[jamesbraid]

The host output had two copies of the same Catmull-Rom resampler: Resampler (stereo, writing into the host ring; HAL2 and RAD1) and FrameResampler (up to eight channels, appending to a vector; the ICE1724). MACE used the stereo one but wanted the frames back to mix its two outputs by position, so every pass built a temporary ring just to read them out again. A fix to the interpolation or the refused-frame rule had to be made twice, and the next PR in this stack (one link from a card to its host stream) needs a single resampler type to hold.

This replaces both with one Resampler for 1 to 8 channels. It calls a closure with each output frame it completes. HostOut::push writes them to the ring and counts refusals as before; MACE adds them into its mix by position; the ICE1724 pushes each frame as it is made, so its intermediate vector goes too. Nothing a guest or the host hears changes.

What changes

One commit, host_audio: resample every card with one Catmull-Rom resampler:

• Resampler::new(in, out, channels) and push(frame, emit) replace both types; HostOut::push takes a frame slice.
• MACE's temporary ring is gone; its add-by-position mix stays (a shorter output still plays alone).
• HAL2's resampler tests move to host_audio.rs next to the type.
• No CHANGELOG entry: not user visible.
• Non-test code: +66 -109, net -43.

What to check

• Bit-exactness is the claim. The f32 operations, the exact rational step and the whole-frame refusal rule are unchanged; the one rewritten line is the history shift, now per channel instead of copying four rows of eight.
• MACE's mix in MaceAudio::play: mixed[k] is added by output position, as before.
• HAL2 cost: the push is a closure over a stack array, no allocation. Measured below; low guest rates cost a few ns more per output frame.

Verified

Golden hashes, taken from the old code before either type was removed. The test drives each card's real path (HAL2 push_to_ring, RAD1 Sink::play, MACE play with two outputs at different rates and passes where one is empty, ICE1724 PcmOut::flush at 2 channels with S/PDIF mixed in and at 8) with a 1 kHz and 440 Hz tone pair, a full-scale square pair and pseudo-random noise, at guest rates 8000, 11025, 22050, 32000, 44100, 48000 to host 48000 and 44100, a frame at a time (HAL2) or in bursts of 1 to 1024, into a roomy ring ("free") and a 41-slot ring that refuses most bursts ("tight"). sha256 of the bytes that reached the ring and the refused-frame count, old code and new code identical for all twelve:

run	sha256
hal2-one/free	dbb439122cb684f91f582ae751a9b8ee8b1cb14e3c86ad326e6ee44f52e2803e
hal2-one/tight	99b0d40ef86a0da5988a87caf4f89ee03c6699c721c814b5da575732fc02d567
hal2-burst/free	dbb439122cb684f91f582ae751a9b8ee8b1cb14e3c86ad326e6ee44f52e2803e
hal2-burst/tight	72024f6f0d5f77794ec696455e44efd569db2e9a45f406df1e5260820b0170e8
rad1-sink/free	dbb439122cb684f91f582ae751a9b8ee8b1cb14e3c86ad326e6ee44f52e2803e
rad1-sink/tight	72024f6f0d5f77794ec696455e44efd569db2e9a45f406df1e5260820b0170e8
mace/free	a3f6ddecab21d8798fe100654765f3d127de6abdf299a4c384ec30f61187ecd7
mace/tight	e850e9d2d67178d3e6ac828f8f644ffbe362df5853570a2004ea58e97a283626
ice1724-2ch/free	6e6676739e2bd1a0a51eef18bbee05127316306996b6587f9787a84d7cb8a644
ice1724-2ch/tight	2ac4a64719dd8e81e97653c34b81fede7acfab52a5c3887fea48fe8e37dae7f3
ice1724-8ch/free	56a1a9b5f3017dc9ab41fe043f5402da5b79e827c2db1419bd15aa7ec8ca30a7
ice1724-8ch/tight	914f935a9af7ede6e7aa8876ce2fa795246561e8a68537976db468822f4746da

The committed test holds the same bytes as blake3 digests; the sha256 values above were taken by dumping the same bytes. Mutation checks, each failing the golden tests: interpolation coefficient changed, phase accumulator reset instead of carried, one channel's history not shifted, refused frames not counted; MACE add-by-position broken fails the MACE test. The HAL2 one-frame, HAL2 burst and RAD1 free runs hash the same, so output is also independent of push size.

• HAL2 per-frame cost, ignored benchmark (cargo test --release -p iris --lib -- --ignored hal2_push_cost --nocapture), x86-64, release, five A B B A pairs, ns per input frame, before to after: 44.1 to 48 kHz 32.4 to 27.7; 11.025 to 48 kHz 69.5 to 79.8; 8 to 48 kHz 82.7 to 106.1. At most 0.03% of a core. arm64 not measured.
• PROM boot chimes through [audio] file, headless release builds, before and after: Indy (embedded PROM): the sounding span of the WAV is byte-identical (200596 words, same sha256). Octane (6.5 PROM): the pre-resample IRIS_AUDIO_DUMP is identical; the WAV holds the same 91261 non-silent frames in the same order (identical sha256), and differs only in where the host's real-time player inserted silence on an underrun, which follows wall-clock time.
• cargo check --tests -p iris, default and --features jitv2; cargo test -p iris --lib for host_audio, host_pcm, hal2, mace_audio, rad1, ice1724, audio_dump and monitor, default and --features developer; cargo test -p iris --test snapshot_schemas. All pass. No resampler state is saved or loaded.

Not run: arm64, Windows and macOS (CI); real-device loopback tone tests; the Indigo2, O2 and Tezro chimes on a guest; the full workspace tests (CI).