[{"data":1,"prerenderedAt":314},["ShallowReactive",2],{"finding:p2-w1-warp-sparse-skip":3,"finding-runs:p2-w1-warp-sparse-skip":199,"finding-related:p2-w1-warp-sparse-skip":228},{"meta":4,"impact":48,"sections":54},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":15,"task_code":26,"commit":27,"branch":28,"config_paths":29,"related_runs":39,"related_findings":43},"p2-w1-warp-sparse-skip","W1 warp sparse skip + exp-skip — backward -35% \u002F replay -11.2% (bit-exact)","backward Simd に (1) \u003Ccode>sigma >= 5.55\u003C\u002Fcode> の exp-skip (α = color_a·exp(-σ) ≤ exp(-5.55) ≈ 0.003887 \u003C 1\u002F255 — 既存 α\u003C1\u002F255 分岐で必ず落ちる iteration の前倒し skip) と (2) \u003Ccode>simd_any(contrib)\u003C\u002Fcode> による all-invalid warp skip (9 simd_sum + lane0 の atomic_fetch_add(+0.0) の静的削除) を追加、forward にも同じ sigma skip (W1b、無条件)。どちらも \u003Cstrong>bit-exact\u003C\u002Fstrong> (single-pixel fixture の byte equality で証明)。replay bench で \u003Cstrong>final regime 21.72 → 19.30 ms\u002Fiter = -11.2% \u002F iter5000 regime 19.98 → 17.76 = -11.1%\u003C\u002Fstrong>、backward_raster 7.18 → 4.65 ms\u002Fcall (\u003Cstrong>-35.2%\u003C\u002Fstrong>、GPU 部 -42〜43%)。12.5k full gate n=3: wallclock mean 261.55 → \u003Cstrong>241.41s (-7.7%)\u003C\u002Fstrong>、PSNR mean \u003Cstrong>36.201 dB\u003C\u002Fstrong> (基準 36.148 比 +0.053、guard ±0.1 内) → default flip。msplat 同日等価 gap 1.56× → \u003Cstrong>1.44×\u003C\u002Fstrong>。","1× 工程表 v2 · W1 · backward\u002Fforward skip","2026-08-16","stable","experiment","positive",[14],1,[16,17,18,19,20,21,22,23,24,25],"1x-roadmap-v2","w1","warp-sparse-skip","exp-skip","backward","forward","simd-any","function-constant","bit-exact","lego","2026-08 1× 工程表 v2 W1","093886f8b50e","main",[30,31,32,33,34,35,36,37,38],"splat\u002Fconfigs\u002F2026-08-16-w1-parity-baseline.toml","splat\u002Fconfigs\u002F2026-08-16-w1-parity-candidate.toml","splat\u002Fconfigs\u002F2026-08-16-w1-bench-final-off.toml","splat\u002Fconfigs\u002F2026-08-16-w1-bench-final-warpany.toml","splat\u002Fconfigs\u002F2026-08-16-w1-bench-iter5000-off.toml","splat\u002Fconfigs\u002F2026-08-16-w1-bench-iter5000-warpany.toml","splat\u002Fconfigs\u002F2026-08-16-lego-w1-warpany-12500.toml","splat\u002Fconfigs\u002F2026-08-16-lego-w1-warpany-12500-rep2.toml","splat\u002Fconfigs\u002F2026-08-16-lego-w1-warpany-12500-rep3.toml",[40,41,42],"lego-w1-warpany-12500","lego-w1-warpany-12500-rep2","lego-w1-warpany-12500-rep3",[44,45,46,47],"p2-s1-pipeline-resident","p2-last-contributor","p2-bench-harness-calibration","p2-msplat-parity-lego",{"summary":49,"rank":50,"verdict":51,"delta_psnr":52,"delta_wallclock":53},"1× 工程表 v2 の第 1 波。msplat \u002F Inria \u002F brush の 3 実装が持ち本実装だけ欠落していた backward の 2 つの skip を、bit-exact の等価性論証つきで移植した。backward_raster は LC (2-BWD) 後の 7.2ms\u002Fcall から 4.65ms\u002Fcall へ -35.2% (kernel share 13.1% に低下)、replay 全体で -11.2%。12.5k full run は 261.55 → 241.41s で msplat 同日等価 gap は 1.56× → 1.44× に縮小。数学的同値の施策のみの累計は 766.07 → 241.41s = -68.5% になった。次の律速は forward (23.6%) \u002F radix (15.0%) で、W2 (encoder\u002FCB 統合) と計測 (counter session) が次段。","high","w1-adopted-default-flip-warp-any","12.5k gate n=3: 36.209 \u002F 36.201 \u002F 36.193 dB → mean 36.201 dB = 基準 36.148 (S1 n=3 mean) 比 +0.053 dB (guard ±0.1 内、n=3 mean 判定 — S1 の教訓に従い n=1 判定なし)","replay bench final regime: 21.724 → 19.300 ms\u002Fiter = **-11.16%** (閾値 max(2σ,1%)=2.7%)。iter5000 regime: 19.984 → 17.759 = **-11.13%** (閾値 3.1%)。backward_raster: 7.183 → 4.652 ms\u002Fcall (**-35.2%**、GPU 部 5.963 → 3.437 = -42.4%)。12.5k full run n=3: 239.74 \u002F 246.93 \u002F 237.57s → mean **241.41s** (S1 基準 261.55s 比 **-7.7%**)",[55,58,61,81,86,88,130,134,136,176,179,181,196],{"type":56,"text":57},"lead","HW 4 軸調査 #1 (他実装 HW 技法比較) で、backward rasterize の inner loop に \u003Cstrong>msplat \u002F Inria 3DGS \u002F brush の 3 実装すべてが持つ 2 つの skip\u003C\u002Fstrong> が本実装だけ欠けていると特定した (msplat 実測では evaluation の 94% が non-contributing)。(1) \u003Cstrong>exp-skip\u003C\u002Fstrong>: splat の color_a は projection の sigmoid 出力 ∈ (0,1) なので α = color_a·exp(-σ) ≤ exp(-5.55) ≈ 0.003887 < 1\u002F255 ≈ 0.003922 — sigma ≥ 5.55 の iteration は既存の α < 1\u002F255 分岐で v_* = 0 のまま必ず落ちる。これを exp 評価前に落としても実行される演算列は不変 (\u003Cstrong>bit-exact\u003C\u002Fstrong>)。(2) \u003Cstrong>warp sparse skip\u003C\u002Fstrong>: all-invalid warp では 9 本の simd_sum が全て +0.0 → lane0 の atomic_fetch_add(+0.0) になるため、\u003Ccode>simd_any(contrib)\u003C\u002Fcode> で reduction ごと省略しても最終 splats_grad は順序非依存で bit-exact。\u003Ccode>simd_any\u003C\u002Fcode> は simd-uniform を返すので non-divergent、done フラグ遷移と \u003Ccode>end = simd_max(my_stop)\u003C\u002Fcode> bound (LC) は不変。",{"type":59,"text":60},"heading","1. 実装 (`raster_backward_sparse = { off, warp_any }`)",{"type":62,"columns":63,"align":66,"rows":68},"table",[64,65],"項目","内容",[67,67],"left",[69,72,75,78],[70,71],"W1a backward (rasterize_backwards.metal)","`rasterize_backwards_simd` のみ。lane ごとに `bool contrib` を立て、状態更新 (pix_acc \u002F T \u002F done) 完了後・reduction 直前に `if (!simd_any(contrib)) continue;`。exp-skip は `sigma >= 0 && sigma \u003C 5.55` への条件強化。Tg variant (#5.33c reject 系) には適用しない",[73,74],"variant 解決","Metal 第 3 条準拠: function constant `USE_WARP_SPARSE_SKIP_FC` (index 1、LC の `USE_LAST_CONTRIBUTOR_FC` と同じ流儀) で PSO specialize。Simd × {off, warp_any} × {full_range, last_contributor} の 2 PSO 追加 (計 8 backward PSO)、kernel 内 runtime 分岐なし",[76,77],"W1b forward (rasterize.metal)","`render_splats_f32` \u002F `f16` の `if (sigma \u003C 0)` を `if (sigma \u003C 0 || sigma >= 5.55)` に (同じ論証で bit-exact、stop_idx \u002F final_T も不変)。分岐 1 個の追加なので flag なしの無条件変更",[79,80],"config 配線","`BackendConfig.raster_backward_sparse = { off, warp_any }`、apply_backend_config の exhaustive destructuring に配線。Simd 専用 (backward = original\u002Ftg との併用は validate が stderr warning、PSO は off が選ばれる)",{"type":82,"label":83,"variant":84,"text":85},"callout","parity — atomic の影響を受けない byte equality まで証明","success","(1) synthetic fixture (24 splat 重ね) で Off vs WarpAny の勾配 tolerance 一致 (abs\u002Frel 1e-4、C7 体系) — FullRange \u002F LastContributor 両スタック。(2) \u003Cstrong>dldr を 1 pixel だけ非ゼロにした fixture で Off と WarpAny の勾配が bit 一致\u003C\u002Fstrong> (`backward_warp_sparse_byte_equality_single_pixel`、両スタック) — warp skip が「+0.0 atomic add の静的削除」、exp-skip が「必ず落ちる iteration の前倒し」であることの直接証明。(3) trainer-level parity (`splat bench --parity`、production-scale Lego): tile_offsets byte 一致 + 全 grad buffer tolerance PASS (残差 max 1.0e-5 は既知の atomic 順序非決定性)。",{"type":59,"text":87},"2. replay bench A\u002FB (final + iter5000 regime、各 n=3、baseline 同日取り直し)",{"type":62,"columns":89,"align":94,"rows":96},[90,91,92,93],"metric","off (baseline)","warp_any","Δ",[67,95,95,95],"right",[97,102,107,112,117,121,126],[98,99,100,101],"**final regime total**","21.724 ms\u002Fiter (σ 1.35%)","19.300 ms\u002Fiter (σ 0.17%)","**-11.16%** (閾値 2.7%)",[103,104,105,106],"final ts_backward_raster","7.183 ms\u002Fcall","4.652 ms\u002Fcall","**-35.2%**",[108,109,110,111],"final ts_raster_bwd_gpu","5.963 ms\u002Fcall","3.437 ms\u002Fcall","**-42.4%**",[113,114,115,116],"**iter5000 regime total**","19.984 ms\u002Fiter (σ 1.57%)","17.759 ms\u002Fiter (σ 0.69%)","**-11.13%** (閾値 3.1%)",[118,119,120,106],"iter5000 ts_backward_raster","6.336 ms\u002Fcall","4.106 ms\u002Fcall",[122,123,124,125],"iter5000 ts_raster_bwd_gpu","5.237 ms\u002Fcall","2.971 ms\u002Fcall","**-43.3%**",[127,128,128,129],"他 kernel (forward\u002Fradix\u002Fssim\u002Fadam)","—","±3.3% 以内 (noise 圏、期待どおり不変)",{"type":82,"label":131,"variant":132,"text":133},"計測 note — config 非依存の interference window を除外","info","計 8 invocation 中 2 window で全 kernel が一様に +5〜25% 遅くなる slow mode を観測 (off \u002F warp_any の両方に発生、backward だけは slow window 内でも -35% を維持)。config 非依存の系統外 interference と判定し、σ が正常な batch (off σ ≤1.6% \u002F warp_any σ ≤0.7%) を採用した。全 batch の median 帯: off 21.27〜21.72 vs warp_any 19.19〜19.73 (final) \u002F off 19.71〜20.03 vs warp_any 17.76〜18.57 (iter5000) で判定は不変。",{"type":59,"text":135},"3. 12.5k full gate (n=3、PSNR guard は n=3 mean 判定)",{"type":62,"columns":137,"align":143,"rows":144},[138,139,140,141,142],"rep","wallclock (training)","PSNR (100 views)","final splats","ts_backward_raster",[67,95,95,95,95],[145,151,157,163,168,172],[146,147,148,149,150],"rep1","239.74s","36.209 dB","410,446","4.531 ms\u002Fcall (13.1%)",[152,153,154,155,156],"rep2","246.93s","36.201 dB","406,770","4.721 ms\u002Fcall (13.2%)",[158,159,160,161,162],"rep3","237.57s","36.193 dB","411,238","4.476 ms\u002Fcall (13.1%)",[164,165,166,128,167],"**n=3 mean**","**241.41s**","**36.201 dB**","**4.58 ms\u002Fcall**",[169,170,171,128,128],"S1 baseline (n=3 mean)","261.55s","36.148 dB",[173,174,175,128,128],"**Δ**","**-7.7%**","+0.053 dB (guard ±0.1 内)",{"type":82,"label":177,"variant":84,"text":178},"verdict: 採用 — `raster_backward_sparse` default を WarpAny に flip","replay 有意閾値 max(2σ,1%) を 4 倍超過、PSNR guard は n=3 mean で通過 (S1 の教訓に従い n=1 判定なし、3 rep とも歴代分布 36.05-36.23 の上側で安定)。backward 期待帯 -25〜50% のど真ん中 (-35.2%)、total 期待帯 -5〜10% を上回る -11.2% (replay)。W1b (forward sigma skip) は無条件変更として同梱 — 同日 S1 post replay (22.06 ms\u002Fiter) と W1 off baseline (21.27〜21.72) の差 -1.5〜-3.6% が W1b + 日内変動に相当するが、bit-exact のため単独 gate は不要。Off は A\u002FB baseline 用に enum 保持。",{"type":59,"text":180},"4. msplat gap の現在地",{"type":62,"columns":182,"align":187,"rows":188},[183,184,185,186],"時点","Lego 12.5k","msplat 同日実測比 (167.7s)","備考",[67,95,95,67],[189,192],[190,170,191,44],"S1 (pipeline 全域 resident 化)","1.56×",[193,165,194,195],"**+ W1 (warp sparse skip + exp-skip)**","**1.44×**","本 finding、-7.7% — 全て数学的同値の施策のみで累計 766.07 → 241.41s = **-68.5%**",{"type":82,"label":197,"variant":132,"text":198},"次段 — W2 (encoder\u002FCB 統合) + counter session","W1 後 breakdown の上位は forward 23.6% \u002F radix_sort 15.0% \u002F backward_raster 13.1% \u002F SSIM ~7% — backward はもはや dominant ではない。1× 工程表 v2 の順序どおり W2 (Stage A: adam CB 統合 + setBytes 化) に進み、W1\u002FW2 着弾後に counter session で limiter を確定する。",[200,214,221],{"id":40,"title":40,"subtitle":201,"date":9,"workspace":202,"tags":203,"verdict":209,"psnr":210,"psnr_unit":-1,"wallclock":211,"splats":212,"summary_url":213,"detail_path":213},"2026-08 1× 工程表 v2 W1 — warp sparse skip + exp-skip (bit-exact) の 12.5k PSNR guard + wallclock gate (rep1\u002F3 — PSNR guard は n=3 mean 判定)","splat",[204,18,19,205,206,207,208],"p2-w1","gpu-gate","lego-12500","brush-compat","premultiplied","partial",36.20934295654297,"3m 59s",410446,"\u002Fruns\u002Flego-w1-warpany-12500\u002F",{"id":41,"title":41,"subtitle":215,"date":9,"workspace":202,"tags":216,"verdict":209,"psnr":217,"psnr_unit":-1,"wallclock":218,"splats":219,"summary_url":220,"detail_path":220},"2026-08 1× 工程表 v2 W1 — warp sparse skip + exp-skip (bit-exact) の 12.5k PSNR guard + wallclock gate (rep2\u002F3 — PSNR guard は n=3 mean 判定)",[204,18,19,205,206,207,208],36.20069122314453,"4m 6s",406770,"\u002Fruns\u002Flego-w1-warpany-12500-rep2\u002F",{"id":42,"title":42,"subtitle":222,"date":9,"workspace":202,"tags":223,"verdict":209,"psnr":224,"psnr_unit":-1,"wallclock":225,"splats":226,"summary_url":227,"detail_path":227},"2026-08 1× 工程表 v2 W1 — warp sparse skip + exp-skip (bit-exact) の 12.5k PSNR guard + wallclock gate (rep3\u002F3 — PSNR guard は n=3 mean 判定)",[204,18,19,205,206,207,208],36.193382263183594,"3m 57s",411238,"\u002Fruns\u002Flego-w1-warpany-12500-rep3\u002F",[229,249,269,293],{"id":46,"title":230,"date":9,"status":10,"polarity":12,"category":231,"axes":232,"tags":234,"task_code":242,"related_runs":243,"delta_psnr":-1,"delta_wallclock":245,"rank":50,"verdict":246,"impact_summary":247,"detail_path":248},"splat bench harness 実装 + calibration — σ_replay 0.83%、full run 比 ~6x ノイズ削減","tooling",[14,233],3,[235,236,237,238,239,240,241,25],"bench-harness","calibration","variance","reproducibility","c7","fixed-workload","gate-infra","splat bench harness (2026-08-16 GPU 専有タスク)",[244],"lego-bench-calib-checkpoints","σ_replay 0.83% (n=10, final-scale workload) vs full run 実測 ~10% — 約 6x ノイズ削減","accepted","`splat bench --config bench.toml` を実装し、同一 config n=10 replay で \u003Cstrong>σ_replay = 0.234 ms\u002Fiter (mean 28.339 ms 比 0.83%)\u003C\u002Fstrong> を実測。full training run で観測されていた ~10% 分散に対し \u003Cstrong>約 6x のノイズ削減\u003C\u002Fstrong>。3-regime (iter 1000 \u002F 5000 \u002F final checkpoint、各 n=5) でも σ 1.2-4.6% の範囲に収まり、splat 数が少ない早期 regime ほど相対分散が大きい傾向を確認。有意判定閾値を \u003Ccode>max(2σ, 1%)\u003C\u002Fcode> と規約化し、以後の gate 判定をこの harness の計測に載せ替える。付随して `--parity` (kernel variant 間の勾配 \u002F binning 一致検証) と `[output] save_ply_at` (checkpoint 保存) を実装、parity mode の実測で production-scale シーンの `sorted_idx` tie-order 非決定性 (~0.2%、真の非決定性であって variant 差ではない) を発見し gate から除外した。","\u002Ffindings\u002Fp2-bench-harness-calibration\u002F",{"id":45,"title":250,"date":9,"status":10,"polarity":12,"category":11,"axes":251,"tags":252,"task_code":258,"related_runs":259,"delta_psnr":264,"delta_wallclock":265,"rank":50,"verdict":266,"impact_summary":267,"detail_path":268},"P2 gap sweep Phase 2 — 2-BWD last_contributor で -15.0% (数学的同値)、STP 再判定は棄却",[14],[253,254,255,20,256,257,23,25],"gap-sweep","phase-2","2-bwd-last-contributor","stop-the-pop","re-eval","2026-08 msplat gap sweep Phase 2 2-BWD \u002F STP 再判定",[260,261,262,263],"lego-lc-smoke-1k","lego-p2lc-12500","lego-stp-c0-12500","lego-stack-lc-stp-12500","LC 12.5k gate: 36.183 → 36.191 dB = +0.008 dB (数学的同値実装、atomic 非決定性 guard ±0.1 内)","LC 12.5k gate: 395.59 → 336.33s = **-14.98%** (backward_raster -40.3%)。STP: -2.51% (gate 未達)、LC+STP スタック: 343.63s (LC 単独より +2.2%)","lc-adopted-minus-15-percent-stp-rejected-final","C0 (-48.4%) に続く Phase 2 第 2 弾で、backward 31.4% の新 bottleneck を kernel 側から -40.3% 削減。\u003Cstrong>Lego 12.5k = 336.33s で msplat 30K (220s) 比 1.53× まで縮小\u003C\u002Fstrong>し、gap sweep の目標 (≤2.0×) を達成した。実装は「forward が既に知っている打ち切り位置を backward に伝えるだけ」で数学的同値 — 品質を 1 bit も動かさずに backward の replay 計算量を ~40% 削った。STP 再判定は負の結果だが価値が高い: Phase 1a の -7.45% の正体が C0 と同じ page fault 削減だったことを分離実証し、「効果の重複を疑って再 bench する」protocol の正しさを裏付けた。次の bottleneck は forward rasterize 9.8% \u002F radix 13.5% \u002F SSIM 7.5% でいずれも trigger 25% 未達 — Phase 2 はここで完了、Phase 3 (再計測 + Pareto v3) へ。","\u002Ffindings\u002Fp2-last-contributor\u002F",{"id":47,"title":270,"date":9,"status":10,"polarity":271,"category":11,"axes":272,"tags":274,"task_code":284,"related_runs":285,"delta_psnr":288,"delta_wallclock":289,"rank":50,"verdict":290,"impact_summary":291,"detail_path":292},"P2 msplat parity — 同日等価比較で msplat 30K = 167.7s \u002F 36.5 dB、gap は 12.5k で 2.10× (matched-wallclock 7.5k は 1.08× で brush +0.94 dB)","mixed",[14,273],2,[253,275,276,277,278,279,280,281,25,282,283],"phase-3","msplat","parity","baseline","cross-eval","eval-convention","thermal-interleave","m4-max","thesis-ready","2026-08 msplat gap sweep Phase 3 (parity 再計測)",[286,287],"lego-parity-noprof-12500","lego-parity-noprof-7500","brush convention (同一 evaluator): 本実装 12.5k +1.79 dB \u002F 7.5k +0.94 dB (vs msplat 30K q8 34.47)。white-bg convention: msplat +1.09 dB (36.28 vs 本実装 12.5k 35.19)","msplat 30K 167.7s vs 本実装 12.5k 352.3s = **2.10×** (両者 profile なし・同日実測)。matched-wallclock: 7.5k 181.4s = 1.08×。従来の 1.53× (spike 値 220s 基準) は撤回","parity-established-gap-2.10x-convention-coupled-psnr","gap sweep の公式 gap 数値を \u003Cstrong>同日・同一条件の実測で確定\u003C\u002Fstrong>: Lego 12.5k vs msplat 30K の wallclock gap は \u003Cstrong>2.10×\u003C\u002Fstrong> (352.3s vs 167.7s、両者 profiling なし) — Phase 2 の 1.53× は spike の古い msplat 値 (219.8s) 起因の過小評価だった。一方 \u003Cstrong>matched-wallclock 断面 (本実装 7.5k = 181.4s、msplat の 1.08×)\u003C\u002Fstrong> では brush convention の同一 evaluator cross-eval で本実装が \u003Cstrong>+0.94 dB\u003C\u002Fstrong> (35.41 vs 34.47)。PSNR 比較は eval convention に強く依存し (white-bg では msplat +1.09 dB)、p1-a-3 の「訓練\u002F評価 convention coupling」を第三者実装との比較で再実証した。msplat の native 品質が spike 比 +1.05 dB 向上した主因は今回の対称条件 (同一 culled init.ply 5207 pts・train 100 view 全量・常時フル解像度) にあり、msplat 側の条件を今回初めて本実装と揃えた結果。卒論 §5 の direct comparison 表はこの数値で置き換える。","\u002Ffindings\u002Fp2-msplat-parity-lego\u002F",{"id":44,"title":294,"date":9,"status":10,"polarity":12,"category":11,"axes":295,"tags":296,"task_code":303,"related_runs":304,"delta_psnr":309,"delta_wallclock":310,"rank":50,"verdict":311,"impact_summary":312,"detail_path":313},"S1 pipeline 全域 resident 化 — replay bench -22.9% (final regime) で採用 (Tier A\u002FB\u002FC)",[14,233],[253,254,297,298,299,300,205,301,235,25,302],"s1","resident","alloc-audit","unified-memory","adopted","attribution","S1 pipeline 全域 resident 化 worktree 統合 + GPU gate (2026-08-16 GPU 専有タスク)",[305,306,307,308,244],"lego-s1-resident-12500","lego-s1-resident-12500-rep2","lego-s1-resident-12500-rep3","lego-gap-baseline-12500","12.5k full run PSNR guard: 基準 36.19 前後 (歴代同 recipe 7 run 分布 36.138-36.226) に対し rep1 36.051 (-0.14、guard 外) \u002F rep2 36.168 \u002F rep3 36.224 dB → n=3 mean 36.148 dB = -0.04 dB (±0.1 dB guard 内、PASS)。rep1 は歴代最低値だが S1 は数学的同値 (parity test 群で証明) のため run 分散の tail と判定","replay bench final regime n=3: 28.630 → 22.059 ms\u002Fiter = **-22.9%** (閾値 max(2σ,1%)=2.05%)。iter5000 regime n=3: 26.113 → 20.405 ms\u002Fiter = **-21.9%** (閾値 1.29%)。12.5k full run: 公式 336.33s → rep1 272.36s \u002F rep2 257.55s \u002F rep3 254.73s = mean 261.55s (**-22.2%**)","adopted-unconditional-no-flag","別 worktree で実装済みだった S1 (pipeline 全域 resident 化) を main へ統合。C3 統合後の tile_bin.rs と conflict したため手 merge し、C3 (rejected flag) との新規相互作用 (S1 の buffer 直結が C3 の panic guard に当たる) はフォールバック + regression test で解決。build warning 0、MSL smoke compile (adam.metal binding 変更) + 新規 S1 test 9 本 (resident byte-stability \u002F grow path \u002F ForwardState parity \u002F Vec vs raw parity \u002F C3 fallback) 含む workspace 全 test pass。replay bench は \u003Cstrong>final regime -22.9% \u002F iter5000 -21.9%\u003C\u002Fstrong> と Phase 2 単一項目として C0 に次ぐ効果で、attribution 計測 (S1 で新設した DispatchAttribution 4 点) が alloc 成分の消滅を直接証明 — C0 の「主犯は first-touch page fault」結論の pipeline 全域への一般化。12.5k full gate は PSNR rep1 が guard 外 (36.051) に出たが n=3 escalation で tail variance と確定 (rep2 36.168 \u002F rep3 36.224、歴代同 recipe 分布 36.14-36.23)。採用 — flag なしの無条件変更として default 化。","\u002Ffindings\u002Fp2-s1-pipeline-resident\u002F",1790227327645]