[mparrett]

The audio half of #255, built from points 1 and 3 of the 2026-10-10 comment.

• Seam (pkg/rt/audio.go). rt.Audio (Play, Stream, Stop, Gain, Enabled) is bound at *audio*. Until a backend sets the root (rt.SetAudioRoot) or a Run binds one (api.WithAudio), it's a silent no-op, the same discipline as *surface*.
• lg namespace. It has audio/tone, audio/sample with :loop-from, audio/stop, audio/gain with an optional fade, and audio/enabled?. Handles are Ints. With no backend, each call returns nil, and stop and gain take a nil handle as a no-op, so a game runs silent without guarding its calls.
• Stream is Go-only. A continuously synthesizing module joins the host's mix and gets back an ordinary handle, so stop and gain work on it.
• Backend (pkg/audio). A pure-Go mixer. With -tags audio_oto, it plays through oto and binds the root at init, and lg_boot_audio.go links it into lg. Release builds set no tag, so lg stays silent and cgo-free.

Checked at this head (on 88a86e1, 2026-10-21):

• go test passes on pkg/..., cmd/... and test/..., and on pkg/audio and pkg/api with -tags audio_oto. vet and the wasm build are clean, and check-generated and check-generated-manifest are OK.
• A tagged lg plays a tone and fades it through its handle. A stock lg runs the same script silently, and audio/enabled? returns false.
• Before the rebase (2026-10-20), undertone's synth played through the mixer as a stream, with lg tones over it, and ducked through its handle.

Not here yet: the browser backend (point 5).

Refs #255

🤖 Generated with Claude Code

[mparrett]

Self-review recommended fixing the four findings below before merging: stream EOF truncates the final buffer, asynchronous device initialization errors are ignored, negative playback gain ramps in the wrong direction, and non-looping samples end without a fade.

Reviewed c0384e7. Three playback bugs were reproduced with headless regression probes; the device-error finding follows the pinned Oto v3.5.0 initialization contract and its asynchronous driver implementation. Normal, race-detector, and audio_oto tests passed for pkg/audio and pkg/api. Physical audio output was not tested.

[mparrett]

All four self-review items are fixed in 5320aa3.

• Stream EOF: a short final read now mixes every frame it returned, then the voice leaves. The new test TestStreamPlaysItsFinalShortBuffer is your 300-into-400 case.
• Device errors: open checks ctx.Err() after ready and returns the error, so Play and Stream fail and Enabled() is false. This one has no test, since it needs a device that fails to open.
• Negative gain: audio/tone and audio/sample reject a negative or infinite :gain, and audio/gain rejects an infinite one. Mixer.add also steps by the gain's absolute value, so a negative gain passed from Go is bounded.
• Sample end: a one-shot fades over its last edgeFrames, or its second half if shorter, and loops are unchanged. TestSampleFadesAtItsEnd uses your 300-frame, 10,000-amplitude case.

The three tests fail on c0384e7 and pass here. pkg/..., cmd/... and test/... pass, along with -race and -tags audio_oto on pkg/audio and pkg/api, the wasm build, and both generated checks.

[nooga]

Approving. I checked that the four self-review findings (stream EOF, async oto init error, negative gain, one-shot sample fade) are fixed in the code, with tests. Default builds carry no oto and the wasm/plan9 builds are clean. As project owner I'm fine with the new public surface (rt.Audio, SetAudioRoot, api.WithAudio).

Nits, none blocking:

• Run go mod tidy: oto/v3 v3.5.0 should be a direct require, and go.sum has stale v3.5.1 lines.
• render calls io.ReadFull on a stream while holding m.mu. A source that returns (0, nil) or short reads forever would spin with the lock held. Consider a single Read per tick with zero-fill on underrun, and close Closer sources on stop or end.
• The tone's end fade sets target but not stopping, so a Gain() call during the last 220 frames overwrites it and the tone ends with a click.
• :freq ##Inf passes validation and turns the phase to NaN; :ms ##Inf has arch-dependent results. Check with math.IsInf.
• Go API only: Mixer.Play panics on odd-length PCM or empty PCM with LoopFrom>=0, and Mixer.Gain doesn't check for NaN.