chore(squad): record native-path placement capture-compat finding

[justinchuby]

Decision inbox drop for the Scribe recording item-3 findings. See #1299. Native path only; #982 untouched.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.70%. Comparing base (2abbbdb) to head (8b3ad44).
⚠️ Report is 46 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff            @@
##           main    #1300       +/-   ##
=========================================
+ Coverage      0   80.70%   +80.70%     
=========================================
  Files         0      364      +364     
  Lines         0   160728   +160728     
  Branches      0   160728   +160728     
=========================================
+ Hits          0   129710   +129710     
- Misses        0    26361    +26361     
- Partials      0     4657     +4657     

Flag	Coverage Δ
mlas	85.22% <ø> (?)
offline	80.61% <ø> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.
see 364 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 Change 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
⚠️	sampling_latency/top_p_per_token	439.01 µs	531.19 µs	+21.0%
⚠️	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	4.12 ms	4.98 ms	+20.9%
⚠️	grammar_masking/llguidance_compute_mask/32	94.32 µs	111.90 µs	+18.6%
⚠️	kv_cache/alloc_dealloc_pages	44.66 µs	52.85 µs	+18.3%
⚠️	qwen3_sampling_processors/top_k_full_sort_baseline	2.44 ms	2.88 ms	+17.9%
⚠️	sampling_latency/min_p_per_token	241.90 µs	279.67 µs	+15.6%
✅	gather/large_f16_threads=1-internal/131072	16.99 µs	19.41 µs	+14.2%
✅	logit_processing/seven_processor_chain_per_step	372.57 µs	420.23 µs	+12.8%
✅	qwen3_sampling_processors/top_k_top_p_fast	751.00 µs	845.42 µs	+12.6%
✅	sampling_latency/top_k_per_token	63.49 µs	70.68 µs	+11.3%
✅	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	7.28 ms	8.09 ms	+11.1%
✅	qwen3_sampling_processors/top_k_partial_selection	162.56 µs	180.49 µs	+11.0%
✅	gather/medium_f32_threads=1-internal/32768	4.53 µs	4.91 µs	+8.4%
✅	matmul/medium_generic_f16_threads=8/32x512x512	47.35 µs	51.24 µs	+8.2%
✅	qwen3_sampling_processors/top_p_fast_after_top_k	586.57 µs	628.44 µs	+7.1%
✅	reduce_mean/small_f32_threads=1-internal/4096	16.80 µs	17.78 µs	+5.8%
✅	add/large_bf16_threads=1-internal/4194304	1.86 ms	1.96 ms	+5.4%
✅	matmul/small_generic_f16_threads=8/1x256x256	40.08 µs	42.16 µs	+5.2%
✅	add/small_f16_threads=1-internal/1024	519.3 ns	543.0 ns	+4.6%
✅	gather/large_bf16_threads=1-internal/131072	18.47 µs	19.19 µs	+3.9%
✅	add/medium_f16_threads=1-internal/262144	118.61 µs	121.55 µs	+2.5%
✅	matmul/medium_generic_bf16_threads=8/32x512x512	660.49 µs	675.94 µs	+2.3%
✅	matmul/large_generic_f32_threads=8/32x1024x1024	7.09 ms	7.24 ms	+2.2%
✅	matmul/large_generic_bf16_threads=1/32x1024x1024	2.40 ms	2.45 ms	+2.1%
✅	gather/large_f32_threads=1-internal/131072	52.28 µs	53.25 µs	+1.8%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.73 ms	2.78 ms	+1.8%
✅	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	139.74 µs	142.19 µs	+1.8%
✅	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	1.37 ms	1.39 ms	+1.5%
✅	reduce_mean/medium_f32_threads=1-internal/65536	277.86 µs	281.89 µs	+1.5%
✅	matmul/large_generic_f32_threads=1/32x1024x1024	11.04 ms	11.19 ms	+1.4%
✅	gather/medium_f16_threads=1-internal/32768	2.81 µs	2.84 µs	+1.2%
✅	gather/small_f32_threads=1-internal/4096	757.6 ns	765.6 ns	+1.1%
✅	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	588.09 µs	593.02 µs	+0.8%
✅	tokenization/decode_tokens_per_second	7.54 ms	7.60 ms	+0.8%
✅	matmul/small_generic_f16_threads=1/1x256x256	39.11 µs	39.40 µs	+0.7%
✅	matmul/large_generic_f16_threads=8/32x1024x1024	120.87 µs	121.73 µs	+0.7%
✅	tokenization/encode_tokens_per_second	445.81 µs	447.93 µs	+0.5%
✅	matmul/small_generic_bf16_threads=8/1x256x256	41.73 µs	41.85 µs	+0.3%
✅	matmul/medium_generic_f32_threads=8/32x512x512	1.85 ms	1.85 ms	+0.2%
✅	gather/small_bf16_threads=1-internal/4096	535.2 ns	534.3 ns	-0.2%
✅	matmul/small_generic_f32_threads=1/1x256x256	48.20 µs	48.08 µs	-0.3%
✅	matmul/medium_generic_f16_threads=1/32x512x512	40.40 µs	40.21 µs	-0.5%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	97.05 µs	96.54 µs	-0.5%
✅	add/large_f16_threads=1-internal/4194304	1.89 ms	1.88 ms	-0.6%
✅	matmul/large_generic_bf16_threads=8/32x1024x1024	2.40 ms	2.37 ms	-1.1%
✅	add/small_f32_threads=1-internal/1024	239.5 ns	236.5 ns	-1.3%
✅	gather/small_f16_threads=1-internal/4096	557.6 ns	549.2 ns	-1.5%
✅	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	120.54 µs	118.69 µs	-1.5%
✅	add/large_f32_threads=1-internal/4194304	971.95 µs	955.44 µs	-1.7%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	660.79 µs	647.35 µs	-2.0%
✅	matmul/small_generic_bf16_threads=1/1x256x256	40.32 µs	39.45 µs	-2.2%
✅	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	121.74 µs	118.50 µs	-2.7%
✅	gather/medium_bf16_threads=1-internal/32768	2.94 µs	2.86 µs	-2.7%
✅	sampling_latency/greedy_per_token	3.72 µs	3.59 µs	-3.6%
✅	matmul/small_generic_f32_threads=8/1x256x256	59.04 µs	56.27 µs	-4.7%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.19 ms	1.12 ms	-5.8%
✅	add/medium_bf16_threads=1-internal/262144	128.21 µs	116.91 µs	-8.8%
🟢	add/small_bf16_threads=1-internal/1024	632.8 ns	512.5 ns	-19.0%
🟢	add/medium_f32_threads=1-internal/262144	36.39 µs	28.63 µs	-21.3%

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: { 5.28 4.61 5.70 }

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)