PLAN.md: record Round 30 benchmark results (610.75 tok/s GPU prefill, 406.98 tok/s 2K prefill, 446.59 tok/s 11K prefill, mm_down3 -11.9%)

Round 30 blitz: W4<T_Q2_0>::dec FP16 0x6400 exponent trick (-10.5% mm_down3 / -37 ms in kprof -> 610.75 tok/s GPU prefill) + gdn_pre/post float4 simd_sum

PLAN.md: record Round 29 benchmark results (mm_gu3 T_IQ2_S -10.0% in kprof, moe_route -26.4%, deq<T_IQ2_S/XXS> -1.1..-6.9%, 24.95 tok/s @10.2K decode)

Round 29 blitz: mm_gu3 T_IQ2_S scale caching (-86.2 ms / -10% in kprof) & T_IQ2_XXS byte loads (-6.5%), moe_route e512/res_mask (-20.6..-26.4%) + deq<T_IQ2_S/XXS> (-1.1..-6.9%)

PLAN.md: record Round 28 benchmark results (mtp.fc_hidden [t][c] stream batching -52.1%, ple.metal tg_sum barrier halving, sample.metal unroll)

Round 28 blitz: mtp.fc_hidden [t][c] stream batching (-52.1% n=2 / -29.3% n=3), ple.metal tg_sum barrier halving (-1.5..-5.8%), sample.metal tg_argmax unroll + router_route_8

PLAN.md: record Round 27 benchmark results (gdn_step_t 5->4 barriers, qsa_merge2 float2(m,l) + simd_max, gemv_tile n_out=512 32,2,4,1)

Round 27 blitz: gdn_step_t 5->4 barriers + dt/ssm_a hoist + red2[IB-1] (-3.0..-7.9%), qsa_merge2/qsa_merge float2(m,l) + simd_max (-8..-20%), gemv_tile n_out=512 32,2,4,1 (-3.2..-25.5%) + Q2_0 640->2560 8,2,4,1

PLAN.md: record Round 26 benchmark results (qsa_attn_mm -23..-30%, gemv IQ4_XS packed_uint4, DW<IQ4_XS/IQ4_NL/Q6_K> unroll)

Round 26 blitz: qsa_attn_mm 32-bit ko/vo + float4 loads (-23.3..-29.9%), gemv IQ4_XS packed_uint4 5-reg codes (-1.5..-5% cold DRAM), DW<IQ4_XS/IQ4_NL/Q6_K> unroll + MM_V3_ONLY fast compile

PLAN.md: record Round 25 benchmark results (mm_down3 Q2_0 64-elem block step + run_q<2/6/10/14> -11.8% in kprof, gdn_post preload)

Round 25 blitz: mm_down3 Q2_0 64-elem block step + run_q<2/6/10/14> (-11.8% in kprof, 361.9->319.2 ms), gdn_post gate/gamma preload (-0.7%), mm_gu3 half4 HS store pointer hoist

PLAN.md: record Round 24 benchmark results (W8<T_Q6_K> +12.8..24.9%, qsa_flash -11.71% @8K, TAB=1024/512)

Round 24 blitz: W8<T_Q6_K> (+12.8..24.9%), qsa_flash 32-bit ko/vo (-11.7% @8K), TAB 1024/512 + gr_norm_h_moec k-outer + gemm_hc_down *0.25f

PLAN.md: record Round 23 benchmark results (42.01 tok/s decode, 24.78 tok/s @10.2K, 3,293.32 ms / 602.1 tok/s GPU kprof)

Round 23 blitz: ksg popcount + gemv_direct_x/gemv_q2_raw_combine + mm_gu3 sign4_iq2xs/ptr-folding -> 42.01 tok/s decode & 602.1 tok/s GPU prefill

PLAN.md: record Round 22 benchmark results (35.66 tok/s coding decode, 24.71 tok/s @10.2K decode, 401.24 tok/s 2K prefill)

Round 22 blitz: gr_down_4_4/_r2 NR=2 xn reuse (-8.1..26.1%), moe_down/moe_gu q2_combine (-7.7..22.5%) + plan_body simd_prefix (-14.6..35.4%), swiglu_prep barrier-free -> 35.66 tok/s coding & 24.71 tok/s @10.2K decode

PLAN.md: record Round 21 benchmark results (41.66 tok/s decode, 36.74 tok/s reasoning, 33.04 tok/s persona_chat, 193.94 ms mtp_ms)

Round 21 blitz: gr_down/gr_up GR_NTOK unroll (-5.4..18.0%), qsa_attn2/qsa_merge2 (-36.8..40.6% / -8.3..14.6%), gdn_step st_off + moe_plan topk10 unroll_count(10) (-9.5..26.6%) -> 41.66 tok/s decode & 193.9 ms mtp_ms

PLAN.md: record Round 20 benchmark results (40.57 tok/s decode, 202.37 ms / 122.18 ms mtp_ms, 186.15 ms gdn_scanr_4)

Round 20 blitz: mtp_load mlock 708 MB experts (mtp_ms 398->122 ms on 11K, 433->202 ms on 2K; 40.57 tok/s decode), qsa_attn (-20..40%) + qsa_merge (-10..13%), gdn_scanr_4 32-bit offsets (186.15 ms kprof)

PLAN.md: record Round 19 benchmark results (40.44 tok/s decode, 212.46 ms mtp_ms)

Round 19 blitz: gemv NTOK/NR/KS unroll (-5.4% to -29.4% at ntok>=2), mtp_pick/slop unroll_count(4) (-27..60%), ple_comb_query_gate_norm + load-time prewarm -> 40.44 tok/s decode & 212.5 ms mtp_ms

PLAN.md: record Round 18 benchmark results (40.03 tok/s decode, 445.79 tok/s 10.18K prefill)

Round 18 blitz: per-stream gr_norm_comb 5.09x (-80%), xprep float4 (-46.8%), moe_combine_sh (-49%), Q6_K moe_gu NE>=3 (-25%), qsa_topk2/qsa_score (-5..9%) -> 40.03 tok/s decode & 445.79 tok/s 10.18K prefill

PLAN.md: record Round 17 benchmark results (39.56 tok/s warm decode, 218.99 ms MTP)

Round 17 blitz: mtp.metal float4 vectorization + Q8_0 gemv_catk, barrier-free simd_shuffle_down in gr_up_f/gr_up3/qsa_merge, gdn_step ib==0 readout pruning + qsa_attn char4 K loads (39.56 tok/s warm decode)

PLAN.md: record Round 16 benchmark results (2,094 dispatches, 1.12 GB scratch, 39.1 tok/s warm decode)

Round 16 blitz: fused QSA Wk+Wv+Wik(+Wiq) single GEMM (-24..-36 dispatches, -5.6 MB scratch), decode resident-MoE + shared-expert wave zip (-48 barriers), gemv TAB>1 guards, ple_moe_gate_norm_c SIMD compaction + ple_conv half4/t>=9

PLAN.md: record Round 15 benchmark results (443.43 tok/s on 10.18K Albedo prompt, 116.58 ms qsa_flash x12)

Round 15 blitz: prompt_chunk router+plan/shexp wave zipping (-190 barriers), fused Wal+Wbe N=96 GEMM (-36 dispatches), QSA PSO pre-warm (-202 ms Chunk-1 stall), qsa_score_mm causal-future skip (-19.4%), qsa_prep 320B TG + 2 barriers

PLAN.md: record Round 14 benchmark results (443.22 tok/s on 10.18K Albedo prompt, 501.5 tok/s on Chunk 1)

Round 14 blitz: qsa_flash/qsa_attn_mm byte-offset + zero-scale hoist (-9.7% at pos0>=2048, 443.2 tok/s on 10.18K Albedo prompt), gdn_scanr_4 hk-locality, compacted moe_comb roff, mm_gu3 IQ2_XS half4 HS store

PLAN.md: record Round 13 benchmark results (600.6 tok/s GPU kprof, 126.8 tok/s coding prefill)

Round 13 blitz: gemm_hc_down + gemm_swiglu_up K4 register epilogues (-144 dispatches, -985 MB DRAM/chunk, 600.6 tok/s GPU kprof), dq_f16 exact-simdgroup 2D grid, SpecGate + mtp_draft PLE prefetch

Add audits/opus_colibri_scout.md: Colibri vs Strata Metal audit + M5 hazard/wait/entropy benchmarks

PLAN.md: record Round 12 benchmark results

Round 12 blitz: qsa_topk 3.85x (cached keys + simdgroup scan + paired emit), gemm_hc_up_reg_32_128_4 + unrolled epilogue (-6.9%), Q6_K hoisted mm_gu3 (-4.3%), 3,375.4 ms GPU kprof (587.5 tok/s GPU)

PLAN.md: record Round 11 benchmark results & 10.18K Albedo prompt speedup

Round 11 blitz: gemm_f16 tile selector (-18% / -234 ms), qsa_attn_mm hoisted rv/rk (-44.5%), branchless W8<T_IQ3_S> (-50.8% / 2.03x) + W8<T_IQ2_XS> k_hoist (-57 ms), 3,397.5 ms GPU kprof (583.7 tok/s GPU) & 2048 prompt chunk on :8080

PLAN.md: record Round 10 benchmark results

Round 10 blitz: qsa_flash 2.2x (-51.9% / -146.1 ms in kprof via unconditional hoisted rv/rk + vec4 epilogue), gdn_pre float4 conv, uint4 mm_gather_h + s_id-free mm_place

PLAN.md: record Round 9 benchmark results

Round 9 blitz: qsa_flash 19.9->7.6 KiB TG memory (4 TGs/core, -17.5% / -51.4 ms), TAB>1 barrier guards, IQ3_S dead sign load removal (392.02 tok/s 2K prefill, 40.11 tok/s decode)

PLAN.md: record Round 8 benchmark results

Round 8 blitz: prompt_chunk sigs/cuts shared-expert+MoE-plan overlap, W8 IQ2_XS/IQ2_XXS opt, gr_norm_h_stream_out, dq_f16+act_f16 half4 (385.1 tok/s 2K prefill, 122.9 tok/s short prefill)

Round 7 blitz: single-pass simd_broadcast gdn_scanr_4 (-22.2%), register gemm_hc_up (-8.9%), ple_hist_adv + 1.13 GB scratch (382.0 tok/s prefill, 40.4 tok/s decode)

Round 6 blitz: qsa_flash -20%, ple_moe_gate_norm_c + gated elim, -286 MB PLE/bo/yr aliasing (1.29 GB scratch), barrier-free gdn_pre/post, gr_norm_h* rv (366.8 tok/s prefill)

Update PLAN.md with Round 5 benchmarks (352.1 tok/s prefill, 1.58 GB scratch)