{
  "generated_at": "2026-09-10T09:49:19.392498+00:00",
  "full_plan_complete": false,
  "plan_snapshot_sha256": "178b8b8b5735bf4c4f30b8cbe027db7764c95e47254c680b876e32d097e50455",
  "scope": "Requirement-level review against cached user plan and active goal. Evidence hashes identify reviewed snapshots; file existence alone does not prove completion. Report is a time-point delivery, not full-plan acceptance.",
  "requirements": [
    {
      "id": "K0",
      "requirement": "环境与完整时间线",
      "status": "部分完成",
      "finding_and_remaining": "环境身份及worker尾部覆盖已核对；全服务阶段与CPU因果链仍不完整。 新增同一运行的nsys+NVTX账本（生产模式）与逐chunk关键路径：480x832 稳态DiT 678 ms/VAE段 169 ms（解码 142，IPC回传 16），4卡同时空闲 58 ms；704x1280 稳态DiT 2138 ms/VAE段 338 ms（解码 290，IPC回传 34），4卡同时空闲 59 ms。时钟锚点与尾部检查通过。仍缺：CPU侧等待的逐线程因果链（请求进入与结果返回的空闲段仅定位到进程边界）与编码/网络/客户端段。",
      "evidence": [
        {
          "path": "analysis/original_nsys_worker_flush_v2_anchor_audit.json",
          "bytes": 1867,
          "sha256": "5703ab5ddee1464319b8b323a054c52324b3b6fa1c3a930b0267d55c76e805aa"
        },
        {
          "path": "analysis/worker_longest_gap_context.json",
          "bytes": 45274,
          "sha256": "18d17ef673d24ddf95c0be2d7b136eee9e7a7292765ed0db987e1ac282969712"
        },
        {
          "path": "analysis/ledger_capture_480x832_v1_anchor_audit.json",
          "bytes": 1868,
          "sha256": "e4880fbf532abb659647407f1c9d0f559c01e829da1946d11d73a122b182b915"
        },
        {
          "path": "analysis/ledger_capture_704x1280_v1_anchor_audit.json",
          "bytes": 1868,
          "sha256": "8ca9a735cd28d1a0140984011c75040baa65437294164ef3704d331cd4ccfc7f"
        },
        {
          "path": "analysis/ledger_480x832_v1/critical_path_v1.json",
          "bytes": 37473,
          "sha256": "599331ad6e88603c2d1e634baebfe454950fc9844e5f21554a51650246fe729d"
        },
        {
          "path": "analysis/ledger_704x1280_v1/critical_path_v1.json",
          "bytes": 34928,
          "sha256": "3b151aaffb1ce83cdf40167c5e35883e1b9397cadc1e1a7e94591939e17a0498"
        }
      ]
    },
    {
      "id": "K1",
      "requirement": "每次调用、shape和上下文追溯",
      "status": "部分完成",
      "finding_and_remaining": "1091749次记录调用有物理GPU/API关联；完整逐层shape对应尚缺。 新增480×832 TP1/SP4的32000次Attention有效长度/阶段/层记录；与原nsys不属同一运行，不能直接逐调用嫁接。 高分辨率新增28个代表输入、90个kernel；两分辨率合计59个算子输入覆盖各自所选线性/卷积/转换结构。PID2019660有未归因的调度备注，保留原失败和后续带备注审核。两分辨率逐调用元数据均已补充，仍非与旧nsys同次运行映射。 新增同一运行逐调用账本：480x832 1,092,983次kernel，87.6%调用/99.9%稳态GPU时间有NVTX上下文，graph内位置由inductor wrapper源码对齐到子算子；704x1280 1,092,185次kernel，87.5%调用/100.0%稳态GPU时间有NVTX上下文，graph内位置由inductor wrapper源码对齐到子算子。shape按eager全算子dispatch采集附加（480x832 稳态GPU时间94.7%有shape；704x1280 稳态GPU时间95.8%有shape），融合kernel的shape为推断值。仍缺：融合kernel精确字节、非同一运行的shape来源说明已保留。 eager全算子dispatch采集的40份输出与对应eager参考逐位一致：480x832、704x1280。",
      "evidence": [
        {
          "path": "analysis/worker_flush_v2_runtime_link_summary.json",
          "bytes": 991,
          "sha256": "e3a0fc2df8ed502e4a0e4ae2ba154e38247392938a6727f0e7f6bb3f60b01567"
        },
        {
          "path": "analysis/worker_launch_resources_summary.json",
          "bytes": 32320,
          "sha256": "e11feb17a2cc20972cbba85641005690b8ad60652967e04946164c6de7cdd68e"
        },
        {
          "path": "analysis/coverage.json",
          "bytes": 7758,
          "sha256": "7a03da4996ad12302c8b309608327c73f7b780a916a8d1640f634543a83ec995"
        },
        {
          "path": "analysis/attention_length_audit_v1.json",
          "bytes": 6341,
          "sha256": "0bb52d404e703ee79529030de65a2fb22e6570bac128355c584981bf70fceb24"
        },
        {
          "path": "analysis/attention_length_model_invariance_v1.json",
          "bytes": 41408,
          "sha256": "a1e2d37f15164e534629d8eaf6be6f2397f2377ab7422befc795c2c497340a85"
        },
        {
          "path": "analysis/missing_operator_704_ncu_audit_v1.json",
          "bytes": 195620,
          "sha256": "f4401c75f2607b2d513134edc6455173e9568ed165ea3f4b8634d01e329f7a6e"
        },
        {
          "path": "analysis/operator_replay_signature_coverage_704_v2.json",
          "bytes": 58361,
          "sha256": "471d764af2767257dc14ebcf3642159a46ed4d886a3c46474cdd68c84661edff"
        },
        {
          "path": "analysis/operator_metadata_704_provenance_audit_v1.json",
          "bytes": 1158,
          "sha256": "e829b7a37116257d8144d0f073528d5354419e9962a252c0d8653fc32ce2784e"
        },
        {
          "path": "analysis/resolution_recorded_work_comparison_v1.json",
          "bytes": 2053,
          "sha256": "ccecb9622ce9d1aaf914af2afce9feed5322176a78aebe14b432141e5eaf7ac6"
        },
        {
          "path": "analysis/ledger_480x832_v1/coverage.json",
          "bytes": 14690,
          "sha256": "f543ab624327eaea2680bb90519a38f55d0ad950a956fd243865120a5fc12056"
        },
        {
          "path": "analysis/ledger_704x1280_v1/coverage.json",
          "bytes": 14690,
          "sha256": "ff24b4f98e94e104913e4606071c2c92d327ec3251b99630ba34eee17bc1c5c3"
        },
        {
          "path": "analysis/ledger_480x832_v1/summary_v1.json",
          "bytes": 202546,
          "sha256": "bb27af636e62ed4a452f5fe6936e79186781c3ce67b78eff46306cd3b14b5f38"
        },
        {
          "path": "analysis/ledger_704x1280_v1/summary_v1.json",
          "bytes": 203808,
          "sha256": "76ea90f85890dc538d52aa7f00dad2f0ee2b6fc46e6445043fb2101cf6fa7d23"
        },
        {
          "path": "analysis/ledger_480x832_v1/k2_coverage_v1.json",
          "bytes": 6540,
          "sha256": "709f59e71d9f1c2ffa5f0a5b9cb36331f8a99358322ebe10bc13d645ca3d3b67"
        },
        {
          "path": "analysis/ledger_704x1280_v1/k2_coverage_v1.json",
          "bytes": 6586,
          "sha256": "492a873103145c3e76fe401b1a3577d07bbba8d65d056325f2e74a287da69045"
        },
        {
          "path": "analysis/full_dispatch_480x832_v1_model_invariance_v1.json",
          "bytes": 40998,
          "sha256": "5238bbf585396ff0fd68c9c80f1d7b708e87bf1e4df13882c5426bae3519c99d"
        },
        {
          "path": "analysis/full_dispatch_704x1280_v1_model_invariance_v1.json",
          "bytes": 35257,
          "sha256": "1764b29a642e1a2f5fe990f92b04be86c7577ddb00f152ddbde7e516240345fe"
        }
      ]
    },
    {
      "id": "K2",
      "requirement": "逐组硬件指标与瓶颈反证",
      "status": "部分完成／已恢复采集",
      "finding_and_remaining": "进程级授权与普通用户调度实测兼容；18个GEMM/VAE/转换真实输入计数器及同卡FA2/FA3对照通过审核。完整逐组覆盖、硬件反证及全模型MFU仍缺。 新增13种真实输入/48个kernel已审核；现有31种算子代表输入覆盖此480×832 TP1/SP4逐调用记录的12种线性、17种卷积与2种转换结构。仍不代表其他配置、所有实际调用或全模型MFU。 高分辨率新增28个代表输入、90个kernel；两分辨率合计59个算子输入覆盖各自所选线性/卷积/转换结构。PID2019660有未归因的调度备注，保留原失败和后续带备注审核。两分辨率逐调用元数据均已补充，仍非与旧nsys同次运行映射。 新增短KV、长KV及cross四组FA2/FA3共8次NCU均通过数值检查，并按输入SHA256匹配已有未插桩回放；两种计时独立呈现，不推导整模型加速。 新增全模型标称FLOPs与模型级MFU（同一运行关键路径为分母）：480x832 DiT 49.0%（BF16 dense）、VAE 14.5%（TF32 dense）、端到端 40.2%；704x1280 DiT 53.1%（BF16 dense）、VAE 15.9%（TF32 dense）、端到端 46.6%。标称口径，非硬件执行FLOPs。 新增69张逐组瓶颈卡（unresolved_no_dominant_limiter 17, compute_bound 29, partial_wave_tail 13, dram_bandwidth_bound 4, l2_bandwidth_bound 6）与账本分组归因：480x832 计数器归因68%/通信12%/推导6%/未归因14%；704x1280 计数器归因78%/通信7%/推导4%/未归因11%。稳态KV长度FA3核心按阈值规则无主导单元，stall/tensor-pipe/L2重采已排队；compiled融合kernel与paged gather无计数器；全模型MFU仍缺。",
      "evidence": [
        {
          "path": "raw/ncu_attention_fa3_pilot_v1/selected_metrics.json",
          "bytes": 6369,
          "sha256": "4fd6985f90f709fbf09c9dbdfc3231ef58bfe5b99a3141670943071c54ee76a4"
        },
        {
          "path": "analysis/capability_counter_probe_audit.json",
          "bytes": 1567,
          "sha256": "aa52e4453e87082f277b07fd82d7f66f3624b2b0f01c6a390c13a0cf5e23c2d6"
        },
        {
          "path": "analysis/capability_operator_ncu_full_audit.json",
          "bytes": 102168,
          "sha256": "213f70d59c3357a7d3179a9d367cfe7a1ea050620462455b91c2f2bbb228dc13"
        },
        {
          "path": "analysis/capability_attention_pair_audit.json",
          "bytes": 7892,
          "sha256": "2d81ace9f64b6d42c2a0bce1f5b6061289792854b14fbcbf2c7548250b1ed7bf"
        },
        {
          "path": "analysis/bottlenecks.json",
          "bytes": 395567,
          "sha256": "26174fe903b677be135571ba9297d8a6c341886edd4550814316a5bcf3911439"
        },
        {
          "path": "analysis/missing_operator_ncu_audit_v1.json",
          "bytes": 99260,
          "sha256": "1cd20f22b13dc6e63fd8ed5890420ce4a45db249f756f777f977848cf494bed8"
        },
        {
          "path": "analysis/operator_replay_signature_coverage_v2.json",
          "bytes": 56827,
          "sha256": "0d5f81feaa1496de662db1ceaed4524e57756f5644b50702288424c0335eab4b"
        },
        {
          "path": "analysis/recorded_model_mac_work_v1.json",
          "bytes": 3668,
          "sha256": "3e460d76891f06a420be84135f125309c0d81c41d3c98252d941f1e9e5199076"
        },
        {
          "path": "analysis/missing_operator_704_ncu_audit_v1.json",
          "bytes": 195620,
          "sha256": "f4401c75f2607b2d513134edc6455173e9568ed165ea3f4b8634d01e329f7a6e"
        },
        {
          "path": "analysis/operator_replay_signature_coverage_704_v2.json",
          "bytes": 58361,
          "sha256": "471d764af2767257dc14ebcf3642159a46ed4d886a3c46474cdd68c84661edff"
        },
        {
          "path": "analysis/operator_metadata_704_provenance_audit_v1.json",
          "bytes": 1158,
          "sha256": "e829b7a37116257d8144d0f073528d5354419e9962a252c0d8653fc32ce2784e"
        },
        {
          "path": "analysis/resolution_recorded_work_comparison_v1.json",
          "bytes": 2053,
          "sha256": "ccecb9622ce9d1aaf914af2afce9feed5322176a78aebe14b432141e5eaf7ac6"
        },
        {
          "path": "analysis/attention_hardware_extension_summary_v1.json",
          "bytes": 4462,
          "sha256": "482460b84c3e89bf4fa276db754dd65115a103d4437a16bf5e9a1a7d7a4a7544"
        },
        {
          "path": "analysis/attention_counter_context_v1.json",
          "bytes": 11348,
          "sha256": "6d7cad4d95ac7384331809534168a6805425ad050e353e0d7bc9603c3eba300f"
        },
        {
          "path": "analysis/model_mfu_480x832_v1.json",
          "bytes": 2003,
          "sha256": "98c3a02d1283d334fec800da4813d00bf9eea35fc33caf2a5302cd8d63786407"
        },
        {
          "path": "analysis/model_mfu_704x1280_v1.json",
          "bytes": 2004,
          "sha256": "9883274059ff2d5500a22474413edda185ab1d549a29052021ca262c59f50a19"
        },
        {
          "path": "analysis/bottleneck_cards_v1.json",
          "bytes": 307792,
          "sha256": "427497e23d4fd8d6d1c4e07686d8f69c1529dc36e906993c1dd7742b03dba224"
        },
        {
          "path": "analysis/ledger_480x832_v1/bottlenecks_ledger_v1.json",
          "bytes": 2259365,
          "sha256": "c1f69a5623d6edc210d038c72761e0cd84d7dcbaffa2c8f2cf7db390b04caa84"
        },
        {
          "path": "analysis/ledger_704x1280_v1/bottlenecks_ledger_v1.json",
          "bytes": 2214173,
          "sha256": "42a432aa19ca61482904b3d046958061ff90f9d148d2f58f4a5a7e7904430a84"
        }
      ]
    },
    {
      "id": "A0",
      "requirement": "真实Shape×Backend矩阵",
      "status": "部分完成",
      "finding_and_remaining": "12配置代表输入已覆盖，1004有效回放、4个math OOM；不是全层覆盖，MFU图尚缺。 新增短KV、长KV及cross四组FA2/FA3共8次NCU均通过数值检查，并按输入SHA256匹配已有未插桩回放；两种计时独立呈现，不推导整模型加速。 新增Attention有效FLOP利用率曲线：1004个聚合点，分母为规格BF16 dense 989.5 TF/s，另标同卡经验上限；回放GPU未逐卡校准，不是整模型MFU。",
      "evidence": [
        {
          "path": "analysis/matrix_coverage.json",
          "bytes": 18072,
          "sha256": "b75fb39cd390ebca0e2e41cd91114699608c70f93495ad8d376e1bcee225ef0c"
        },
        {
          "path": "analysis/fa3_interface_dispatch_comparison.json",
          "bytes": 512239,
          "sha256": "d4d3438424f9aaa5413e98302b0f8a8111654f1e8187b7a9edeb24d00570bb24"
        },
        {
          "path": "analysis/attention_hardware_extension_summary_v1.json",
          "bytes": 4462,
          "sha256": "482460b84c3e89bf4fa276db754dd65115a103d4437a16bf5e9a1a7d7a4a7544"
        },
        {
          "path": "analysis/attention_counter_context_v1.json",
          "bytes": 11348,
          "sha256": "6d7cad4d95ac7384331809534168a6805425ad050e353e0d7bc9603c3eba300f"
        },
        {
          "path": "analysis/attention_mfu_points_v1.json",
          "bytes": 618990,
          "sha256": "7c7d610f064a7499298bb39a53560b9b6269591120358eb4e57212476d105e1f"
        }
      ]
    },
    {
      "id": "G0",
      "requirement": "其他重要kernel图谱",
      "status": "部分完成",
      "finding_and_remaining": "18个真实输入有NCU、shape和标称FLOPs/DRAM流量图；新增同卡三启动连续访问带宽参考和NCU流量核对；同卡计算27点和12个同卡算子经验参考图已补充；完整模型映射、L2口径及完整Roofline仍缺。 新增13种真实输入/48个kernel已审核；现有31种算子代表输入覆盖此480×832 TP1/SP4逐调用记录的12种线性、17种卷积与2种转换结构。仍不代表其他配置、所有实际调用或全模型MFU。 小工作集三启动Graph配对和NCU流量已审核；Graph平均调用开销与普通提交显著不同，存在DRAM流量，仍不能认领纯L2峰值。 高分辨率新增28个代表输入、90个kernel；两分辨率合计59个算子输入覆盖各自所选线性/卷积/转换结构。PID2019660有未归因的调度备注，保留原失败和后续带备注审核。两分辨率逐调用元数据均已补充，仍非与旧nsys同次运行映射。 新增全模型标称FLOPs与模型级MFU（同一运行关键路径为分母）：480x832 DiT 49.0%（BF16 dense）、VAE 14.5%（TF32 dense）、端到端 40.2%；704x1280 DiT 53.1%（BF16 dense）、VAE 15.9%（TF32 dense）、端到端 46.6%。标称口径，非硬件执行FLOPs。 新增DRAM层Roofline（67个点，含Attention NCU）；L2层计数器未采集，重采已排队；完整模型映射改由同一运行账本承担。",
      "evidence": [
        {
          "path": "analysis/operator_replay_v2_audit.json",
          "bytes": 89771,
          "sha256": "f2daaae4ccf2a02d7e865a596e28169322898abbe1b21266a68dd72c5b8ea0a1"
        },
        {
          "path": "analysis/local_gemm_calibration_audit.json",
          "bytes": 29909,
          "sha256": "7eb630888171ab6636ef4a3ee8399c66d2493f471afda6b003652e719eb9c0d4"
        },
        {
          "path": "analysis/operator_work_traffic.json",
          "bytes": 39000,
          "sha256": "87c4c6d61626f3be61835643966ad4e0aefc75d04a7980a36748f114975758c6"
        },
        {
          "path": "analysis/local_bandwidth_audit.json",
          "bytes": 24690,
          "sha256": "2478941ab4958ba7f56312ec3df6e17a93a3c6d1bd6a3ef43e356b812602b87d"
        },
        {
          "path": "analysis/local_bandwidth_ncu_audit.json",
          "bytes": 3216,
          "sha256": "c5c9b85b8c4aa8511faf0ea664da8eddababeb9a68abb19293989f814b1584e2"
        },
        {
          "path": "analysis/same_gpu_reference.json",
          "bytes": 12724,
          "sha256": "51b6fffe05d7268953932407000c9053e5361c338d1207276aba413e4db360b0"
        },
        {
          "path": "analysis/missing_operator_ncu_audit_v1.json",
          "bytes": 99260,
          "sha256": "1cd20f22b13dc6e63fd8ed5890420ce4a45db249f756f777f977848cf494bed8"
        },
        {
          "path": "analysis/operator_replay_signature_coverage_v2.json",
          "bytes": 56827,
          "sha256": "0d5f81feaa1496de662db1ceaed4524e57756f5644b50702288424c0335eab4b"
        },
        {
          "path": "analysis/recorded_model_mac_work_v1.json",
          "bytes": 3668,
          "sha256": "3e460d76891f06a420be84135f125309c0d81c41d3c98252d941f1e9e5199076"
        },
        {
          "path": "analysis/warm_graph_pair_audit_v1.json",
          "bytes": 44446,
          "sha256": "6b3744076de4242a458635a2eaf133c6fb803b55a5cb55538e2d0ef6f2554535"
        },
        {
          "path": "analysis/warm_working_set_ncu_audit_v1.json",
          "bytes": 3221,
          "sha256": "47c15a59b4314baba3a7b5682fc540418f06192097e68442f698012a6c86fdc7"
        },
        {
          "path": "analysis/missing_operator_704_ncu_audit_v1.json",
          "bytes": 195620,
          "sha256": "f4401c75f2607b2d513134edc6455173e9568ed165ea3f4b8634d01e329f7a6e"
        },
        {
          "path": "analysis/operator_replay_signature_coverage_704_v2.json",
          "bytes": 58361,
          "sha256": "471d764af2767257dc14ebcf3642159a46ed4d886a3c46474cdd68c84661edff"
        },
        {
          "path": "analysis/operator_metadata_704_provenance_audit_v1.json",
          "bytes": 1158,
          "sha256": "e829b7a37116257d8144d0f073528d5354419e9962a252c0d8653fc32ce2784e"
        },
        {
          "path": "analysis/resolution_recorded_work_comparison_v1.json",
          "bytes": 2053,
          "sha256": "ccecb9622ce9d1aaf914af2afce9feed5322176a78aebe14b432141e5eaf7ac6"
        },
        {
          "path": "analysis/model_mfu_480x832_v1.json",
          "bytes": 2003,
          "sha256": "98c3a02d1283d334fec800da4813d00bf9eea35fc33caf2a5302cd8d63786407"
        },
        {
          "path": "analysis/model_mfu_704x1280_v1.json",
          "bytes": 2004,
          "sha256": "9883274059ff2d5500a22474413edda185ab1d549a29052021ca262c59f50a19"
        },
        {
          "path": "analysis/operator_roofline_v1.json",
          "bytes": 64857,
          "sha256": "7df66e54ab4149d51b622d9b6e9eae1a1e7688373fac0791f165750a5a835b2d"
        }
      ]
    },
    {
      "id": "G1",
      "requirement": "整模型优化验收",
      "status": "未通过",
      "finding_and_remaining": "cuDNN cross候选40份输出均未通过门槛；尚无接受的整模型加速。 新SDPA Flash候选的6个实际实现输入已完成独立数值和kernel审核；该结果不能替代整模型验收。 本次检查未确认模型任务仍存活，不能仅凭历史状态认定在运行。 新候选保存输出数值检查未通过；单次A/B不构成稳定性能验收。",
      "evidence": [
        {
          "path": "analysis/cross_candidate_eager_r0_comparison.json",
          "bytes": 43280,
          "sha256": "489b316b15752136c50d0fad9825806de4f62111dd5683c545b27f47382e80da"
        },
        {
          "path": "analysis/execution_control_pilot_timings.json",
          "bytes": 5896,
          "sha256": "3643831214810f3467fc2be23ddc7ad6406e0bfecdee3f815d5d69efe79a7d59"
        },
        {
          "path": "analysis/cross_sdpa_flash_impl_audit.json",
          "bytes": 6795,
          "sha256": "358c219afb2ca0acd290c83341746627d419f5fc9a5f95c9cbc0740cc9f8f397"
        },
        {
          "path": "raw/cross_sdpa_reallocation_v2.json",
          "bytes": 1151,
          "sha256": "8f201f8f86642f3a836ebc3f1f0f5d2d12b93c7c712c1cf4a7c33c8781a11ca3"
        },
        {
          "path": "analysis/cross_sdpa_model_comparison_v2.json",
          "bytes": 42896,
          "sha256": "e6f2490a498cf3df1dc68bd14de73aa7943eae1b6a90cd3527772a0409fee3e6"
        },
        {
          "path": "analysis/cross_sdpa_model_audit_v2.json",
          "bytes": 2249,
          "sha256": "49d560f056a1c9a1e95bc0a132cc4403a9ac60931f4db11eadc27fb5ee291c9a"
        }
      ]
    },
    {
      "id": "H0",
      "requirement": "参考优化适用性",
      "status": "部分完成",
      "finding_and_remaining": "已做本地源码映射；所列外部PR/RFC状态及本地提交祖先关系已追加核验，完整编译融合及transport运行关联仍缺。",
      "evidence": [
        {
          "path": "analysis/optimization_transfer_matrix.json",
          "bytes": 8194,
          "sha256": "233957ec847494043154e504e65f383b8e7967e594d0c21d75f6bfc906c0c609"
        },
        {
          "path": "analysis/h3_external_reference_review.json",
          "bytes": 2852,
          "sha256": "f8fa4ee765e9205ea02e234068fbd1158fa7099bfd7f909fec8b645e6a8bfba6"
        },
        {
          "path": "analysis/h3_reference_local_ancestry.json",
          "bytes": 1446,
          "sha256": "96abca00f5513e2447dfed2dff0b68dc27f9db7695c30eef314d0fb7a15ec0ce"
        }
      ]
    },
    {
      "id": "H1",
      "requirement": "Attention与通信布局消融",
      "status": "未完成",
      "finding_and_remaining": "独立后端回放不能替代通信布局的单变量A/B。",
      "evidence": [
        {
          "path": "analysis/optimization_transfer_matrix.json",
          "bytes": 8194,
          "sha256": "233957ec847494043154e504e65f383b8e7967e594d0c21d75f6bfc906c0c609"
        }
      ]
    },
    {
      "id": "H2",
      "requirement": "DiT融合与局部数据构造消融",
      "status": "未完成",
      "finding_and_remaining": "尚无通过数值门槛的融合增量收益验证。",
      "evidence": [
        {
          "path": "analysis/optimization_transfer_matrix.json",
          "bytes": 8194,
          "sha256": "233957ec847494043154e504e65f383b8e7967e594d0c21d75f6bfc906c0c609"
        }
      ]
    },
    {
      "id": "H3",
      "requirement": "解码算子与空间并行消融",
      "status": "未完成",
      "finding_and_remaining": "新增含全部转换成本的局部布局A/B：数值逐值一致但慢约22.42%；仅单输入局部反证，完整解码优化消融仍缺。 新增时间维卷积全成本布局A/B同样逐值一致但慢14.82%；布局转换51.9%时间占比未直接转化为收益。",
      "evidence": [
        {
          "path": "analysis/optimization_transfer_matrix.json",
          "bytes": 8194,
          "sha256": "233957ec847494043154e504e65f383b8e7967e594d0c21d75f6bfc906c0c609"
        },
        {
          "path": "analysis/missing_vae_components_v1.json",
          "bytes": 21278,
          "sha256": "dd125c18588b573223bc007c4c65c77394a13ea0065c97270144040c6de69f8d"
        },
        {
          "path": "analysis/vae_layout_temporal_ab_audit_v1.json",
          "bytes": 1852,
          "sha256": "e8fd62d86e734f4e078aa865f95de9f99c6b4756ac40c1e16d03e08b13049ba5"
        }
      ]
    },
    {
      "id": "H4",
      "requirement": "输出与传输优化",
      "status": "未完成",
      "finding_and_remaining": "已有uint8路径与交接计时，尚无受控输出优化及编码后媒体验收。 已补充六阶段生产者/消费者表：冻结路径只回传元数据，不含像素IPC；CPU-ready不能替代编码或客户端交付。输出融合真实输入补采已提交，尚未验收。 新增固定FFV1无损CPU编码往返，3份视频和351帧逐字节审核通过；仍非模型服务集成、像素IPC或有损质量验收。 交接终态：输出补采与局部融合A/B已完成；40模型输出、20输入转换与775边界字节一致，D2H-inclusive局部耗时降低23.76%/31.84%；未集成模型。用户要求停止新增实验并交接。",
      "evidence": [
        {
          "path": "analysis/clean_handoff_summary.json",
          "bytes": 20802,
          "sha256": "8539d05aa1230bfa088d539d026cc34b2a143a517fa2d9e836e07adf6d46c879"
        },
        {
          "path": "analysis/optimization_transfer_matrix.json",
          "bytes": 8194,
          "sha256": "233957ec847494043154e504e65f383b8e7967e594d0c21d75f6bfc906c0c609"
        },
        {
          "path": "analysis/output_boundary_inventory_v1.json",
          "bytes": 3145,
          "sha256": "a8516cddcd4b8ab166ad71079d36222b72f68eab05ee2ee3d5133acf5b20a67b"
        },
        {
          "path": "raw/output_conversion_protocol_v1.json",
          "bytes": 779,
          "sha256": "664a8597732da6ff6dba45265852892c0ca672cabf9d5a36f19bf69feb0b0322"
        },
        {
          "path": "analysis/lossless_media_roundtrip_audit_v1.json",
          "bytes": 51358,
          "sha256": "64bd978fe17c57c9e1f646c3c6036f6cf5bb33a195ba388fdbaad6628c4f080f"
        },
        {
          "path": "analysis/output_conversion_ab_audit_v1.json",
          "bytes": 9823,
          "sha256": "7af0bbf21d0827fc60f4329cd1a0673cd6a09ee10d4e115d64a464aeb0e2b1b4"
        },
        {
          "path": "analysis/output_float_capture_model_invariance_v1.json",
          "bytes": 41121,
          "sha256": "db5681d5b4b0baa92c0a2acde67dce203eecf6b6f55378b16c8cc46ac2961c86"
        },
        {
          "path": "HANDOFF_CN.md",
          "bytes": 14355,
          "sha256": "9f0f7b09021d5168d50de592d838907d474da2079d69f0ebc6e3b8c7dd985280"
        },
        {
          "path": "HANDOFF_STATE.json",
          "bytes": 1746,
          "sha256": "17c71fd4c5acb0e611f40278bd890fe0716349fa7466d1165498f3db9d4dd875"
        }
      ]
    },
    {
      "id": "H5",
      "requirement": "流水线重叠、背压与取消",
      "status": "未完成",
      "finding_and_remaining": "时间线重叠不等于优化收益；慢消费者、取消和组合调度验收尚缺。",
      "evidence": [
        {
          "path": "analysis/worker_chunk_activity.json",
          "bytes": 6911,
          "sha256": "1c92a4f660df9af27ebd73981958020ddf9071a69332229200ea8da6ef68fd25"
        },
        {
          "path": "analysis/optimization_transfer_matrix.json",
          "bytes": 8194,
          "sha256": "233957ec847494043154e504e65f383b8e7967e594d0c21d75f6bfc906c0c609"
        }
      ]
    },
    {
      "id": "TAIL",
      "requirement": "重复运行与尾延迟",
      "status": "部分完成",
      "finding_and_remaining": "两分辨率各3次、每次1000间隔；新增连续会话块重采样显示部分P99区间敏感，数值控制与未来稳定性仍未完成。",
      "evidence": [
        {
          "path": "raw/tail_timing_audit_workflow_v1.json",
          "bytes": 6259,
          "sha256": "53acde1c9876a32f2298ba7e87e92c71c0328cd035324771b34022454d8b5819"
        },
        {
          "path": "analysis/clean_handoff_summary.json",
          "bytes": 20802,
          "sha256": "8539d05aa1230bfa088d539d026cc34b2a143a517fa2d9e836e07adf6d46c879"
        },
        {
          "path": "analysis/tail_six_launch_hash_repeatability.json",
          "bytes": 39974,
          "sha256": "36b2229ab13e92989bd38501ba0fc2a2825a60552c7078204214a00dde260259"
        },
        {
          "path": "analysis/tail_block_sensitivity.json",
          "bytes": 20642,
          "sha256": "3ecfcda5cb0766e29a565d01f10753a23c7c0c3bac316b6da7172d78ffa0fb8c"
        }
      ]
    },
    {
      "id": "NUM",
      "requirement": "数值可重复性与候选门槛",
      "status": "部分完成",
      "finding_and_remaining": "eager两次40份输出逐位一致；两种编译模式重复及cuDNN候选均失败，原因未最终定位。 新参考与历史40份逐值一致；原版和候选两条轨迹各160次同输入比较及FP64抽样通过，诊断最终输出各40份与对应未插桩运行逐值一致。跨轨迹Q从第2次出现差异，仍缺完整逐层因果证明。 新增分步诊断：default编译模式（无CUDA Graph）两次启动0/40通过；reduce-overhead加库级确定性控制（cuBLAS工作区、cuDNN确定性、use_deterministic_algorithms）两次启动40/40通过、40/40逐位一致：跨启动差异定位为库级非确定性而非编译/Graph。单一控制拆分与性能代价计时排队中。",
      "evidence": [
        {
          "path": "analysis/execution_control_eager_repeat_comparison.json",
          "bytes": 41449,
          "sha256": "cd9d6484db2413a297e93d0c32f099b22ca348aa644ef9440e79b21a5f5017d4"
        },
        {
          "path": "analysis/execution_error_progression.json",
          "bytes": 7549,
          "sha256": "2740e7ee34dbfd861e395592e5077d8a688c5367736cb421e8bb3101f1b99995"
        },
        {
          "path": "analysis/cross_sdpa_fresh_reference_repeat_v2.json",
          "bytes": 41203,
          "sha256": "fa7065b24a5f3302dfeaadb2b43b3b0965ba2aefb3bb925f2b9f7392af60013e"
        },
        {
          "path": "analysis/cross_sdpa_shadow_audit_v1.json",
          "bytes": 174462,
          "sha256": "21a655e4a6cc0196f110f1c10b4f9c44e5249301e1ad0d77483e8d59464ee8fe"
        },
        {
          "path": "analysis/cross_sdpa_candidate_shadow_audit_v1.json",
          "bytes": 174521,
          "sha256": "471098761fde9698e45968ca96427db35b31dce7068220b038855d844e98a4c1"
        },
        {
          "path": "analysis/cross_shadow_trajectory_comparison_v1.json",
          "bytes": 168109,
          "sha256": "ce2e327d0d6ad8f44f6fce608591f42d810c03c9972190844d574118e755a513"
        },
        {
          "path": "analysis/compile_bisect_A_default_mode_repeat_comparison_v1.json",
          "bytes": 43870,
          "sha256": "2b1e2e606035334058cbe59c8fa5069ec30e0f4299f240b898b95976c875dfbd"
        },
        {
          "path": "analysis/compile_bisect_B_reduce_overhead_deterministic_repeat_comparison_v1.json",
          "bytes": 43335,
          "sha256": "78a6687f740848fdb9c43abfdcee798d12414bac27d6dd69128d3b0b559e4940"
        },
        {
          "path": "raw/compile_bisect_workflow_v1.json",
          "bytes": 6251,
          "sha256": "9f71c25df686f0051a99b332a653af7dfe0051bc2b4505feceb0a1fd63997c4b"
        }
      ]
    },
    {
      "id": "RAW",
      "requirement": "全量原始数据与复现",
      "status": "部分完成",
      "finding_and_remaining": "已存在文件有索引、源快照与复算脚本；长尾后续轮次未保存完整输出，无法补回。",
      "evidence": [
        {
          "path": "analysis/delivery_archive_audit_20260910T0020Z.json",
          "bytes": 597,
          "sha256": "c99041e6e21008591abbf5f3abdf79048615965e2780590b57c281ebea81c1b4"
        },
        {
          "path": "REPRODUCE_CN.md",
          "bytes": 24979,
          "sha256": "2daee1e396c9f829e3b950426ab5d49b3007766fbebdcba430027b9d0ba851be"
        },
        {
          "path": "analysis/raw_data_index.json",
          "bytes": 10666269,
          "sha256": "6c5ce1fb000ec6a43ddc0b53c7df929b3bcf9710500cb5681d60be204f533348"
        }
      ]
    },
    {
      "id": "WEB",
      "requirement": "首轮报告与证据访问",
      "status": "首版已交付／持续更新",
      "finding_and_remaining": "中文报告及看板已发布，82个报告内资源链接通过内容核对；不代表完整计划完成。",
      "evidence": [
        {
          "path": "REPORT_CN.md",
          "bytes": 203415,
          "sha256": "02dc3524839397b98e17aeeb88d42ad5ff49cd017fca1f3fa660ecb479eade79"
        },
        {
          "path": "analysis/public_report_asset_audit.json",
          "bytes": 39245,
          "sha256": "5112ee1ab42c5a7a862217e48d2685513f558a3fd971d66dacedbbabaf9e613a"
        }
      ]
    }
  ]
}
