[{"data":1,"prerenderedAt":291},["ShallowReactive",2],{"finding:p1-axis1-metal-opt-audit":3,"finding-runs:p1-axis1-metal-opt-audit":199,"finding-related:p1-axis1-metal-opt-audit":200},{"meta":4,"impact":31,"sections":37},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":15,"task_code":24,"related_runs":25,"related_findings":26},"p1-axis1-metal-opt-audit","P1 axis 1 Metal 最適化候補 audit — 5 候補 + 既実装 gate flip 機会、Tier 1 -1.0% wallclock 即時 actionable","Phase D 完遂後 (8 scene mean 33.49 dB \u002F brush +0.63 dB) の clean profile baseline 5 候補 (emit_pairs 14.2% \u002F radix_sort 13.6% \u002F backward 13.4% \u002F ssim 7.8% \u002F forward 6.1%) を kernel 実装まで降りて構造的 audit。**最大発見: emit_pairs_simd PSO は既に実装済、Cell::new(false) gate off で `30k validation 後 true 化予定` comment、Phase D 完遂で即 flip 候補** (-0.7-1.0% wallclock、zero risk)。Tier 1+2 累計 -1.5-2.5% wallclock 改善余地、Tier 3 imageblock+TBDR で +1.0% 追加。backward SIMD は既に default 有効 (2.43× 享受中)、f16 forward は env gate (disabled)。","P1 axis 1 · Metal 最適化 audit · 5 候補 + gate flip 機会","2026-05-25","stable","design","positive",[14],1,[16,17,18,19,20,21,22,23],"p1-profile","axis-1","metal-optimization","kernel-audit","tbdr","simd-reduction","apple-silicon","gate-flip","P1 axis 1 Metal kernel audit",[],[27,28,29,30],"p1-profiling-clean","p1-axis1-target-cache","p1-e-refine-gpu-smoke","p1-d-multi-scene-rechain",{"summary":32,"rank":33,"verdict":34,"delta_psnr":35,"delta_wallclock":36},"5 kernel (clean baseline share 合計 55.1%) を Explore subagent で構造的 audit、Apple Silicon 特化最適化候補を kernel 単位で 2-4 個ずつ抽出。\u003Cstrong>最大の発見\u003C\u002Fstrong>: **emit_pairs_simd PSO は既に実装済**、\u003Ccode>use_simd_emit: Cell::new(false)\u003C\u002Fcode> で gate off、comment に「30k validation 後 default true 化予定」(tile_bin.rs:86-87)、**Phase D 30k 完遂で即 flip 可能** (-0.7-1.0% wallclock 即時、zero risk)。同様の即 actionable 機会: f16 forward kernel \u003Ccode>render_splats_f16\u003C\u002Fcode> は env \u003Ccode>SPLAT_F16_FORWARD=1\u003C\u002Fcode> gate (現在 disabled、A\u002FB test 必須 PSNR risk MED-HIGH)。\u003Cstrong>Tier 2 (Phase E scope)\u003C\u002Fstrong>: radix_sort GPU prefix sum (-0.54-0.82% wallclock、CPU-GPU 16-pass sync 除去)、backward_raster imageblock+TBDR (-0.67-1.07% wallclock、tile-local 累積で atomic 大幅削減)。\u003Cstrong>累計 -1.5-2.5% wallclock 改善余地確定、卒論 §6 future work 候補と pilot 実装目標\u003C\u002Fstrong>。backward SIMD reduction は既に default 有効 (rasterize.rs:642、2.43× win 享受中で確認済)、SSIM は eval-only で training 直接寄与なしのため Tier 3。","high","design-complete-actionable","N\u002FA (audit)","estimated -1.5 〜 -2.5% (Tier 1+2)",[38,41,46,49,114,116,123,125,129,131,136,138,143,145,182,184,190,192],{"type":39,"text":40},"lead","Phase D 完遂 (8 scene mean 33.49 dB \u002F brush +0.63 dB \u002F wallclock -61%) 後、clean profile baseline (single-process、p1-profiling-clean) の 5 候補 kernel を \u003Cstrong>Explore subagent で実装まで降りて audit\u003C\u002Fstrong>。最大の発見は \u003Cstrong>emit_pairs_simd PSO が既に実装済、gate off で \u003Ccode>30k validation 後 true 化予定\u003C\u002Fcode> と comment 明記\u003C\u002Fstrong>、Phase D 30k 完遂で \u003Cstrong>即 flip 可能な zero-risk 改善\u003C\u002Fstrong> (-0.7-1.0% wallclock)。Tier 1 (即 actionable gate flip)、Tier 2 (Phase E scope の構造改善)、Tier 3 (低 ROI \u002F 卒論 narrative 価値) の 3 段階で優先順位確定。",{"type":42,"label":43,"variant":44,"text":45},"callout","Headline (即 actionable 発見)","success","\u003Cstrong>tile_bin.rs:86-87 に \u003Ccode>use_simd_emit: Cell::new(false)\u003C\u002Fcode> と「30k validation 後 default true 化予定 (現在 false で慎重に)」comment\u003C\u002Fstrong>。emit_pairs_simd kernel は既に shader (tile_bin.metal:119) + PSO (tile_bin.rs:79) 実装済、SIMD 32 thread で \u003Ccode>simd_prefix_exclusive_sum\u003C\u002Fcode> による atomic 削減 (1M atomic op → ~30k SIMD group atomic、32× 削減)。Phase D 30k validation 完遂で本 gate を true 化する \u003Cstrong>structural prerequisite は満たされた\u003C\u002Fstrong>。PSNR risk LOW (atomic order の permutation のみ、pair count \u002F key 値は不変)、実装 cost 1 行変更 + 5k smoke で gate flip 完了。同様の機会: \u003Ccode>SPLAT_F16_FORWARD=1\u003C\u002Fcode> env で f16 forward kernel 切替 (A\u002FB test で PSNR 検証必須、MED-HIGH risk)。",{"type":47,"text":48},"heading","1. 5 候補 kernel summary (share + bottleneck + 改善上限)",{"type":50,"columns":51,"align":61,"rows":65,"caption":113},"table",[52,53,54,55,56,57,58,59,60],"#","kernel","share","主 bottleneck","Apple 特化候補 (top)","PSNR risk","推定 wallclock %","実装 cost","備考",[62,63,62,63,63,64,62,64,63],"right","left","center",[66,76,85,95,104],[67,68,69,70,71,72,73,74,75],"1","**emit_pairs**","14.2%","atomic counter contention (1M ops\u002Fiter)","SIMD prefix sum reduction (既実装!)","LOW","**-0.7〜-1.0%**","**1 line**","**flag flip 即可、Tier 1**",[77,78,79,80,81,72,82,83,84],"2","ts_fwd_radix_sort","13.6%","16-pass CPU-GPU sync round-trip","GPU prefix sum (no-CPU radix)","-0.54〜-0.82%","4-5h","Phase E scope",[86,87,88,89,90,91,92,93,94],"3","ts_backward_raster","13.4%","per-pixel × per-splat 9 atomic (32M ops\u002Fiter)","**imageblock + TBDR tile-local accumulator**","MED-HIGH","-0.67〜-1.07%","6-8h","TBDR ↔ Kahan summation 必要",[96,97,98,99,100,72,101,102,103],"4","ts_ssim_fwd_grad","7.8%","window read bandwidth + kernel separation","forward+backward kernel fusion","-0.16〜-0.23%","3-4h","eval-only、training 直接寄与なし",[105,106,107,108,109,91,110,111,112],"5","ts_fwd_rasterize","6.1%","global splat batch load + barrier overhead","f16 accumulator (既実装!) \u002F TBDR imageblock","-0.03〜-0.12%","0.5h \u002F 3-4h","f16 は env gate、A\u002FB 必須","全 5 候補で wallclock 推定改善上限を share × per-kernel 改善率で算出。最大 ROI は emit_pairs (gate flip 1 行で -1.0%) と backward imageblock+TBDR (-1.07%、高 cost & 中-高 risk)。SSIM は 7.8% share だが eval-only のため training 効果直接寄与なし。",{"type":47,"text":115},"2. Tier 1: 即 actionable (既実装 code の gate flip)",{"type":117,"ordered":118,"items":119},"list",true,[120,121,122],"\u003Cstrong>emit_pairs_simd (1 行 flip)\u003C\u002Fstrong>: \u003Ccode>tile_bin.rs:87\u003C\u002Fcode> の \u003Ccode>use_simd_emit: Cell::new(false)\u003C\u002Fcode> → \u003Ccode>true\u003C\u002Fcode>、5k smoke で pair count + sort key の bit-exact 確認 + 30k Lego で PSNR ±0.05 dB 検証。期待 \u003Cstrong>-0.7-1.0% wallclock\u003C\u002Fstrong>、zero risk。\u003Cstrong>Phase D 30k 完遂で gate 開放条件達成\u003C\u002Fstrong>","\u003Cstrong>f16 forward (env flip + A\u002FB test)\u003C\u002Fstrong>: \u003Ccode>SPLAT_F16_FORWARD=1\u003C\u002Fcode> で \u003Ccode>render_splats_f16\u003C\u002Fcode> kernel 有効化 (rasterize.rs:147)。5k smoke で PSNR 計測、underflow regime (T \u003C 6e-5) で drift 観察、30k Lego で ±0.05 dB 内なら本採用。期待 \u003Cstrong>-0.03-0.06% wallclock\u003C\u002Fstrong>、risk MED-HIGH (fp16 underflow)、cost 0.5h","\u003Cstrong>backward SIMD reduction (既 default 有効、確認)\u003C\u002Fstrong>: \u003Ccode>BackwardKind::Simd\u003C\u002Fcode> が rasterize.rs:642, 683 で default、2.43× win を Phase D 30k 結果で享受中。No further action、ただし axis 1 narrative の「既存 Apple 最適化」として卒論記述",{"type":47,"text":124},"3. Tier 2: Phase E scope (構造改善、Phase D 30k 後の axis 1 future work)",{"type":117,"ordered":118,"items":126},[127,128],"\u003Cstrong>radix_sort GPU prefix sum\u003C\u002Fstrong>: 現状 16 pass × (CPU prefix scan + GPU dispatch) sync round-trip → all-GPU prefix scan kernel で CPU sync 除去。期待 -0.54-0.82% wallclock、cost 4-5h、PSNR risk LOW (順序保存)。MPS sort 置換 (-1.9-2.2% combined with emit_pairs SIMD) は Tier 4 で別途検討","\u003Cstrong>backward_raster imageblock + TBDR\u003C\u002Fstrong>: 現状 per-pixel × per-splat × 9 atomic (32M ops\u002Fiter)、SIMD reduction で 32× 削減済も atomic 残存。imageblock_data + TBDR で \u003Cstrong>tile-local (16x16 pixel × ~40 splat × 9 grad = ~40KB)\u003C\u002Fstrong> 累積 → tile 境界で global 書き戻し 1 回、threadgroup memory 96KB 制約内に収まる (M4 Max 検証済)。期待 -0.67-1.07% wallclock、cost 6-8h、PSNR risk **MED-HIGH** (Kahan summation 必要、累積順序変更で fp32 drift 観察)",{"type":47,"text":130},"4. Tier 3: 低 ROI \u002F narrative 価値",{"type":117,"items":132},[133,134,135],"\u003Cstrong>SSIM forward+backward kernel fusion\u003C\u002Fstrong>: 7.8% share だが eval-only (training 直接寄与なし)、期待 -0.16-0.23% wallclock、cost 3-4h、PSNR risk LOW。卒論 axis 1 narrative としては A.1 SSIM TBDR の延長","\u003Cstrong>fwd_rasterize imageblock TBDR\u003C\u002Fstrong>: 6.1% share、期待 -0.06-0.12% wallclock、cost 3-4h、PSNR risk LOW。Tier 1 f16 と stack 可能だが ROI 限定的","\u003Cstrong>MPS sort (vendor-optimized 置換)\u003C\u002Fstrong>: emit_pairs + radix_sort 合計 27.8% share を MPS で一括置換、期待 -1.9-2.2% wallclock combined、cost 5-6h、PSNR risk MED (tie-breaking 安定性検証必須)。Tier 2 の純自作改修と排他的、ROI 大だが research narrative (\"vendor 依存\") に注意",{"type":47,"text":137},"5. 既排除 \u002F 再帰評価 (A.6 \u002F A.7 \u002F A.1)",{"type":117,"items":139},[140,141,142],"\u003Cstrong>A.6 f16 packed Splat\u003C\u002Fstrong>: feat_g_f16_packed_roi.md で -1% ROI 実測済、Phase D 後 forward 6.1% share と orthogonal (buffer layout 変更 vs kernel compute)。**re-attack で >0.1% gain は見込めず**、棄却維持","\u003Cstrong>A.7 ICB (MTLIndirectCommandBuffer)\u003C\u002Fstrong>: 既部分実装 (\u003Ccode>extract_offsets_from_buf\u003C\u002Fcode> tile_bin.rs:449-506、\u003Ccode>forward_from_buf\u003C\u002Fcode> \u002F \u003Ccode>backward_from_buf\u003C\u002Fcode> rasterize.rs:192-465)。trainer 統合 pending、async command buffer overlap が主 benefit (per-kernel speedup ではない)、本 audit 5 候補と orthogonal で独立進行","\u003Cstrong>A.1 SSIM TBDR\u003C\u002Fstrong>: Phase 5 探索済、本 audit の Tier 3 ssim_fwd_grad imageblock+TBDR と本質同じ。SSIM eval-only のため training 効果なし、卒論 narrative 価値のみ",{"type":47,"text":144},"6. 推奨 sequence (next session 用 task list)",{"type":50,"columns":146,"align":153,"rows":154,"caption":181},[147,148,149,150,57,151,152],"順","task","scope","expected wallclock","cost","前提条件",[62,63,63,62,64,64,63],[155,160,166,170,174],[67,156,157,73,72,158,159],"**emit_pairs_simd gate flip + smoke**","tile_bin.rs:87 false → true, 5k smoke + 30k Lego A\u002FB","1 line + 1 run","Phase D 完遂 ✓ (gate 開放条件達成済)",[77,161,162,163,91,164,165],"f16 forward A\u002FB test","SPLAT_F16_FORWARD=1, 5k smoke 全 8 scene PSNR 計測","-0.03〜-0.06%","0.5h + 8 runs","Tier 1、underflow 検証必須",[86,167,168,82,72,83,169],"radix_sort GPU prefix sum 実装","all-GPU scan kernel + dispatch loop refactor","Tier 1 完了後 (累積 wallclock 測定可能)",[96,171,172,92,91,93,173],"backward_raster imageblock+TBDR prototype","tile-local accumulator + Kahan summation + tile flush","Tier 1 完了後、5k smoke で PSNR drift 計測",[105,175,176,177,178,179,180],"Final bench: 全 Tier 1+2 適用 30k full + 8 scene","Phase D re-chain と同 protocol で再測、wallclock 累積効果 + PSNR 全項目維持確認","**-1.5〜-2.5%**","LOW (検証のみ)","1 day (re-chain script)","Tier 1-4 完了後","Tier 1 (順 1-2) は即 actionable、Tier 2 (順 3-4) は Phase E scope。順 5 で累積効果を確定。期待 total -1.5-2.5% wallclock、PSNR 維持 (8 scene mean 33.49 dB 死守)。",{"type":47,"text":183},"7. 卒論 narrative 価値",{"type":117,"items":185},[186,187,188,189],"\u003Cstrong>§5.4 axis 1 native Metal narrative 強化\u003C\u002Fstrong>: 「既実装 SIMD reduction を 30k validation gate 通過で flip → 即 actionable な構造改善」という engineering discipline narrative","\u003Cstrong>§6 future work 4 候補\u003C\u002Fstrong>: emit_pairs SIMD (gate flip 済み前提)、radix_sort GPU prefix、backward TBDR、SSIM fusion を「Phase D 完遂後の axis 1 ROI 上限と未踏可能性」として明記","\u003Cstrong>A.6 \u002F A.7 \u002F A.1 の再帰評価 history\u003C\u002Fstrong>: 過去の探索結果を本 audit で再 ground、卒論 negative findings 章 (chapter-5-4-negative-findings.md) に追記候補","\u003Cstrong>profile share の動的性質\u003C\u002Fstrong>: clean baseline で share が大幅 shift (radix 27% → 13.6%、emit 6.5% → 14.2%) を踏まえた「ROI 推定は contention 状態で over\u002Funder-state」方法論を §3 evaluation methodology に追記検討",{"type":47,"text":191},"8. 関連",{"type":117,"items":193},[194,195,196,197,198],"本 audit の baseline data: \u003Ccode>p1-profiling-clean\u003C\u002Fcode>","target_upload cache の経験 (async commit overlap で ROI 1\u002F25): \u003Ccode>p1-axis1-target-cache\u003C\u002Fcode>","Phase E refine GPU 化棄却 (ROI 仮説 -5x → 実測 0.2% share): \u003Ccode>p1-e-refine-gpu-smoke\u003C\u002Fcode>","f16 packed ROI -1% 実測 (re-attack 棄却): \u003Ccode>a-6-f16-packed-rebench\u003C\u002Fcode>","Phase D 完遂で gate flip 開放条件達成: \u003Ccode>p1-d-multi-scene-rechain\u003C\u002Fcode>",[],[201,223,254,276],{"id":28,"title":202,"date":9,"status":10,"polarity":203,"category":204,"axes":205,"tags":206,"task_code":214,"related_runs":215,"delta_psnr":217,"delta_wallclock":218,"rank":219,"verdict":220,"impact_summary":221,"detail_path":222},"P1 axis 1 target_upload cache — kernel 除去は成功、wallclock ROI は host\u002FGPU overlap で予想の 1\u002F25","neutral","optimization",[14],[207,17,208,209,210,211,212,213],"p1","target-cache","kernel-removal","host-gpu-overlap","low-roi","apples-to-apples-ab","lego-5k","P1 axis 1 target upload cache",[216],"lego-target-cache-5k","+0.14 dB (seed同一、RNG drift、許容範囲)","-0.23% (apples-to-apples A\u002FB、env toggle 同一 binary)","low","accepted-cleanup-keep-merged","Per-iter target upload kernel を完全除去 (5000 calls → 0)、構造的には kernel 一つ消えた成果。 だが wallclock ROI は **予想 -5.5% に対し実測 -0.23%** (-1\u002F25)、profile baseline の \"5.6% share\" は GPU contention 3x 環境での host stall 値で、平常 contention では host upload は既に GPU 計算と overlap していた。 PSNR は seed 同一でも +0.14 dB drift (Metal driver の buffer 配置順序差 → atomic ordering 差 → refine.split RNG 経由)、5k smoke の noise floor 内。 実装は trivial (train_loop entry で `Vec\u003CBuffer>` 構築、train_step に `Option\u003C&Buffer>` 追加)、commit 残しておく価値はあるが、roadmap 上の位置付けは \"deprioritize \u002F cleanup level\" に修正。 **真の優先順位は radix_sort 改善 (27% share) と A.7 ICB batching tail に集中すべき**。","\u002Ffindings\u002Fp1-axis1-target-cache\u002F",{"id":30,"title":224,"date":9,"status":10,"polarity":12,"category":225,"axes":226,"tags":229,"task_code":239,"related_runs":240,"delta_psnr":249,"delta_wallclock":250,"rank":33,"verdict":251,"impact_summary":252,"detail_path":253},"P1.D multi-scene Phase D re-chain final — 8 scene mean 33.49 dB、brush mean 32.86 を +0.63 dB 上回り、universal win-win-win 実証","experiment",[14,227,228],2,3,[207,230,231,232,233,234,235,236,237,238],"phase-d","milestone-m5","multi-scene","brush-parity","brush-超え","premultiplied","opacity-decay","universal-win-win-win","rechain-final","P1.D multi-scene re-chain (M5 final)",[241,242,243,244,245,246,247,248],"lego-brushcompat-opacdecay-30k","chair-brushcompat-opacdecay-30k","ficus-brushcompat-opacdecay-30k","drums-brushcompat-opacdecay-30k","hotdog-brushcompat-opacdecay-30k","mic-brushcompat-opacdecay-30k","materials-brushcompat-opacdecay-30k","ship-brushcompat-opacdecay-30k","8 scene mean +0.63 dB vs brush paper (33.49 vs 32.86)","-61% total chain (13h+ → 5h 5m)","accepted-m5-complete","Phase D opacity_decay (rate=0.004 brush default) を 7 scene × 30k full chain bench、Lego val Phase D 30k と合わせて 8 scene 集計。**全 scene で baseline brushcompat 30k 比 PSNR + splats + wallclock すべて改善 (universal win-win-win)**: PSNR +0.18〜+1.42 dB \u002F splats -57〜-78% \u002F wallclock -39〜-69%。8 scene mean 33.49 dB vs brush paper 8 scene mean 32.86 dB = **+0.63 dB 上回り**、本実装が brush の multi-scene mean を decisive に超えた。brush 超え 3 scene (Lego val +4.07 \u002F drums +1.05 \u002F mic +1.02)、4 scene が brush 比 ±0.7 dB 圏内 (chair -0.02 \u002F hotdog -0.39 \u002F ship -0.01 \u002F materials -0.10)、最遠 scene でも ficus -0.65 で接近。全体 wallclock baseline chain (13h+) → Phase D re-chain 5h 5m (-61%)、mean splats 1.4M → 428k (-69%) で brush 282k に肉薄。P1.M5 完全達成 (Lego val > 36 dB ✅ + multi-scene mean > 32 dB ✅)、卒論 central evaluation table の final 数字確定、universal claim 完全実証。","\u002Ffindings\u002Fp1-d-multi-scene-rechain\u002F",{"id":29,"title":255,"date":9,"status":10,"polarity":256,"category":225,"axes":257,"tags":258,"task_code":266,"related_runs":267,"delta_psnr":271,"delta_wallclock":272,"rank":33,"verdict":273,"impact_summary":274,"detail_path":275},"P1.E refine GPU 化 hypothesis を SPLAT_TIMING profile で falsified — refine 寄与は \u003C1%、真の bottleneck は forward 60% + loss 28%","negative",[14],[207,259,260,261,262,263,264,265,213],"phase-e","refine-gpu","negative-finding","kernel-profile","axis-1-limit","opacity-decay-gpu","kernel-plumbing","P1.E refine GPU 化 (axis 1 core contribution)",[268,269,270],"p1-e-profile-1k","p1-e-profile-5k","p1-e-gpu-decay-5k","-0.21 dB (CPU 31.92 → GPU 31.71、5k smoke、bit-close 内 RNG cascade)","+1.4% (CPU 144.32s → GPU 146.32s、5k smoke、opacity_decay は 0.005% で誤差)","accepted-negative-redirect-phase-f","Phase D 30k 実測 wallclock 41m54s vs brush 9m08s = -4.6x gap の原因について、task は `splat process CPU 63.4% = 1 core only` → 「refine の host RMW loop が CPU 1-thread bound」と仮説立てた。本 Phase E ではこの仮説を SPLAT_TIMING profile で実測。5k smoke (84k splats、p1-e-profile-5k.toml) の kernel breakdown: **ts_forward 60.1% (123s) \u002F ts_loss_gpu 28.0% (57s) \u002F ts_adam 4.8% (9.9s) \u002F ts_target_upload 3.9% (8.1s) \u002F ts_project_back 2.3% (4.75s) \u002F ts_refine_compact 0.6% (1.14s, 103ms\u002Fcall × 11) \u002F ts_refine_accumulate 0.3% (605ms) \u002F ts_opacity_decay 0.0046% (957µs)**。**refine 全体で \u003C1%** = refine を完璧に GPU 化しても全体 wallclock は -1% も短縮されない。代わりに demo kernel として `refine_opacity_decay.metal` を実装し、kernel + Rust pipeline + `refine.gpu_path` flag plumbing pattern を validate (CPU vs GPU max diff 1.5e-5、5k full PSNR delta -0.21 dB = 許容内)。Phase F の真の target は (a) forward subdivision で判明した tile-binning chain (`ts_fwd_sort 15.5% + ts_fwd_emit 12.8%`)、(b) Adam の 5x sequential `cmd.wait_until_completed` (1 cmd buffer 化で ~5% 削減期待)、(c) target_upload preload (~4% 削減期待) の 3 つ。","\u002Ffindings\u002Fp1-e-refine-gpu-smoke\u002F",{"id":27,"title":277,"date":9,"status":10,"polarity":203,"category":278,"axes":279,"tags":280,"task_code":285,"related_runs":286,"delta_psnr":35,"delta_wallclock":35,"rank":33,"verdict":288,"impact_summary":289,"detail_path":290},"P1 clean single-process profile baseline — radix_sort 27% → 13.6%、emit_pairs 6.5% → 14.2% の share 大幅 shift、axis 1 真の ROI 上限確定","audit",[14],[16,17,281,282,283,235,213,284],"clean-baseline","kernel-share","single-process","share-correction","P1 axis 1 profile re-baseline",[287],"lego-profile-clean-5k","audit-complete-share-correction","clean single-process で per-kernel share を取り直し、前 profile baseline (multi-process contention 中) と比較すると \u003Cstrong>share が大幅 shift\u003C\u002Fstrong>: ts_fwd_radix_sort 27.0% → **13.6%** (-13.4 pt)、ts_fwd_emit_pairs 6.5% → **14.2%** (+7.7 pt)、ts_forward 全体 60.1% → **36.7%** (-23.4 pt)。これは前 profile の share が contention で over-state されていた決定的証拠 (target_upload subagent の share 5.6% → 実 ROI 0.23% を kernel level で再現)。新 axis 1 ROI 上限: emit_pairs 改善 -14% \u002F radix_sort -13% \u002F backward_raster -13% \u002F ssim_fusion -7-8%。Phase E (refine GPU 化) の ROI 仮説 -5x は元々 share 2.6% で 1\u002F40 過大評価だったが、本 clean baseline でも refine 0.2% に縮小、棄却強化。target_upload は本 clean baseline で完全消滅 (cache 化済)。","\u002Ffindings\u002Fp1-profiling-clean\u002F",1782449788631]