[joshuaswarren]

Honeykrisp integration branch: Apple Silicon (AGX) correctness + coopmat + G13G CDM trim

Brings omacom/mesa to the line the Omarchy M+ team actually runs and measures on M1-class hardware. 43 commits on top of the Sep 12 stock sync. Three parts:

1. Correctness fixes (upstreamable — see the series branch)

• agx: keep the mod-32 shift count when fusing ushr+iand into ubfe — wrong-results fix for the shift-then-mask fusion
• agx: represent the fused shift-then-mask as ubfe, not ubitfield_extract
• nir: pass FP_MATH_CTRL to the builder generator — the NIR generator bug that affects every Mesa driver, not just AGX
• agx: correctly rounded fdiv/frcp, faithful log/log2, exact sin/cos reduction + FTZ handling + the late fdiv lowering

A clean 6-patch series with these commits only (no receipts, no perf-only changes) is staged at joshuaswarren/mesa:upstream/correctness for forwarding to gitlab.freedesktop.org.

2. VK_KHR_cooperative_matrix for AGX

Enabled by default (opt-out via AGX_SIMDMAT=0): 16x16x16 and fp16 shapes, hardware 8x8x8 fp32 via simd_matrix_fmadd32, ISA disassembly support, software lowering outside full subgroups, shader-cache key handling, and the regression tests. This is what mlx-omarchy's cooperative-matrix path needs on M1.

3. Per-launch CDM barrier trim — G13G only

Trims the kitchen-sink CDM barrier after each compute launch on t8103 (M1) to the designed bit set: +6.3% ctx1024 decode, pins held 24/24, runtime suite 22 cases / 6,189 assertions green (six interleaved A/B rounds, mlx-omarchy v0.6.0 wheel untouched). Deliberately not enabled on G13X (M1 Max): the port measured +13% short / −3.2% long-context and was reverted — this branch carries that history (5deac1c8068 restores the G13X sink).

Validation

• mlx-omarchy v0.6.0 release gate on this line: pinned ctx1024 decode 102.29 tok/s, Parakeet 3× warm byte-identical (104/104)
• 25-suite mlx-omarchy battery, stock vs merged
• oMLX runtime suite: 22 cases / 6,189 assertions, 0 failures
• Five miscompiles reproduced and fixed; standalone repros included (the repro/ commits)

Not in this branch

• The hk/* experiment branches (G13X barrier ports, default-off diagnostic knobs, trig-invariance, agx-wait-batching, etc.) stay unmerged until measured — they are not part of this PR
• Internal receipts/ evidence commits ride along for team review; they will be stripped before any freedesktop submission