chore(squad): qwen3.5-2b native ORT-win + general head-size kernel lever

[justinchuby]

Wallace's different-architecture GPU probe result + coordinator standing-table update.

• qwen3.5-2b (hybrid linear-attn, head_dim=256) native beats ORT in BOTH framings: eager 1.10x (within noise), capture 1.65x (clean, fallbacks=0). ORT CUDA EP can't place LinearAttention/CausalConvWithState on GPU (25 Memcpy nodes, graph blocked); native runs+captures the whole hybrid graph.
• New lever (in flight, Deckard): general head-size fused GQA decode kernel — head_dim=256 falls out of the head<=128 fast path to gqa_attention_reference_f32 = 31% of decode. Generalized (per user directive) over head sizes {64,80,96,112,128,192,256} + asymmetric dual-head-size audit (Gemma/DeepSeek MLA).
• Support gaps: Phi-3.5-mini generic-cpu If op; qwen3.5-9b multimodal harness.

Squad-state only (.squad/decisions.md, identity/now.md, agents/wallace/history.md).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 79.96%. Comparing base (f3c3c22) to head (18d8219).
⚠️ Report is 57 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1269      +/-   ##
==========================================
- Coverage   80.44%   79.96%   -0.48%     
==========================================
  Files         359      361       +2     
  Lines      157898   157076     -822     
  Branches   157898   157076     -822     
==========================================
- Hits       127027   125613    -1414     
- Misses      26255    26829     +574     
- Partials     4616     4634      +18     

Flag	Coverage Δ
mlas	?
offline	79.96% <ø> (-0.38%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 38 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

🔴 Benchmark Regression Detected

Comparison of criterion micro-benchmarks: PR head vs merge-base, measured on the same runner in the same job (base first → PR second).

ℹ️ Absolute times are informational only — they vary with runner load. The % change column is the reliable signal because both sides ran under identical conditions.

Status	Scenario	Base	PR	Change
🔴	grammar_masking/llguidance_compute_mask/32	77.69 µs	128.55 µs	+65.5%
🔴	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	3.67 ms	5.63 ms	+53.3%
🔴	qwen3_sampling_processors/top_k_partial_selection	145.33 µs	214.64 µs	+47.7%
🔴	qwen3_sampling_processors/top_k_full_sort_baseline	2.14 ms	3.01 ms	+40.7%
🔴	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	131.88 µs	180.64 µs	+37.0%
🔴	qwen3_sampling_processors/top_p_fast_after_top_k	525.17 µs	715.91 µs	+36.3%
🔴	logit_processing/seven_processor_chain_per_step	322.26 µs	430.15 µs	+33.5%
⚠️	qwen3_sampling_processors/top_k_top_p_fast	662.88 µs	792.26 µs	+19.5%
⚠️	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	5.83 ms	6.86 ms	+17.6%
✅	kv_cache/alloc_dealloc_pages	39.30 µs	45.19 µs	+15.0%
✅	matmul/large_generic_f16_threads=8/32x1024x1024	109.30 µs	125.23 µs	+14.6%
✅	gather/large_bf16_threads=1-internal/131072	13.72 µs	15.59 µs	+13.6%
✅	matmul/small_generic_f16_threads=1/1x256x256	37.99 µs	42.85 µs	+12.8%
✅	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	607.09 µs	679.63 µs	+11.9%
✅	matmul/medium_generic_f16_threads=1/32x512x512	38.76 µs	42.08 µs	+8.6%
✅	gather/small_f16_threads=1-internal/4096	476.7 ns	516.2 ns	+8.3%
✅	tokenization/encode_tokens_per_second	384.46 µs	408.19 µs	+6.2%
✅	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	102.58 µs	108.54 µs	+5.8%
✅	matmul/small_generic_f32_threads=8/1x256x256	49.13 µs	51.71 µs	+5.3%
✅	matmul/medium_generic_f16_threads=8/32x512x512	43.23 µs	45.09 µs	+4.3%
✅	sampling_latency/greedy_per_token	3.27 µs	3.41 µs	+4.3%
✅	add/medium_f32_threads=1-internal/262144	25.12 µs	26.01 µs	+3.5%
✅	matmul/small_generic_f16_threads=8/1x256x256	37.73 µs	38.95 µs	+3.2%
✅	matmul/medium_generic_f32_threads=8/32x512x512	1.91 ms	1.97 ms	+2.9%
✅	add/medium_f16_threads=1-internal/262144	107.95 µs	111.04 µs	+2.9%
✅	add/large_f16_threads=1-internal/4194304	1.71 ms	1.76 ms	+2.7%
✅	gather/small_bf16_threads=1-internal/4096	477.4 ns	488.8 ns	+2.4%
✅	add/large_f32_threads=1-internal/4194304	862.83 µs	879.76 µs	+2.0%
✅	matmul/large_generic_f32_threads=1/32x1024x1024	9.59 ms	9.76 ms	+1.8%
✅	matmul/large_generic_f32_threads=8/32x1024x1024	7.55 ms	7.68 ms	+1.8%
✅	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	138.11 µs	139.95 µs	+1.3%
✅	matmul/large_generic_bf16_threads=1/32x1024x1024	2.18 ms	2.21 ms	+1.1%
✅	gather/large_f16_threads=1-internal/131072	14.61 µs	14.76 µs	+1.0%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	86.45 µs	87.29 µs	+1.0%
✅	matmul/small_generic_bf16_threads=1/1x256x256	37.60 µs	37.95 µs	+0.9%
✅	gather/small_f32_threads=1-internal/4096	670.7 ns	675.2 ns	+0.7%
✅	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	1.29 ms	1.30 ms	+0.6%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.35 ms	2.35 ms	+0.1%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	547.71 µs	547.90 µs	+0.0%
✅	sampling_latency/min_p_per_token	209.16 µs	207.21 µs	-0.9%
✅	reduce_mean/medium_f32_threads=1-internal/65536	252.60 µs	248.76 µs	-1.5%
✅	sampling_latency/top_k_per_token	53.35 µs	52.29 µs	-2.0%
✅	gather/medium_f32_threads=1-internal/32768	4.24 µs	4.14 µs	-2.4%
✅	matmul/large_generic_bf16_threads=8/32x1024x1024	2.44 ms	2.38 ms	-2.4%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.02 ms	991.45 µs	-2.5%
✅	tokenization/decode_tokens_per_second	6.73 ms	6.55 ms	-2.7%
✅	gather/medium_f16_threads=1-internal/32768	2.52 µs	2.42 µs	-3.6%
✅	matmul/small_generic_bf16_threads=8/1x256x256	39.53 µs	38.06 µs	-3.7%
✅	matmul/small_generic_f32_threads=1/1x256x256	49.03 µs	47.02 µs	-4.1%
✅	sampling_latency/top_p_per_token	390.08 µs	372.99 µs	-4.4%
✅	add/small_f16_threads=1-internal/1024	469.6 ns	448.8 ns	-4.4%
✅	matmul/medium_generic_bf16_threads=8/32x512x512	678.85 µs	648.07 µs	-4.5%
✅	add/medium_bf16_threads=1-internal/262144	110.99 µs	104.83 µs	-5.6%
✅	add/small_bf16_threads=1-internal/1024	468.5 ns	442.0 ns	-5.7%
✅	gather/medium_bf16_threads=1-internal/32768	2.54 µs	2.39 µs	-5.9%
✅	gather/large_f32_threads=1-internal/131072	40.92 µs	37.35 µs	-8.7%
✅	reduce_mean/small_f32_threads=1-internal/4096	16.86 µs	15.33 µs	-9.1%
✅	add/small_f32_threads=1-internal/1024	226.8 ns	198.3 ns	-12.6%
🟢	add/large_bf16_threads=1-internal/4194304	2.03 ms	1.72 ms	-15.0%

Visual flags: ⚠️ ≥ 15% slower, 🔴 ≥ 30% slower — calibrated against measured runner noise (~27% worst-case on multi-threaded matmul)

Host info

CPU: Apple M1 (Virtual)
Cores: 3
OS: Darwin 25.5.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 5.51 4.54 5.37 }

What this cannot catch

• Regressions in code paths not covered by these benchmarks (e.g., end-to-end decode with a real model)
• Sub-threshold regressions that compound over multiple PRs
• Performance changes that only manifest under GPU execution
• Latency changes in the ORT integration path (these benchmarks exercise the native Rust kernels)