[justinchuby]

Scribe consolidation for the Marlin int4 GEMM program landing (#960/#961/#962).

• Merge 5 decision inbox drops (deckard/pris/sebastian kernel+bench+numerics + chew/gaff reviews) into one consolidated Marlin program entry; corrected per-model attribution (glm block-128 → gate_up fusion no-op; 2.16× was the general split-K retune).
• Archive resolved docs(research): speculative-capture feasibility — CONDITIONAL-GO gated behind Marlin #957/research(decode): Lever B Phase-0 capture-stability probe — gated-GO on Increment-0 #948/research(decode): Lever B Increment-0 — decisive captured M=K number → NO-GO, promote Marlin #949 spec-capture + Lever B entries to decisions-archive/2026-08.md (decisions.md 45,991 → 42,104 B, under the 50 KB gate).
• Orchestration logs (deckard/sebastian/pris/chew/gaff), session log, cross-agent history updates; summarized sebastian history (>8 chronicle threshold).

State-only change. Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 78.99%. Comparing base (df6d3af) to head (68256ac).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #972      +/-   ##
==========================================
- Coverage   79.63%   78.99%   -0.64%     
==========================================
  Files         356      356              
  Lines      150211   150211              
  Branches   150211   150211              
==========================================
- Hits       119620   118659     -961     
- Misses      25935    26899     +964     
+ Partials     4656     4653       -3     

Flag	Coverage Δ
cli-ort-linux	83.52% <ø> (ø)
cli-ort-windows	83.11% <ø> (ø)
mlas	80.87% <ø> (-0.10%)	⬇️
offline	78.79% <ø> (-0.68%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 7 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]

✅ Benchmarks — No Regression

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
✅	logit_processing/seven_processor_chain_per_step	297.77 µs	330.61 µs	+11.0%
✅	kv_cache/alloc_dealloc_pages	36.81 µs	40.27 µs	+9.4%
✅	grammar_masking/llguidance_compute_mask/32	71.44 µs	78.01 µs	+9.2%
✅	gather/medium_f32_threads=1-internal/32768	3.62 µs	3.89 µs	+7.6%
✅	qwen3_sampling_processors/top_k_top_p_fast	632.74 µs	673.31 µs	+6.4%
✅	qwen3_sampling_processors/top_k_partial_selection	129.82 µs	137.17 µs	+5.7%
✅	sampling_latency/min_p_per_token	191.52 µs	194.43 µs	+1.5%
✅	qwen3_sampling_processors/top_p_fast_after_top_k	487.35 µs	490.68 µs	+0.7%
✅	sampling_latency/top_p_per_token	351.51 µs	352.54 µs	+0.3%
✅	tokenization/encode_tokens_per_second	351.13 µs	351.32 µs	+0.1%
✅	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	5.79 ms	5.79 ms	+0.0%
✅	qwen3_sampling_processors/top_k_full_sort_baseline	1.97 ms	1.95 ms	-1.2%
✅	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	3.28 ms	3.23 ms	-1.7%
✅	matmul/small_generic_bf16_threads=8/1x256x256	31.39 µs	29.54 µs	-5.9%
✅	gather/medium_bf16_threads=1-internal/32768	2.40 µs	2.24 µs	-6.9%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	527.77 µs	490.21 µs	-7.1%
✅	sampling_latency/top_k_per_token	52.10 µs	48.32 µs	-7.2%
✅	matmul/small_generic_f16_threads=8/1x256x256	31.72 µs	29.24 µs	-7.8%
✅	gather/small_bf16_threads=1-internal/4096	478.5 ns	441.0 ns	-7.8%
✅	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	79.58 µs	72.51 µs	-8.9%
✅	reduce_mean/medium_f32_threads=1-internal/65536	253.15 µs	226.98 µs	-10.3%
✅	add/large_bf16_threads=1-internal/4194304	48.68 ms	43.62 ms	-10.4%
✅	matmul/medium_generic_f16_threads=8/32x512x512	31.87 µs	28.55 µs	-10.4%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.40 ms	2.15 ms	-10.5%
✅	gather/small_f32_threads=1-internal/4096	695.0 ns	615.2 ns	-11.5%
✅	add/large_f16_threads=1-internal/4194304	44.76 ms	39.44 ms	-11.9%
✅	matmul/medium_generic_bf16_threads=8/32x512x512	410.74 µs	360.50 µs	-12.2%
✅	add/large_f32_threads=1-internal/4194304	43.03 ms	37.49 ms	-12.9%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.05 ms	912.64 µs	-13.3%
✅	sampling_latency/greedy_per_token	3.46 µs	2.99 µs	-13.6%
✅	reduce_mean/small_f32_threads=1-internal/4096	18.34 µs	15.83 µs	-13.7%
✅	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	45.19 µs	39.00 µs	-13.7%
✅	gather/small_f16_threads=1-internal/4096	519.0 ns	445.8 ns	-14.1%
✅	add/small_bf16_threads=1-internal/1024	15.06 µs	12.81 µs	-14.9%
🟢	tokenization/decode_tokens_per_second	6.74 ms	5.72 ms	-15.1%
🟢	matmul/small_generic_f32_threads=1/1x256x256	40.18 µs	33.90 µs	-15.6%
🟢	add/medium_f32_threads=1-internal/262144	2.80 ms	2.35 ms	-15.9%
🟢	add/medium_f16_threads=1-internal/262144	2.88 ms	2.41 ms	-16.2%
🟢	matmul/large_generic_bf16_threads=1/32x1024x1024	2.18 ms	1.82 ms	-16.2%
🟢	matmul/small_generic_bf16_threads=1/1x256x256	35.04 µs	28.96 µs	-17.4%
🟢	matmul/medium_generic_f32_threads=8/32x512x512	1.09 ms	897.93 µs	-17.8%
🟢	matmul/small_generic_f32_threads=8/1x256x256	41.45 µs	34.01 µs	-18.0%
🟢	add/medium_bf16_threads=1-internal/262144	3.02 ms	2.44 ms	-19.1%
🟢	matmul/medium_generic_f16_threads=1/32x512x512	34.02 µs	27.46 µs	-19.3%
🟢	matmul/small_generic_f16_threads=1/1x256x256	35.04 µs	27.97 µs	-20.2%
🟢	gather/medium_f16_threads=1-internal/32768	2.79 µs	2.22 µs	-20.2%
🟢	add/small_f16_threads=1-internal/1024	15.75 µs	12.14 µs	-22.9%
🟢	matmul/large_generic_f16_threads=8/32x1024x1024	102.23 µs	77.32 µs	-24.4%
🟢	matmul/large_generic_bf16_threads=8/32x1024x1024	1.69 ms	1.24 ms	-26.4%
🟢	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	475.71 µs	345.28 µs	-27.4%
🟢	matmul/large_generic_f32_threads=1/32x1024x1024	11.92 ms	8.65 ms	-27.4%
🟢	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	752.33 µs	499.02 µs	-33.7%
🟢	gather/large_f32_threads=1-internal/131072	37.57 µs	24.86 µs	-33.8%
🟢	matmul/large_generic_f16_threads=1/32x1024x1024	111.35 µs	72.63 µs	-34.8%
🟢	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	66.98 µs	42.43 µs	-36.7%
🟢	gather/large_f16_threads=1-internal/131072	17.97 µs	10.48 µs	-41.7%
🟢	matmul/large_generic_f32_threads=8/32x1024x1024	6.68 ms	3.58 ms	-46.5%
🟢	gather/large_bf16_threads=1-internal/131072	20.59 µs	10.43 µs	-49.4%
🟢	add/small_f32_threads=1-internal/1024	474.6 ns	184.5 ns	-61.1%

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: { 2.89 3.55 5.04 }

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)