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

[spacedevin]

wave <name> <32 hex nibbles> was never part of the grammar. It was a host-registered top-level statement — missing from DECK_GRAMMAR.md and from Highlight.tish's TOP_KEYS — so every host that wanted it registered its own copy: tish-gba, where it originated, and deck-player. Deckard never implemented it at all.

src/deckfile/Parser.tish already named the consequence:

Without it a host that needs one extra statement — tish-gba's wave <name> <32 hex nibbles> — has to fork the parser, which is how the GBA implementation ended up a separate grammar rather than a subset.

A statement every host implements is a language feature in the wrong place. And the corpus that exists to stop exactly this drift never saw a wave line: conformance parses with no host registrations, so hostStatements is {} in every expected file, while profiles.json carried wave under gba.extensions.topLevel.

What changed

wave is core. Parsed in Parser.tish before the host dispatch, so a host registering wave is now shadowed rather than authoritative (there is a test for that). Removed from the GBA profile's extension list.

New generative form. The hex is a poor way to say what a table is — timbre is harmonic content, so the digits had to be computed elsewhere and pasted in:

wave organ 8beffecbbbbaa9888776554444310014
wave organ harmonics 1 0.5 0.33 0.2

Those two lines are the same sound. a1 is the fundamental, a2 the octave above it, a3 the twelfth; the result is normalized to fill the 4-bit range, so only the ratios matter.

Expanded at parse time, exactly as steps euclid is — and for the same reason. A host reads levels and never learns which spelling produced it, so no host needs changing to keep working. The source form is kept alongside (mode, harmonics, hex) so an emitter can write harmonics back instead of flattening it to hex, which is the trap euclid avoids by retaining hits/len.

Notes for review

• The clamps in the additive path are removed rather than covered: normalizing by the peak puts every sample in 0..15 by construction, so they were dead code the 100% gate correctly flagged. What is guarded instead is the case that was reachable — a non-finite amplitude, which would have put NaN in the table.
• The conformance diff on the 12 existing cases is purely additive: one "waves": [] line each, nothing else moved.
• docs/DECK_EXTENSION.md used wave as its worked example of the host-statement hook. That would now tell hosts to register something the language owns, so the example is a cue statement instead.

Verification

• npm test — build, 100% line/function/statement coverage, 13 conformance cases, Tish VM suite, JS smoke.
• npm run test:rust — the regenerated crate passes the same corpus, plus a new facade test asserting the two spellings produce identical levels.
• packages/player — 15 tests pass, including a new one asserting a harmonics table reaches the audio graph identically to its hex literal.
• The five wavetables published on chuggie.dev reproduce byte-for-byte from harmonics, so existing documented tables are unchanged.

Downstream PRs follow in schlopai/chuggie (bake), spacedevin/deckard (first-ever wave support), and schlopai/chuggie.dev (docs). This must merge and release first — the others depend on the published deckfile crate and npm packages.

[spacedevin]

Pushed two follow-ups:

Fixed the conformance case. 013-wave originally used two type wave tracks. The GBA profile test in schlopai/chuggie caught it — the LR35902 has one wave channel — so the case now stays inside the subset it claims to be in. Fixed the case rather than loosening the profile.

Expanded the examples page (docs/EXAMPLES.md, which is what https://spacedevin.github.io/deck/ publishes). It was eight blocks, all Game Boy chip songs, and none of them used a wave table — so the page never showed this feature and read as if the language were GBA-only. Now 14 blocks across five new sections: wavetables (both spellings side by side so you can hear they are one sound), designing a timbre from harmonics, instrument voices, basicOsc, and mix/fx — which nothing demonstrated despite the channel strip applying to every generator.

Everything stays within gameBoyDmg / gbaDirectSound / basicOsc. The other 30 generator ids parse fine but the player substitutes a plain oscillator, and on a page where every block has a Play button that would be a demo that lies. The intro says so now.

test/examples.mjs is new and wired into npm test: it parses every block on the page and fails on a substitution or ignored feature, not just a parse error, so the Play buttons cannot quietly start lying.