PCM-clocked avatar core

This component is the allocation-free, platform-independent part of the StackChan talking head. It consumes PCM that has actually completed speaker DMA, produces a compact semantic pose, and renders one of four 160×120 RGB565 sprite atlases into caller-owned memory. It owns no task, queue, display, network connection, or framebuffer.

That boundary is deliberate. Network arrival is not playout: animating from a WebSocket callback makes the mouth lead whenever the speaker buffer grows. Conversely, allowing the renderer to retain audio creates the same accumulating delay that the voice pipeline is designed to destroy. A board platform may drop visual observations and jump to the newest physical playout frame, but it must never delay audio to preserve animation continuity.

The engine and four generated atlases were extracted from the measured StackChan prototype in ~/src/github.com/iterate/stackchan/experiments/02-minimal-realtime-aec/firmware-ws. The generated atlas headers record their CC0 1.0 dedication for the generator's contribution. Their source packs describe the images as project-generated, AI-assisted original artwork; no third-party game artwork is included here.

The renderer snapshot is intentionally a one-shot operation. A display task may share a core with a lower-priority analyzer, so spinning on an in-progress seqlock can deadlock the writer it preempted. On a race, the caller retains its last good pose and tries on the next display tick.