to the pre-change baseline.

* style: satisfy rustfmt and clippy in the wave bake

CI runs `cargo fmt --check` on every crate and `cargo clippy --all-targets -D
warnings` on this one; the wave changes passed neither.

Formatting was mechanical. Clippy found two needless range loops and one real
problem: `assert!(b >> 4 <= 15 && (b & 0x0f) <= 15)` is always true, because a
nibble of a u8 cannot exceed 15. The assertion proved nothing.

Replaced with the invariant actually worth pinning — that packing is lossless
and high-nibble-first. Getting that order wrong still yields 16 plausible bytes,
so nothing downstream would catch it; the ROM would just play a scrambled
waveform.

Tests still 11/11 and the corpus digest is unchanged across all 54 decks.