chore(decisions): consolidate GAP-3 native-paged-decode decisions (#565-#568)

[justinchuby]

Consolidates the accumulated GAP-3 native-paged-decode decision inbox (7 notes) into the canonical .squad/decisions.md, clearing the drop-box.

Folds in the decisions from the 4 merged GAP-3 PRs plus the Scan-capture deferral:

• feat(pipeline): wire pure-native multi-component decode (GAP-3 Inc-A) #565 Inc-A — native multi-component pipeline construction
• feat(pipeline): native present-KV mirroring for paged native decode (GAP-3 Inc-C) #566 Inc-C — host present-KV mirroring → paged native decode
• feat(engine): device-resident present-KV read-out for paged native CUDA decode (GAP-3 Inc-D) #567 Inc-D — device-resident f32 CUDA present-KV read-out → paged
• feat(engine): f16 device present-KV read-out for paged native CUDA decode (GAP-3 Inc-D.1) #568 Inc-D.1 — f16 device present-KV read-out → paged
• Inc-C test-rigor fix (byte-equality oracle pattern)
• Scan-capture slice-1b deferral (needs per-EP device-graph handle-keyed registry; resumption trigger recorded)
• Scan slice-1a (feat(engine): flag-gated single-trip Scan inline dual-path (27B Scan-capture slice 1a) #564)

decisions.md 20465 → 28227 bytes (under the 30KB healthy threshold; no archiving needed). Inbox now README-only. Docs-only change under .squad/.

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 81.32%. Comparing base (c1a2450) to head (86d9c75).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #569      +/-   ##
==========================================
+ Coverage   80.60%   81.32%   +0.72%     
==========================================
  Files         315      315              
  Lines      123574   123574              
  Branches   123574   123574              
==========================================
+ Hits        99606   100501     +895     
+ Misses      19916    19015     -901     
- Partials     4052     4058       +6     

Flag	Coverage Δ
cli-ort-linux	83.27% <ø> (ø)
cli-ort-windows	82.67% <ø> (-0.11%)	⬇️
mlas	78.72% <ø> (+0.81%)	⬆️
offline	81.28% <ø> (+0.75%)	⬆️

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]

🔴 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	631.13 µs	1.36 ms	+116.0%
🔴	matmul/large_generic_f32_threads=1/32x1024x1024	10.64 ms	16.69 ms	+56.9%
🔴	matmul/medium_generic_f16_threads=8/32x512x512	49.80 µs	73.53 µs	+47.7%
🔴	logit_processing/seven_processor_chain_per_step	1.34 ms	1.84 ms	+37.7%
🔴	matmul/large_generic_f16_threads=8/32x1024x1024	110.71 µs	150.40 µs	+35.8%
🔴	matmul/medium_generic_bf16_threads=1/32x512x512	591.66 µs	801.43 µs	+35.5%
🔴	sampling_latency/top_p_per_token	1.09 ms	1.44 ms	+31.9%
🔴	matmul/large_generic_bf16_threads=1/32x1024x1024	2.36 ms	3.09 ms	+31.3%
⚠️	matmul/large_generic_f16_threads=1/32x1024x1024	98.78 µs	128.32 µs	+29.9%
⚠️	grammar_masking/llguidance_compute_mask/32	84.84 µs	107.70 µs	+26.9%
⚠️	tokenization/encode_tokens_per_second	440.29 µs	557.57 µs	+26.6%
⚠️	tokenization/decode_tokens_per_second	6.94 ms	8.71 ms	+25.6%
⚠️	sampling_latency/top_k_per_token	550.94 µs	681.97 µs	+23.8%
⚠️	kv_cache/alloc_dealloc_pages	40.51 µs	49.62 µs	+22.5%
⚠️	sampling_latency/greedy_per_token	3.55 µs	4.27 µs	+20.3%
⚠️	sampling_latency/min_p_per_token	379.56 µs	447.11 µs	+17.8%
✅	matmul/large_generic_f32_threads=8/32x1024x1024	6.18 ms	6.93 ms	+12.1%
✅	matmul/small_generic_bf16_threads=8/1x256x256	39.17 µs	43.23 µs	+10.4%
✅	add/large_f32_threads=1-internal/4194304	45.80 ms	48.43 ms	+5.7%
✅	matmul/medium_generic_f16_threads=1/32x512x512	40.07 µs	41.25 µs	+2.9%
✅	matmul/large_generic_bf16_threads=8/32x1024x1024	2.19 ms	2.23 ms	+2.1%
✅	add/medium_f16_threads=1-internal/262144	2.88 ms	2.93 ms	+1.9%
✅	add/small_f32_threads=1-internal/1024	215.3 ns	216.8 ns	+0.7%
✅	gather/small_f16_threads=1-internal/4096	545.2 ns	549.0 ns	+0.7%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.73 ms	2.74 ms	+0.3%
✅	gather/large_f16_threads=1-internal/131072	19.75 µs	19.77 µs	+0.1%
✅	matmul/small_generic_f16_threads=1/1x256x256	40.47 µs	40.50 µs	+0.1%
✅	add/small_f16_threads=1-internal/1024	14.25 µs	14.19 µs	-0.4%
✅	matmul/small_generic_f32_threads=1/1x256x256	51.41 µs	51.16 µs	-0.5%
✅	add/large_bf16_threads=1-internal/4194304	48.40 ms	47.70 ms	-1.5%
✅	add/medium_bf16_threads=1-internal/262144	3.05 ms	2.98 ms	-2.1%
✅	add/small_bf16_threads=1-internal/1024	14.83 µs	14.35 µs	-3.3%
✅	add/medium_f32_threads=1-internal/262144	2.95 ms	2.85 ms	-3.3%
✅	add/large_f16_threads=1-internal/4194304	49.57 ms	47.31 ms	-4.6%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.18 ms	1.10 ms	-6.2%
✅	matmul/medium_generic_f32_threads=8/32x512x512	1.76 ms	1.65 ms	-6.3%
✅	gather/small_bf16_threads=1-internal/4096	582.9 ns	545.9 ns	-6.4%
✅	reduce_mean/medium_f32_threads=1-internal/65536	292.42 µs	269.52 µs	-7.8%
✅	gather/medium_bf16_threads=1-internal/32768	3.15 µs	2.90 µs	-7.9%
✅	matmul/small_generic_bf16_threads=1/1x256x256	43.94 µs	40.12 µs	-8.7%
✅	gather/large_bf16_threads=1-internal/131072	19.17 µs	17.46 µs	-8.9%
✅	matmul/small_generic_f16_threads=8/1x256x256	44.54 µs	40.33 µs	-9.5%
✅	gather/small_f32_threads=1-internal/4096	807.1 ns	721.7 ns	-10.6%
✅	matmul/small_generic_f32_threads=8/1x256x256	62.00 µs	55.41 µs	-10.6%
✅	reduce_mean/small_f32_threads=1-internal/4096	18.96 µs	16.75 µs	-11.6%
🟢	gather/medium_f32_threads=1-internal/32768	5.59 µs	4.60 µs	-17.7%
🟢	gather/medium_f16_threads=1-internal/32768	3.65 µs	2.99 µs	-18.0%
🟢	gather/large_f32_threads=1-internal/131072	63.44 µs	39.35 µs	-38.0%

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 25.4.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 5.72 5.68 7.10 }

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)