audio-small (batch 2026-10-24-plan5): audrec:u, svc 0x23, shader IMAD and LEA; I3DL2 already present

Linux cloud session. Nothing here was compiled or run on macOS.

I3DL2 reverb (audren): no change needed

The effect already exists: audio/src/reverb.{c,h} is the I3DL2 reverb (effect type 5, Reverb3dParam), wired into mix.c (mixer_update_effects, the e->type==5 branch of the effect loop) and covered by audio/tests/test_reverb3d.c (impulse response, RT60, dry path, linearity, stability, validation; see 2026-10-24-audio-i3dl2-audin-review.md). It landed in the base commit before this batch's coverage numbers were taken, so the +0.39 is already earned. I changed no audio code. test_reverb3d was re-run here (54 checks, 0 failed), built without hwopus.c because libopus headers are missing on this host. The other effect types (1 mix, 2 aux, 3 delay, 4 reverb, 6 biquad, 7 limiter, 8 capture, 9 compressor) are all handled in mix.c too.

audrec:u

runtime/src/services/audrec.c, audrec_core.[ch], registered in services.c, test runtime/tests/test_svc_audrec.c. OpenFinalOutputRecorder (cmd 0) returns {48000, 2, PCM16, Stopped} and an IFinalOutputRecorder (cmds 0-9: state, start, stop, append, buffer event, released buffers, contains, end time, and the Auto variants). The host has no tap on the mix, so a started recorder records silence; appended buffers are zero-filled and complete at once. Ids and layouts are from memory (unverified), like audin.c. Result code on a full queue / unknown buffer reuses audio module 153 description 8 (a guess).

svc 0x23 SendAsyncRequestWithUserBuffer

runtime/src/kernel.c: delivers like 0x22 (through the registered 0x22 handler, args remapped x1/x2/w3 -> x0/x1/w2) and returns an already-signalled readable event in w1 (the writable end is closed). A request to a guest server thread blocks until the server replies (no background delivery). Host test: runtime/tests/host/test_svc_async.c (make -C runtime test-host).

Maxwell IMAD and LEA

gfx/src/shader/maxwell_translate.c, tests in gfx/tests/t_xlate.m. Layouts are from memory and no word in the LAS or SMO corpora uses them; the layouts are documented in comments at the decoder. IMAD: reg 0x5a, cbuf 0x4a, imm 0x34, and the 0x52 (B = REG_39, C = cbuf) form; bit 48 signed, 49 negate c, 51 negate product, 53 .HI. IMAD.CC / .X are refused cleanly. LEA.LO (reg 0x5bd0, cbuf 0x4bd0, imm 0x36d0; scale [39,44), bit 45 negate B, .CC 47, .X 48) and LEA.HI (0x5bd8, 0x4bd8; scale [28,33), c = REG_39, .X 48). No other op was touched. TRANSLATOR_VERSION is NOT bumped: these opcodes failed as "unknown instruction" before, so no cached MSL changes (it still reads 6u).

Verified vs not

tested on Linux: test_svc_audrec (61 checks), test_svc_async (7), make test-host (all), t_xlate_linux.sh (ok, Metal compile checks skipped), test_reverb3d. Not run: any macOS build, Metal compile of the new MSL, a real guest using audrec:u / svc 0x23 / IMAD / LEA.