Byte-identity, engine revisions, and streaming

The guarantees behind "audio as a pure function": what makes a render reproducible, how engine revisions let fidelity improve without ever changing an older sound, and which documents can stream live.

Determinism

• Rendering is a pure function of (graph, seed, sample_rate) — a document renders byte-identical every time. With engine: 5 (the default for new documents and songs) that identity holds across platforms; engine ≤ 4 documents keep their historical per-platform renders bit-for-byte (platform libm's last bits differ between macOS-arm64 and linux-x86_64, though integer-RNG, PolyBLEP, and rational-filter content is identical everywhere).
• The document's top-level seed drives every noise source, dust train, and Karplus-Strong pluck burst, so takes are reproducible; change seed for a different-but-equivalent roll.
• Because the document is the artifact, version your .json files and you can always reproduce the exact WAV — no separate session log needed.

Pick an engine revision

A document carries two independent version numbers: version is the schema version (document structure); engine is the DSP-kernel revision (which audio kernels render it). They are split so a fidelity upgrade never changes the bytes of an older sound.

engine	what changed
omitted (0)	the original kernels — byte-for-byte forever
1	anti-aliased drive (ADAA)
2	per-node structurally-seeded RNG for noise/dust (decorrelated siblings; byte-identical streaming randomness)
3	the inharmonic additive piano voice (stretched partials, per-partial decay, hammer spectrum, detuned unison pair)
4	corrected mixer output stage (joint stereo loudness normalization, gated BS.1770, oversampled true-peak) and per-note humanize jitter
5	deterministic transcendental kernels (det: fdlibm-grade sin/cos/exp/ln/powf/tanh in pure f64) replace platform libm everywhere in the render path; convolve runs a fixed-order radix-2 FFT with deterministic twiddles ⇒ renders byte-identically on every platform

• To modernise an existing sound, set "engine": 5 — its output will change; that's the point.
• To keep a legacy sound bit-exact, leave engine off.
• New documents and songs stamp 5 by default.

What streams live

The streaming renderer pulls a document block-by-block, byte-identical to the offline render. StreamGraph::blockers reports what a document trips, each blocker naming the fix:

streams natively	falls back to the buffer-backed Player (byte-identical, whole-buffer)
every node with constant filter/EQ cutoffs and gain amounts	a normalize output stage (whole-buffer op)
all modulators on source params (closed forms of the sample index; rand carries its walk)	a filter/EQ/gain carrying a modulated cutoff or amount
tremolo (a closed form of the sample index)	loop playback; a stereo (Haas/Wide) treatment (write-time ops)
noise/dust/seq under engine ≥ 2 (structurally-seeded RNG)	RNG nodes under engine < 2; the sampler seq
a schema-v2 tracks root — sidechains, buses, automation	a schema-v1 tracks root

convolve / granular are offline-only whole-buffer effects — no streaming form exists; bounce them offline and keep the streamed graph causal.