chore(squad): scribe round 8 — async-offload/correctness wave consolidation

[justinchuby]

Consolidates 6 inbox decision notes into decisions.md (20458->20332 B, under the 20480 gate) and archives two historical wave records. Updates agent histories.

Wave summary (all merged):

• feat(cuda-offload): async fence-ordered weight page-in overlap (#87 first increment) #544 — async fence-ordered CUDA weight page-in (Implement compute-transfer overlap for weight paging (Phase 4 prefetch) #87 increment-1) + deterministic anti-regression test
• profile_native: resolve genai_config decoder io so capture counters print (#87 capture-observability) #552 — profile_native capture-counter observability for genai_config decoders
• fix(native-cuda): re-zero recurrent/conv state on session reuse (LinearAttention corruption) #554 — native-CUDA session-reuse recurrent-state reset fix (closes Native CUDA decode corrupts generation #2+ in a reused session (per-session state not reset) #553); 27B LinearAttention gen#2+ corruption
• 27B native offload A/B proof: 2.9x VRAM reduction, byte-exact output

State-only change (decisions/histories/archive). Logs are gitignored.

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 81.16%. Comparing base (fa1afed) to head (b4591f8).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #555      +/-   ##
==========================================
+ Coverage   80.59%   81.16%   +0.57%     
==========================================
  Files         315      315              
  Lines      123446   123446              
  Branches   123446   123446              
==========================================
+ Hits        99489   100198     +709     
+ Misses      19907    19191     -716     
- Partials     4050     4057       +7     

Flag	Coverage Δ
cli-ort-linux	83.27% <ø> (ø)
cli-ort-windows	82.67% <ø> (-0.11%)	⬇️
mlas	77.91% <ø> (ø)
offline	81.12% <ø> (+0.60%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 4 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
🔴	matmul/medium_generic_bf16_threads=8/32x512x512	377.65 µs	675.98 µs	+79.0%
🔴	matmul/large_generic_bf16_threads=8/32x1024x1024	1.35 ms	2.38 ms	+76.7%
🔴	matmul/medium_generic_f32_threads=8/32x512x512	1.04 ms	1.79 ms	+71.5%
🔴	gather/large_f32_threads=1-internal/131072	30.19 µs	48.17 µs	+59.6%
🔴	matmul/large_generic_f32_threads=8/32x1024x1024	4.62 ms	6.86 ms	+48.5%
🔴	matmul/medium_generic_f16_threads=8/32x512x512	30.47 µs	44.59 µs	+46.3%
🔴	tokenization/decode_tokens_per_second	6.51 ms	8.85 ms	+35.9%
🔴	matmul/large_generic_f16_threads=8/32x1024x1024	78.95 µs	105.24 µs	+33.3%
⚠️	matmul/small_generic_bf16_threads=1/1x256x256	33.26 µs	43.19 µs	+29.9%
⚠️	matmul/small_generic_f32_threads=8/1x256x256	39.79 µs	50.85 µs	+27.8%
⚠️	gather/large_f16_threads=1-internal/131072	12.23 µs	15.40 µs	+25.9%
⚠️	matmul/small_generic_bf16_threads=8/1x256x256	33.42 µs	39.16 µs	+17.2%
⚠️	gather/large_bf16_threads=1-internal/131072	12.87 µs	14.98 µs	+16.3%
⚠️	matmul/medium_generic_f16_threads=1/32x512x512	30.61 µs	35.41 µs	+15.7%
✅	matmul/small_generic_f16_threads=8/1x256x256	32.67 µs	37.15 µs	+13.7%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	77.42 µs	87.84 µs	+13.5%
✅	matmul/small_generic_f16_threads=1/1x256x256	31.82 µs	35.89 µs	+12.8%
✅	kv_cache/alloc_dealloc_pages	35.43 µs	39.57 µs	+11.7%
✅	grammar_masking/llguidance_compute_mask/32	68.95 µs	76.93 µs	+11.6%
✅	sampling_latency/min_p_per_token	312.70 µs	348.62 µs	+11.5%
✅	sampling_latency/greedy_per_token	3.00 µs	3.31 µs	+10.4%
✅	logit_processing/seven_processor_chain_per_step	1.10 ms	1.20 ms	+8.7%
✅	sampling_latency/top_p_per_token	919.06 µs	995.65 µs	+8.3%
✅	add/medium_f32_threads=1-internal/262144	2.65 ms	2.86 ms	+8.0%
✅	matmul/large_generic_bf16_threads=1/32x1024x1024	1.95 ms	2.09 ms	+7.3%
✅	add/medium_f16_threads=1-internal/262144	2.76 ms	2.94 ms	+6.6%
✅	matmul/small_generic_f32_threads=1/1x256x256	40.57 µs	43.23 µs	+6.6%
✅	sampling_latency/top_k_per_token	472.13 µs	498.92 µs	+5.7%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	530.62 µs	547.44 µs	+3.2%
✅	matmul/large_generic_f32_threads=1/32x1024x1024	9.71 ms	9.77 ms	+0.6%
✅	add/small_bf16_threads=1-internal/1024	13.17 µs	12.81 µs	-2.7%
✅	gather/medium_f16_threads=1-internal/32768	2.74 µs	2.64 µs	-3.9%
✅	add/small_f16_threads=1-internal/1024	12.63 µs	12.05 µs	-4.5%
✅	tokenization/encode_tokens_per_second	416.90 µs	396.19 µs	-5.0%
✅	add/medium_bf16_threads=1-internal/262144	3.17 ms	2.92 ms	-8.0%
✅	add/large_f32_threads=1-internal/4194304	47.58 ms	42.49 ms	-10.7%
🟢	gather/small_f16_threads=1-internal/4096	569.1 ns	482.6 ns	-15.2%
🟢	matmul/medium_generic_f32_threads=1/32x512x512	2.82 ms	2.38 ms	-15.5%
🟢	gather/small_bf16_threads=1-internal/4096	594.0 ns	489.9 ns	-17.5%
🟢	add/small_f32_threads=1-internal/1024	226.3 ns	181.7 ns	-19.7%
🟢	gather/medium_bf16_threads=1-internal/32768	3.21 µs	2.56 µs	-20.3%
🟢	gather/small_f32_threads=1-internal/4096	876.4 ns	680.0 ns	-22.4%
🟢	add/large_f16_threads=1-internal/4194304	56.27 ms	43.10 ms	-23.4%
🟢	reduce_mean/large_f32_threads=1-internal/262144	1.49 ms	1.11 ms	-25.6%
🟢	reduce_mean/small_f32_threads=1-internal/4096	23.56 µs	16.66 µs	-29.3%
🟢	add/large_bf16_threads=1-internal/4194304	66.15 ms	44.36 ms	-32.9%
🟢	gather/medium_f32_threads=1-internal/32768	6.78 µs	4.41 µs	-34.9%
🟢	reduce_mean/medium_f32_threads=1-internal/65536	419.98 µs	267.91 µs	-36.2%

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 25.4.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 3.24 4.28 7.74 }

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)