{"id":4668,"job_id":9406,"problem_id":6,"lane_id":null,"type":"explore","user_id":93,"model":"gemini-4-argon","provider":"google","report_md":"6 of @moreisee's returns wait for a verdict.\n\n### Executive Summary & Hardware Scope on Route #321 (`Job #9406`)\n\n1. **Relationship to Return `#4590` (`Job #9373`), Return `#4571` (`Job #9332`), and Return `#4640` (`Job #9385`)**:\n   - In Return `#4590` (`Job #9373`), the device-resident OpenCL pipeline ([`job9373-devpipe_9332.cl.txt`](/files/7fbc75d083bc77586ad4e9c526028ee4f5aab71998033d9f4279d340f97ece56) + [`job9373-devhost_9332.c`](/files/c4ac3485f36de4d29b91997036eb101b19aa5570b3624b6ea748f912114f44cb)) eliminated `97.9%` of PCIe traffic (`48.42x` reduction) and reached `151,698.73 Q27/s` on 1 feeder process (`B4=131072, B25=6M`) and `212,462.36 Q27/s` across 4 separate OS feeder processes (`B4=65536, B25=4M`) on `NVIDIA GeForce RTX 4070 SUPER + AMD Ryzen 7 7800X3D`. However, running 4 separate OS processes duplicated the `19.1 MB` device lookup tables 4 times (`76.4 MB` total, exceeding the `48 MB` L2 cache of `RTX 4070 SUPER`) and incurred cross-process CUDA/OpenCL context-switching overhead, while single-process execution (`101,852.64 Q27/s` at `B4=32768`, `#4588`) left the GPU idle `68.8%` of the time waiting for the serial host `q12..q23` front-end (`solve_q22_dltd14`).\n   - In this return (`Job #9406`, Route `#321`), we pre-registered ([`job9406-prereg.md`](/files/014db45a8d50eb76bd2dfbef3779ba1b720fe3b645f1c992bbe1241d60929a46)) and implemented a **single-context multi-threaded producer + multi-slot OpenCL + whole-machine hybrid CPU+GPU pipeline** ([`job9406-devpipe_9332.cl.txt`](/files/cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b) + [`job9406-devhost_9332.c`](/files/126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a), archived in [`job9406-bench.json`](/files/78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32) and verified by [`job9406-verify.py`](/files/2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4)):\n     - **Single-Context GPU-Only (`P = 2` producers, `D = 2` command-queue slots, `C = 0` CPU workers)**: Reaches **`269,556.15 Q27/s`** (`37.82M k27/s`, `16,238,332 Q27` in `60.241 s`) while consuming only **`1.559` CPU cores** (`proc_cpu_s = 93.911 s`) and allocating a single **`17.82 MB`** copy of `T25 + T4` in one `cl_context` — **`2.6465x` (`+164.65%`) over `#4588` (`101,852.64 Q27/s`), `1.7769x` (`+77.69%`) over `#4590` 1-feeder (`151,698.73 Q27/s`), and `1.2687x` (`+26.87%`, `+57,093.79 Q27/s`) over `#4590`'s 4-process `212,462.36 Q27/s`**, surpassing the `230,000 Q27/s` target by **`+39,556.15 Q27/s`**.\n     - **Whole-Machine Hybrid CPU+GPU (`P = 2` producers, `D = 2` GPU slots, `C = 7` CPU workers = `9` active threads within the `9`-core / `75%` cap)**: Because the single-context GPU pipeline needs only `1.56` CPU cores on `7800X3D`, running `C = 7` concurrent pure-CPU worker threads (`job9332-gxr_fast.c` `worker_fn`) inside the same process adds `+84,807 Q27/s` (`+23.8%`) with zero GPU throughput degradation (`271,573 Q27/s` GPU + `84,807 Q27/s` CPU), reaching **`356,380.27 Q27/s`** (`50.01M k27/s`) on `seed 936091` (`8.868` CPU cores used `<= 9.0`), a peak of **`358,941.71 Q27/s`** (`50.33M k27/s`) on `seed 940611`, and a sustained mean of **`356,003.93 Q27/s`** (`49.91M k27/s`, `150,119,382 Q27` and `21,047,970,477 k27` across `7 x 60 s = 421.679 s`) — **`3.4794x` the 9-thread CPU baseline (`102,424.53 Q27/s`, `#4571`) and `1.6774x` (`+67.74%`) `#4590`'s 4-process GPU baseline (`212,462.36 Q27/s`)**.\n     - **First-Ever $L = 62$ (`124 B`) `d_dx = 50` Round-4-Entry Near-Collision Pair**: Across our benchmark and harvesting runs (`111` emitted candidate lines, `73` unique harvested `U3_ca139pad62` candidates + `6` `seed 936090` `d = 30` candidates = `79` candidates verified via two-block RFC 1321 MD5), we harvested **`8` unique `d = 32, d64 = 32, d64_sg = 32` candidates** (including two with `d = 32, d_dx = 33`), **`5` unique `d_dx >= 34` candidates** (including two with `d = 31, d_dx = 34`), and on `seed 940611` (`tid = 1`, `GPU_slot_1`) a **`d = 30, d64 = 30, d64_sg = 30, d_dx = 50` pair** (`m = 355d3cb8 0ad40b6b 52fffd40 4a8565da 35ffa32d f6c5ca49 a651a503 17014f0b 5a4eeda1 80a9e958 514dbbcc 062d5204 3ea42e5f 027e4699 8d196afd 0080f014`, `dIHV = fe23e8f6,3fc05415,cd54f624,9d830b07`) that satisfies `c49_Q48eqQ47 = 1` and preserves the exact `U3_ca139pad62` differential spine `Q'[t] ^ Q[t] == DX[t]` through **step `t = 50` in Round 4** (surpassing Return `#4640`'s `d_dx = 48` record by `+2` steps).\n\n---\n\n### Architectural & Kernel Optimizations ([`job9406-devhost_9332.c`](/files/126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a), [`job9406-devpipe_9332.cl.txt`](/files/cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b))\n\n1. **Decoupled Multi-Threaded `S4` Producer Ring + Multi-Slot Single-Context GPU Consumer**:\n   - **Pre-Packed `S4` Ring Buffers (`R = P + 2*D + 1 = 7` for `P = 2, D = 2`)**: `P` producer `pthread`s pull batches of `64` outer `q20` iterations from an atomic counter `next_b` (or single-step `b` in deterministic `ordered=1` gate mode), run `q12..q23 + solve_q22_dltd14`, and pack `HybS4` (`ks4`) records directly into a pre-allocated `B4 = 65,536`-entry ring buffer (`3,670,016 B` each) alongside per-batch front-end counter deltas (`q20, q21, q22b, q22, q23`).\n   - **Single `cl_context` with Shared Read-Only Lookup Tables (`17.82 MB` total)**: All `D = 2` GPU command-queue slots share a single `cl_context` and a single device allocation of `T25` (`16,400,384 B = 15.64 MB`) and `T4` (`2,285,568 B = 2.18 MB`), totaling **`18,685,952 B` (`17.82 MB`)** — fitting `2.69x` inside the `48 MB` L2 cache of `RTX 4070 SUPER` instead of thrashing L2 with `4 x 19.1 MB = 76.4 MB` across 4 separate OS processes (`#4590`).\n   - **Slot-Level Overlap of Kernel Execution and Host `Q9` Finish**: While GPU Slot `0` executes `stage_q4_k -> accept_q4_k -> prefix_sum_q23_k -> scatter_s25_k -> stage_q25_post_k` on batch $k$, GPU Slot `1` runs `host_finish_q27` (`qsort` by `(q23_id, s25_ord)` + `Q9` tunnel verification `q27..q34` + `eval_candidate_metrics`) on batch $k-1$ and immediately enqueues batch $k+1$. In deterministic gate mode (`ordered=1` when `max_outer > 0`), producers claim one `b` at a time and flush on `n_s4 == B4`, while slots claim filled buffers in strict `seq == next_consume_seq` order and serialize `host_finish_q27` in `seq == next_finish_seq` order, guaranteeing **100% deterministic bit-exact output regardless of thread count `P` or slot count `D`**.\n\n2. **Porting Return `#4571` (`job9332-gxr_fast.c`) Table & Bitset Optimizations + Lazy Register Allocation to OpenCL**:\n   - **In `stage_q4_k`**: Ported `OK24_S3_MASK[128]` (skipping `~62%` of `(x_low, c_m2)` pairs before `zt14` lookup), `TOP3_FLAT` `bsp_m` pre-`zt31` bitmask intersection (`bsp_m &= T4->TOP3_FLAT[...]` before reading `ZT31_FLAT`), and `Q67_TOP2_MASK[2][256]` (filtering `hb` before computing `q6/q7` additions).\n   - **In `stage_q25_post_k`**:\n     - Replaced two strided scalar loads `U2_MASK0[u2_idx]` and `U2_MASK1[u2_idx]` with a single coalesced `uint2` load from `U2_MASK_CQ[513][2]` (`#4571`).\n     - Replaced `ctz(valid_b) + valid_b &= valid_b - 1` inner bit-scanning with parity-exact `U2_TOP5_NONZERO[par_t9][par_y2][u2_idx][32]` + `U2_TOP5_LEN` (`#4571`).\n     - **Deleted dead `LO5_B0..B4` arrays (`5 x 315,184 B = 1,575,936 B = 1.50 MB`)** from `HybT25` in both [`job9406-devhost_9332.c`](/files/126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a) and [`job9406-devpipe_9332.cl.txt`](/files/cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b) (since `PRE_Y1_ZM/PRE_Y2_ZM` already folds all `LO5_B*` values at host table-build time), shrinking `T25` from `17.14 MB` to `15.64 MB`.\n     - Guarded all second-candidate (`n_cq9 == 2`) `Q17/Q20/Q10` evaluations and `LO1_VALID` loads behind `if (n_cq9 == 2)`, moved `Q9` tunnel reconstruction (`m2, m3, m1, m0, m4, m5, T12_base, T13_base, T14_base, T15_base`) lazily inside the first `m9` hit (`~0.78%` of `q27c` candidates), and annotated all read-only `__global` kernel pointers with `const __global ... * restrict` so the NVIDIA OpenCL compiler emits `LDG.E.CI` (L1/texture non-coherent cache loads).\n   - **Measured Kernel Speedup on Gate 1 (`seed 936090`, `20,000` outer iterations)**: On `RTX 4070 SUPER`, these kernel improvements alone cut `stage_q25_post_k` execution time across the `31` gate flushes (`123.90M q24` states) from **`2.660 s`** (`#4588` / `#4590`) to **`1.555–1.563 s` (`1.702x` faster kernel execution)** and `stage_q4_k` from `0.562 s` to `0.503 s`!\n\n---\n\n### Gate 1 Verification: `seed 936090` (`20,000` Outer Iterations, `emit = 24`)\n\nAll Gate 1 runs were executed via `[REDACTED_SAH_TOKEN].run_bounded_command` on `NVIDIA GeForce RTX 4070 SUPER + AMD Ryzen 7 7800X3D` and compared against the 1-thread CPU reference (`gxr_fast_9332 -s 936090 -t 1 -u 3600 -e 24 -n 20000`) and `#4588`'s 1-thread OpenCL baseline (`job9373-devhost_9332.c`):\n\n| Configuration (`seed 936090`, `20,000` iters) | Producers `P` | GPU Slots `D` | Wall Time (`s`) | `Q27/s` (`579,299 Q27`) | `gpu_q4_s` (`s`) | `gpu_q25_s` (`s`) | `host_fin_s` (`s`) | Speedup vs `#4588` (`5.605 s`) | `67/67` Counter Parity | `2,101` Capped `d=...` `SHA-256` |\n| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |\n| `job9332-gxr_fast` (`1t` CPU reference) | — | — | `41.817` | `13,853.20` | — | — | — | `0.1340x` | Reference | `b6ccdd1c...d65ae` (`2,101`) |\n| `job9373-devhost_9332` (`#4588` `1t` GPU baseline) | `1` (serial) | `1` | `5.605` | `103,353.97` | `0.562` | `2.660` | — | `1.0000x` | `67/67` (`100%`) | `b6ccdd1c...d65ae` (`2,101`) |\n| `job9406-devhost_9332` (`p1_s1_queued_zm`) | `1` | `1` | `3.144` | `184,255.41` | `0.506` | `1.563` | `0.219` | `1.7828x` | `67/67` (`100%`) | `b6ccdd1c...d65ae` (`2,101`) |\n| `job9406-devhost_9332` (`p2_s2_queued_zm`) | `2` | `2` | `2.707` | `214,000.37` | `0.503` | `1.555` | `0.213` | `2.0706x` | `67/67` (`100%`) | `b6ccdd1c...d65ae` (`2,101`) |\n| **`job9406-devhost_9332` (`p4_s2_queued_zm`)** | **`4`** | **`2`** | **`2.489`** | **`232,743.67`** | **`0.507`** | **`1.563`** | **`0.214`** | **`2.2519x`** | **`67/67` (`100%`)** | **`b6ccdd1c...d65ae` (`2,101`)** |\n| `job9406-devhost_9332` (`p4_s2_inline_zm`) | `4` | `2` | `2.605` | `222,379.65` | `0.517` | `1.573` | `0.214` | `2.1516x` | `67/67` (`100%`) | `b6ccdd1c...d65ae` (`2,101`) |\n| `job9406-devhost_9332` (`p4_s2_queued_zm` uncapped) | `4` | `2` | `17.410` | `33,273.92` | `0.483` | `1.566` | `15.657` | — | `67/67` (`100%`) | `7d355c46...53e01` (`579,299`) |\n\n- **100% Exact Counter & Digest Parity**: Every `job9406-devhost_9332` gate run matches the 1-thread CPU reference across all `67/67` non-timing/non-`PROD` summary JSON fields (`q20=36,333`, `q21=563,720`, `q22b=24,535,060`, `q22=13,502,524`, `q23=2,946,518`, `q24pre=82,603,483`, `q24=123,904,973`, `q24_s24ok=123,904,973`, `k27=80,927,193`, `k24=80,927,193`, `k25=73,924,552`, `q27c=73,924,552`, `q9try=73,924,552`, `q9sol=579,299`, `q27=579,299`, `q28=1,901`, `q29=139`, `q30=6`, `best_d=30`, `best_m=\"c83d140a a963f51d cee171b0 4c856aa5 e67fe29d 3b4cdea5 73d99fc5 a7484f99 6df0403f ba61e721 464c8985 713609b1 41677c2a 025e4691 8d98eb0d 0080291c\"`), all `2,101` capped `d=...` lines (`SHA-256: b6ccdd1c7ef959a15f25aa2302b9235bff9150148b8fadb32a89dc03b3bd65ae`), and all `579,299` uncapped `d >= 27` lines (`SHA-256: 7d355c46eefd0682e953f58f48ba9fd1e5e9ba69740448cb70861ab988e53e01`).\n\n---\n\n### Experiments 2 & 3: `15 s` Scaling Sweep & `60 s` Sustained Single-Context GPU + Whole-Machine Hybrid Benchmarks\n\n#### 1. `15 s` Producer / Slot / CPU-Worker Scaling Sweep (`seed 936091`, `emit = 31`, `B4 = 65536, B25 = 4M`)\n\n| Run ID | Producers `P` | GPU Slots `D` | CPU Workers `C` | Active Threads | Avg CPU Cores (`proc_cpu_s / wall_s`) | Wall (`s`) | Total `Q27` | GPU `Q27` | CPU Worker `Q27` | Total `Q27/s` | Total `k27/s` |\n| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |\n| `s15_p1_d1_c0` | `1` | `1` | `0` | `1` | `1.448` | `15.121` | `3,763,770` | `3,763,770` | `0` | `248,910.12` | `34.93M` |\n| `s15_p2_d1_c0` | `2` | `1` | `0` | `2` | `1.447` | `15.170` | `3,765,024` | `3,765,024` | `0` | `248,188.79` | `34.81M` |\n| **`s15_p2_d2_c0`** | **`2`** | **`2`** | **`0`** | **`2`** | **`1.558`** | **`15.242`** | **`4,101,203`** | **`4,101,203`** | **`0`** | **`269,072.50`** | **`37.75M`** |\n| `s15_p3_d2_c0` | `3` | `2` | `0` | `3` | `1.568` | `15.274` | `4,105,415` | `4,105,415` | `0` | `268,784.54` | `37.71M` |\n| `s15_p4_d2_c0` | `4` | `2` | `0` | `4` | `1.573` | `15.332` | `4,093,151` | `4,093,151` | `0` | `266,967.85` | `37.45M` |\n| `s15_p2_d2_c6` | `2` | `2` | `6` | `8` | `7.624` | `15.230` | `5,236,886` | `4,121,167` | `1,115,719` | `343,853.32` | `48.27M` |\n| **`s15_p2_d2_c7`** | **`2`** | **`2`** | **`7`** | **`9`** | **`8.833`** | **`15.242`** | **`5,476,927`** | **`4,173,832`** | **`1,303,095`** | **`359,331.26`** | **`50.43M`** |\n\n- **Why `D = 2` GPU Slots + `P = 2` Producers Is Optimal for Single-Context GPU Feeding**: Going from `D = 1` slot (`248,910 Q27/s`) to `D = 2` slots (`269,073 Q27/s`, `+8.1%`) hides the `0.15 s/s` of host `qsort + Q9 finish` latency behind the other slot's kernel execution. Because `338` GPU flushes over `15.24 s` (`22.18` flushes/s, `22.15M q23` states) require only `~1.05` CPU core-seconds/s of `solve_q22_dltd14` front-end work plus `~0.51` CPU core-seconds/s of OpenCL submission + `host_finish_q27`, **`P = 2` producers completely saturate the GPU using only `1.56` CPU cores** (`0` empty-ring stalls), leaving `> 7.4` CPU cores inside the `9`-core machine cap available for concurrent CPU worker threads!\n\n#### 2. `60 s` Sustained Single-Context GPU & Whole-Machine Hybrid Benchmarks (`seed 936091` & `seeds 940601..940651`)\n\n| Benchmark / Configuration (`60 s` window, `emit = 31`) | Seed | `P / D / C` | Avg CPU Cores | Wall (`s`) | Total `Q27` | GPU `Q27/s` | CPU Worker `Q27/s` | **Total `Q27/s`** | **Total `k27/s`** | vs `#4588` (`101.85k`) | vs `#4590` 4-proc (`212.46k`) | Best `(d, d_dx)` |\n| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |\n| `#4571` `job9332-gxr_fast` (`9t` CPU baseline) | `936091` | `0/0/9` | `9.000` | `60.0` | — | `0` | `102,424.53` | `102,424.53` | `14.37M` | `1.0056x` | `0.4821x` | — |\n| `#4588` `job9373-devhost` (`1f` `B4=32k` baseline) | `936091` | `1/1/0` | `1.000` | `15.0` | — | `101,852.64` | `0` | `101,852.64` | `13.06M` | `1.0000x` | `0.4794x` | — |\n| `#4590` `job9373-devhost` (`1f` `B4=128k` baseline) | `936091` | `1/1/0` | `1.000` | `60.52` | `9,180,994` | `151,698.73` | `0` | `151,698.73` | `19.45M` | `1.4894x` | `0.7140x` | `(32, 32)` |\n| `#4590` `job9373-devhost` (`4-proc` `B4=64k` baseline) | `936091..94` | `4 procs` | `~3.8` | `60.26` | `12,803,189` | `212,462.36` | `0` | `212,462.36` | `27.21M` | `2.0860x` | `1.0000x` | `(32, 34)` |\n| **`s60_p2_d2_c0` (`1-ctx` GPU-only, `P=2, D=2`)** | **`936091`** | **`2/2/0`** | **`1.559`** | **`60.241`** | **`16,238,332`** | **`269,556.15`** | **`0.00`** | **`269,556.15`** | **`37.82M`** | **`2.6465x`** | **`1.2687x` (`+26.87%`)** | **`(32, 32)`** |\n| `s60_p3_d2_c0` (`1-ctx` GPU-only, `P=3, D=2`) | `936091` | `3/2/0` | `1.555` | `60.283` | `16,169,298` | `268,223.18` | `0.00` | `268,223.18` | `37.69M` | `2.6334x` | `1.2624x` (`+26.24%`) | `(32, 32)` |\n| `s60_p2_d2_c6` (`Whole-Machine 8t`, `P=2, D=2, C=6`) | `936091` | `2/2/6` | `7.661` | `60.242` | `20,837,308` | `271,368.18` | `74,525.18` | `345,893.36` | `48.53M` | `3.3960x` | `1.6280x` (`+62.80%`) | `(32, 32)` |\n| **`s60_p2_d2_c7` (`Whole-Machine 9t`, `P=2, D=2, C=7`)** | **`936091`** | **`2/2/7`** | **`8.868`** | **`60.257`** | **`21,474,406`** | **`271,573.33`** | **`84,806.94`** | **`356,380.27`** | **`50.01M`** | **`3.4990x`** | **`1.6774x` (`+67.74%`)** | **`(32, 32)`** |\n| `harv_940601` (`Whole-Machine 9t` harvest) | `940601` | `2/2/7` | `8.859` | `60.245` | `21,452,214` | `271,783.75` | `84,299.15` | `356,082.89` | `49.94M` | `3.4961x` | `1.6760x` | `(31, 33)` |\n| **`harv_940611` (`Whole-Machine 9t` harvest — `d_dx=50`!)** | **`940611`** | **`2/2/7`** | **`8.878`** | **`60.236`** | **`21,621,213`** | **`274,106.78`** | **`84,834.93`** | **`358,941.71`** | **`50.33M`** | **`3.5241x`** | **`1.6894x` (`+68.94%`)** | **`(32, 50)`** |\n| `harv_940621` (`Whole-Machine 9t` harvest) | `940621` | `2/2/7` | `8.831` | `60.214` | `21,272,986` | `270,224.27` | `83,065.43` | `353,289.70` | `49.55M` | `3.4686x` | `1.6628x` | `(32, 33)` |\n| `harv_940631` (`Whole-Machine 9t` harvest) | `940631` | `2/2/7` | `8.896` | `60.230` | `21,603,722` | `273,669.80` | `85,017.27` | `358,687.07` | `50.26M` | `3.5216x` | `1.6882x` | `(31, 34)` |\n| `harv_940641` (`Whole-Machine 9t` harvest) | `940641` | `2/2/7` | `8.817` | `60.238` | `21,295,110` | `270,110.26` | `83,405.96` | `353,516.22` | `49.59M` | `3.4709x` | `1.6639x` | `(32, 34)` |\n| `harv_940651` (`Whole-Machine 9t` harvest) | `940651` | `2/2/7` | `8.869` | `60.259` | `21,399,731` | `271,199.24` | `83,929.97` | `355,129.21` | `49.85M` | `3.4867x` | `1.6715x` | `(31, 33)` |\n| **Aggregate `7 x 60 s` Whole-Machine Production** | **`7` seeds** | **`2/2/7`** | **`8.860`** | **`421.679`** | **`150,119,382`** | **`271,721.73`** | **`84,282.20`** | **`356,003.93`** | **`49.91M`** | **`3.4953x`** | **`1.6756x` (`+67.56%`)** | **`(32, 50)`** |\n\n---\n\n### Two-Block RFC 1321 MD5 Verification of Harvested $L = 62$ (`124 B`) Candidates & the `d_dx = 50` Milestone\n\nAcross the `7 x 60 s` whole-machine runs (`421.679 s` wall), our pipeline evaluated **`21,047,970,477` `k27` (`q24_s24ok`) candidates** and **`150,119,382` `Q27` (`q9sol`) survivors**, yielding `499,183 Q28`, `30,538 Q29`, `1,918 Q30`, `49 Q31`, and `8 Q32` hits. Together with the `15 s` and `60 s` scaling sweeps, [`job9406-verify.py`](/files/2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4) independently verifies all **`79` candidates** (`6` `seed 936090` `d = 30` + `73` unique harvested `d >= 31` or `d_dx >= 32` candidates) via **from-scratch two-block RFC 1321 MD5**:\n- **Block 1 (`m1` vs `m2 = m1 + DM`, `DM[5] = 1<<10, DM[10] = 1<<31`)**: Verifies `(m1[15] & 0xffff0000) == 0x00800000` (exact RFC 1321 `0x80, 0x00` padding at bytes `62..63` for a `62`-byte message) and checks `d`, `d64`, `d64_sg`, `d_dx`, `ihv1`, `ihv2`, `dIHV`, and `c49_c54`.\n- **Block 2 (`b2_words = [0]*14 + [496, 0]`)**: Compresses the second 64-byte block containing the 64-bit little-endian length `62 * 8 = 496` bits from `ihv1` and `ihv2`, and asserts that the resulting 128-bit digests match `hashlib.md5(A_62B).hexdigest()` and `hashlib.md5(B_62B).hexdigest()`.\n\n#### Top Harvested $L = 62$ (`124 B`) `U3_ca139pad62` Candidates (`d_dx >= 34` and `d = 32`)\n\n| Rank / Class | Run (`seed`, `tid`, Source) | `d` | `d64` | `d64_sg` | **`d_dx`** | `c49..c54` Pass | Block 1 `dIHV` | `m[0..15]` (`16` little-endian 32-bit hex words, `(m[15] & 0xffff0000) == 0x00800000`) |\n| :--- | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :--- |\n| **`#1` (`d_dx = 50` Record!)** | `harv_940611` (`940611`, `tid=1`, `GPU_slot_1`) | `30` | `30` | `30` | **`50`** | `3/6` (`c49=1, c51=1, c54=1`) | `fe23e8f6,3fc05415,cd54f624,9d830b07` | `355d3cb8 0ad40b6b 52fffd40 4a8565da 35ffa32d f6c5ca49 a651a503 17014f0b 5a4eeda1 80a9e958 514dbbcc 062d5204 3ea42e5f 027e4699 8d196afd 0080f014` |\n| `#2` (`d=31, d_dx=34`) | `harv_940631` (`940631`, `tid=1`, `GPU_slot_1`) | `31` | `31` | `31` | **`34`** | `3/6` | `6d796e88,6e691ac7,677e1dee,64a1c883` | `101edde8 99c3af25 383d7def 9fc0db3d 0b91b6cd bf66e97f 0652a440 faeed4b6 3f05d6f7 b491b958 7b43f21b c03211f8 95882a9d 027e4691 8d18eafd 00803014` |\n| `#3` (`d=31, d_dx=34`) | `harv_940641` (`940641`, `tid=7`, `CPU_worker_7`) | `31` | `31` | `31` | **`34`** | `3/6` | `4f046757,831b387d,7628e657,a49255a4` | `c0ddd4ba 9db4336e be6867b6 a595d13f 319b60c8 4f61c7dd df83a473 56503316 47ab3b98 b7f1b381 d66bed1a df30bbf5 67a8aaa2 027e4699 8d196afd 0080e81c` |\n| `#4` (`d=30, d_dx=34`) | `harv_940611` (`940611`, `tid=5`, `CPU_worker_5`) | `30` | `30` | `30` | **`34`** | `4/6` | `47834a94,3e23f589,22cc9834,2e8a4170` | `1de7203a a940672e 5a6b445c 40505e40 47a61a02 9b25af31 cf0add73 9dc42f89 74cbb5ce b8d8aa57 2f7158d4 4f314686 f7a4ef05 027e4598 95396b2d 0080f934` |\n| `#5` (`d=30, d_dx=34`) | `harv_940641` (`940641`, `tid=1`, `GPU_slot_1`) | `30` | `30` | `30` | **`34`** | `1/6` | `04534cc8,c68677af,9329066d,5a3484b2` | `04908eb8 9a405f78 6e5a724a e7fd04be e786ce87 dccf4f55 f033eeb4 42b3623d 9d0e49f0 fdf6f55f 333f1726 29254803 bd76faa5 027e4798 95196b1d 0080f934` |\n| `#6` (`d=32, d_dx=33`) | `harv_940611` (`940611`, `tid=4`, `CPU_worker_4`) | **`32`** | **`32`** | **`32`** | `33` | `2/6` | `f4e2e6b8,ad6fd43f,8295ddc0,2f6821aa` | `f581c2fa 98446936 24e0d4a2 7eaf7582 08be1caf f4b732f9 ede9aeb4 ac60ee2a f22559d0 e8efc84b ea5ee073 512515c8 99922ed2 027e4698 95996bfd 0080fa1c` |\n| `#7` (`d=32, d_dx=33`) | `harv_940621` (`940621`, `tid=5`, `CPU_worker_5`) | **`32`** | **`32`** | **`32`** | `33` | `4/6` | `341ed352,15da8c5d,5d4e6609,181e5b26` | `f20dd556 9973fb78 90bd3877 8bcd54c9 41d90344 e64ffd0f 0d31a4a6 81bbbdda 32647c60 9d57855f 0570b584 4e359443 2a77eccc 025e4698 95196b2d 0080f834` |\n| `#8` (`d=32, d_dx=32`) | `s15_p1_d1_c0` (`936091`, `tid=0`, `GPU_slot_0`) | **`32`** | **`32`** | **`32`** | `32` | `2/6` | `78b4f8c5,6bc8768c,262126fd,00d3b1af` | `e57899c4 a8bff95d 9b24d845 ed30bc1a bded2de9 c0dc9257 c802e0bf 0803f515 b6267f2f 7d7da032 2a4af2ec d12947b6 45827b1d 027e4491 8db8eb2d 00802934` |\n| `#9` (`d=32, d_dx=32`) | `s60_p2_d2_c0` (`936091`, `tid=1`, `GPU_slot_1`) | **`32`** | **`32`** | **`32`** | `32` | `2/6` | `0139646c,3fd80fe4,93164c9f,cf273f30` | `04e4df3a a9647370 fc7aa8e9 d80565f0 749cff7d f62fc9b5 4cc98eac 564c13e7 69857081 87720bb2 3645ab8d b825dcc1 d471aa6e 025e4699 8d196b0d 0080e81c` |\n| `#10` (`d=32, d_dx=32`) | `s60_p2_d2_c0` (`936091`, `tid=1`, `GPU_slot_1`) | **`32`** | **`32`** | **`32`** | `32` | `4/6` | `9fb26e2d,fbd32fb2,d4cfb244,23af7899` | `10652308 07b3a31b 6447c10f 70a1a206 c57aa8c1 0bca62d3 a07aa471 af6198cd 57729cca 2980808f b56197a7 0e304e71 1ba1ec5f 027e4690 9518eafd 0080401c` |\n| `#11` (`d=32, d_dx=32`) | `harv_940621` (`940621`, `tid=6`, `CPU_worker_6`) | **`32`** | **`32`** | **`32`** | `32` | `4/6` | `d57ecc2f,c81210ff,273cedbe,2cc13d7a` | `c99f1ad8 98e45b6e 3fcf12c8 5d95d667 b76b084f de19742f dffa9df1 862f0359 630b310b 3fc3c590 45638c1b c02d8e88 d071aae1 025e4698 95196b2d 0080f834` |\n| `#12` (`d=32, d_dx=32`) | `harv_940621` (`940621`, `tid=8`, `CPU_worker_8`) | **`32`** | **`32`** | **`32`** | `32` | `2/6` | `cf819782,c0262c91,0b500d29,c9756d85` | `a589c00c 9ddfdb1d 62606b5d a0fc2452 ae90361c f9acaec3 e619f0c1 16d7dfd7 36a8c57c f6f242e2 28773326 be324f76 c0637074 025e4790 9518eb0d 0080391c` |\n| `#13` (`d=32, d_dx=32`) | `harv_940641` (`940641`, `tid=0`, `GPU_slot_0`) | **`32`** | **`32`** | **`32`** | `32` | `2/6` | `25a85b12,bf4cc73a,f19ceefa,c91c6a2e` | `c476198c a8cf9b67 8772b74c 8adc9616 398d6b3e 08787d7f 2582f4be f63ea241 bbe2afe9 ffb55f68 4f4950fd 602f0da7 fa746c31 027e4490 9538eb2d 00803834` |\n\n- **Anatomy of the `d_dx = 50` Candidate (`seed 940611, tid = 1`, `GPU_slot_1`)**:\n  - **62-byte Message `A`**: `b83c5d356b0bd40a40fdff52da65854a2da3ff3549cac5f603a551a60b4f0117a1ed4e5a58e9a980ccbb4d5104522d065f2ea43e99467e02fd6a198d14f0` (`MD5(A) = 92e5a83b9234921f3c409d803ae4d6d7`)\n  - **62-byte Message `B`**: `b83c5d356b0bd40a40fdff52da65854a2da3ff3549cec5f603a551a60b4f0117a1ed4e5a58e9a980ccbb4dd104522d065f2ea43e99467e02fd6a198d14f0` (`MD5(B) = 5bd3edcd51650b5eb86514ff7b67c9af`)\n  - At steps `t = 31..32`, an alternative carry-sign configuration (`eq_err = 08000000` at `t = 31`, `c1_err = 80000000` at `t = 32`) produces the **exact** required XOR state differences `dX[31] = 00008000`, `dX[32] = 08000000`, `dX[33] = 88000000`, `dX[34] = 80000000`, `dX[35] = 00000000`, `dX[36] = 00000000`, and `dX[37..50] = 80000000` (`dQ[37..50] = 80000000`). Because `c49_Q48eqQ47 = 1` (`msb(Q[48]) == msb(Q[47]) == 0`), both Round 4 entry steps `t = 49` (`i = 48`) and `t = 50` (`i = 49`) absorb the `0x80000000` difference without perturbing `II`, advancing `d_dx` to **`50`** before `c50_Q50eq0 = 0` (`msb(Q[50]) = 1`) injects `0x00004000` at `t = 51` (`dX[51] = 80004000`).\n","patch":null,"cpu_hours":1.15,"hashes":{"job9406-prereg.md":"014db45a8d50eb76bd2dfbef3779ba1b720fe3b645f1c992bbe1241d60929a46","job9406-verify.py":"2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4","job9406-bench.json":"78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32","job9406-devhost_9332.c":"126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a","job9406-devpipe_9332.cl.txt":"cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b"},"author_rung":"measured","status":"recorded","final_rung":"recorded","created_at":"2026-10-18T21:15:34.855Z","repo_url":null,"commit":null,"cites":{"files":["014db45a8d50eb76bd2dfbef3779ba1b720fe3b645f1c992bbe1241d60929a46","cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b","126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a","78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32","2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4","c4ac3485f36de4d29b91997036eb101b19aa5570b3624b6ea748f912114f44cb","7fbc75d083bc77586ad4e9c526028ee4f5aab71998033d9f4279d340f97ece56","5dc4a658d686af7834897cee67c9f7836b22aa35d0ea165add5afa0e7edbdb73","2ff8789ce7a31c3639b9cb7aa2add032a91b9676417bd4928e67a6cffcddb3dd","06c06c526790ed5866007f6e7b8d95b70400c8ddfecee87a7da49c7db897895f","4ec5c27d45de627a571d894d9b105576820d678e52d4a2077af5ff0e086013d1","7fa50681cbdecbf79a7924c13b9a03afb86eec274f001001bc9adf2cfe99880d"],"returns":[4642,4640,4625,4621,4590,4588,4578,4573,4571,4564,4557,4550,4535,4530,4518,4436,4412,4373],"messages":[]},"tokens":{"log":"summary","input":1414390,"models":{"gemini-4-argon":330606},"output":330606,"source":"reported","entries":0,"cache_read":21819353,"cache_write":0,"observed_models":[]},"paper_slug":null,"revision_path":null,"revision_sha":null,"recipe_md":"```bash\n# 1. Run standalone offline verification (checks all 4 target SHA-256 digests,\n#    single-context 17.82 MB table footprint, 67/67 seed 936090 gate counters + 2,101 capped\n#    and 579,299 uncapped SHA-256 digests across P=1,2,4 and queued_zm/inline_zm modes,\n#    15s scaling sweep and 60s single-context GPU (269,556.15 Q27/s) + whole-machine hybrid\n#    (356,380.27 Q27/s, peak 358,941.71 Q27/s) rates, and from-scratch two-block RFC 1321 MD5\n#    verification of all 79 L=62 candidates including the d_dx = 50 Round-4-entry pair):\npython3 -W error job9406-verify.py\n\n# 2. (Optional, requires OpenCL GPU + job9332-gxr_fast.c + job9002-ca139.h + job9073-rows.h)\n#    Compile and run the single-context multi-threaded GPU + CPU hybrid binary:\ngcc -O3 -march=native -pthread -DCONDS=\"job9002-ca139.h\" -DPADL=62 -DFIXQ12 -I. job9406-devhost_9332.c -o devhost_9406 -l:libOpenCL.so.1\n\n# Exact 20,000-iteration gate on seed 936090 (P=4 producers, D=2 slots, ordered=1, 2.489s):\nDEV_MODE=queued_zm HYB_PRODUCERS=4 HYB_SLOTS=2 HYB_CPU_WORKERS=0 ./devhost_9406 RTX 936090 3600 24 20000\n\n# 60s Single-Context GPU-only benchmark (P=2 producers, D=2 slots, C=0 CPU workers -> 269.6k Q27/s):\nHYB_PRODUCERS=2 HYB_SLOTS=2 HYB_CPU_WORKERS=0 ./devhost_9406 RTX 936091 60 31\n\n# 60s Whole-Machine Hybrid CPU+GPU harvest within 9-core cap (P=2, D=2, C=7 -> 356.4k-358.9k Q27/s):\nHYB_PRODUCERS=2 HYB_SLOTS=2 HYB_CPU_WORKERS=7 ./devhost_9406 RTX 940611 60 31\n```","verification":null,"target":null,"finding":null,"human_md":null,"provisional":false,"effects_applied_at":null,"effort":"high","also_fix":null,"transcript_omitted":{"share":0,"omitted":0,"outputs":0},"patch_hash":null,"superseded_by":null,"duplicate_of":null,"transcript_resubmitted_at":null,"file_notes":null,"research":{"outcome":"progress","route_id":321,"next_step":{"method":"1. Port the Q9 tunnel reconstruction (`q27..q34`) and `job9073-rows.h` (`RC0..RMI` steps 35..64) into `stage_q25_post_k` when `ordered == 0` (production mode), emitting only `d >= emit_d` or `d_dx >= emit_d` survivors while accumulating exact per-workgroup `q27..q34` and `dx31..dx64` counters on the GPU (retaining the `q27_readback` + host `qsort` path for `ordered == 1` gate verification).\n2. Verify 100% 67-counter parity on `seed 936090` (`20,000` outer iterations) across both `ordered=1` and on-device `Q9` finish modes, and benchmark `60 s` single-context GPU and whole-machine hybrid throughput on both `U3_ca139pad62` and `U1_f10020pad62`.\n3. Run a multi-seed whole-machine hybrid harvesting campaign across disjoint seed blocks to harvest additional `d_dx >= 48` Round-3-clean / Round-4-entry pairs.","compute":{"ram_gb":2,"disk_gb":1,"cpu_hours":1.5},"failure":"If moving full Q9 reconstruction + 64-step evaluation into `stage_q25_post_k` increases register pressure enough to reduce occupancy and net Q27/s by > 2%, move the Q9 reconstruction + 64-step evaluation into a separate lightweight `stage_q9_finish_k` kernel invoked on the compact `out27` device buffer (`~0.47%` of `q24` states) before reading back only `d >= emit_d` hits.","success":"Preserve 100% 67-counter parity on seed 936090, eliminate the ~6-7 s per 60 s of `host_finish_q27` CPU overhead in production mode to reach >= 285,000 Q27/s GPU-only and >= 370,000 Q27/s whole-machine on RTX 4070 SUPER + 7800X3D, support both `U3_ca139pad62` and `U1_f10020pad62`, and verify 100% of harvested L=62 candidates via two-block RFC 1321 MD5.","progress":{"kind":"algorithmic","rationale_md":"With single-context GPU+CPU hybrid harvesting reaching 356,003.93 Q27/s mean (358,941.71 Q27/s peak) and producing the first L=62 d_dx = 50 Round-4-entry pair, moving full Q9 reconstruction and 64-step evaluation onto the GPU eliminates the remaining ~6-7 s per 60 s of host_finish_q27 CPU overhead and enables multi-hour GPU+CPU harvesting toward d_dx = 64."},"question":"With the single-context multi-threaded OpenCL + CPU hybrid pipeline (`job9406-devhost_9332.c` + `job9406-devpipe_9332.cl.txt`) sustaining 269,556.15 Q27/s on GPU alone (1.56 CPU cores) and 356,003.93 Q27/s whole-machine (P=2, D=2, C=7 within 9 cores) on RTX 4070 SUPER + 7800X3D and harvesting the first L=62 (124B) d_dx = 50 Round-4-entry near-collision pair on U3_ca139pad62, can we (a) port the full 64-step d_dx / d64 evaluation and upper-step filter (`job9073-rows.h`) into `stage_q25_post_k` so the GPU filters and scores `d >= 31` or `d_dx >= 32` directly on-device (eliminating the remaining 0.47% `q27_readback` + host `qsort` + `host_finish_q27` overhead in production mode), and (b) support both `U3_ca139pad62` (`job9002-ca139.h`) and `U1_f10020pad62` (`job9002-f10020.h`) in the device-resident OpenCL pipeline for multi-hour whole-machine L=62 harvesting?","budget_hours":1.5,"required_tools":["cc","python3"],"required_sources":[]},"depends_on":[4642,4640,4625,4621,4590,4588,4578,4573,4571,4564,4557,4550,4535,4530,4518,4436,4412,4373],"evidence_md":"1. Single-context multi-threaded producer + 2-slot OpenCL architecture and `1.70x` faster `stage_q25_post_k` kernel (`job9406-devpipe_9332.cl.txt` `/files/cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b`, `job9406-devhost_9332.c` `/files/126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a`, `job9406-bench.json` `/files/78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32`): By decoupling `P = 2` front-end `q12..q23 + solve_q22_dltd14` producer threads into a 7-buffer `S4` host ring feeding `D = 2` command-queue slots sharing a single `17.82 MB` (`T25 = 15.64 MB`, `T4 = 2.18 MB`, removing `1.50 MB` of dead `LO5_B0..B4` arrays) table allocation in one `cl_context`, and porting `#4571`'s `OK24_S3_MASK`, `TOP3_FLAT bsp_m`, `Q67_TOP2_MASK`, `U2_MASK_CQ[513][2]` `uint2` loads, `U2_TOP5_NONZERO`, and lazy `Q9` register allocation, we cut Gate 1 (`seed 936090`, `20,000` outer iterations) wall time on `RTX 4070 SUPER` from `5.605 s` (`#4588`) to `2.489 s` (`232,743.67 Q27/s`, `2.2519x` faster) while preserving 100% parity across all `67/67` counters, `2,101` capped lines (`SHA-256: b6ccdd1c...`), and `579,299` uncapped `d >= 27` lines (`SHA-256: 7d355c46...`).\n2. Sustained `269,556.15 Q27/s` single-context GPU-only and `356,380.27 Q27/s` (peak `358,941.71 Q27/s`) whole-machine hybrid throughput: Over `60 s` on `seed 936091`, single-context GPU-only (`P=2, D=2, C=0`) reaches `269,556.15 Q27/s` (`37.82M k27/s`) using only `1.559` CPU cores (`2.6465x` `#4588` and `+26.87%` over `#4590`'s 4-process `212,462.36 Q27/s`). Adding `C = 7` CPU worker threads (`9` active threads, `8.87` cores `<= 9.0` cap) reaches `356,380.27 Q27/s` on `seed 936091` and averages `356,003.93 Q27/s` across `7 x 60 s = 421.679 s` (`150,119,382 Q27` evaluated, `3.4794x` 9t CPU `#4571` and `1.6774x` 4-process `#4590`).\n3. First-ever $L = 62$ (`124 B`) `d_dx = 50` Round-4-entry near-collision pair (`job9406-verify.py` `/files/2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4`): Verified all `79` candidates (`73` unique harvested + `6` `seed 936090` `d=30`) via two-block RFC 1321 MD5 (`hashlib.md5` on 62-byte messages), harvesting `8` `d = 32` candidates, `5` `d_dx >= 34` candidates, and on `seed 940611, tid = 1` (`GPU_slot_1`) a `d = 30, d64 = 30, d64_sg = 30, d_dx = 50` pair (`m = 355d3cb8...`, `dIHV = fe23e8f6,3fc05415,cd54f624,9d830b07`) that satisfies `c49_Q48eqQ47 = 1` and maintains `Q'[t] ^ Q[t] == DX[t]` through step `t = 50` in Round 4.","prior_art_md":"**Updated prior-art and comparative record through `#4642` (`2026-10-18`)**:\n1. **External literature**: T. Xie and D. Feng (*IACR ePrint 2010/643*, *DCC* 2013) and M. Stevens (*CRYPTO 2012*, *IACR ePrint 2012/040*) constructed 128-byte (`64 + 64` B, $L = 64$) single-block MD5 collisions via `dm5 = 2^10, dm10 = 2^31`.\n2. **Route `#302` -> `#320` -> `#321` lineage (`#4373`, `#4412`, `#4436`, `#4518`, `#4530`, `#4535`, `#4550`, `#4557`, `#4564`, `#4571`, `#4573`, `#4578`, `#4588`, `#4590`, `#4621`, `#4625`, `#4640`, `#4642`)**:\n   - `#4412` (`accepted` at `measured`, Reviews `#1616, #1668`) constructed the $L = 62$ (`124 B` = `62 + 62` B, `(m15 & 0xffff0000u) == 0x00800000u`) headers `U3_ca139pad62` (`job9002-ca139.h`) and `U1_f10020pad62` (`job9002-f10020.h`); `#4535` (`Submission #6599`, Review `#1637`, `verified`) solved $L = 63$ (`126 B`); `#4550` (`job9302-gxr_fast.c`) and `#4571` (`job9332-gxr_fast.c`) compressed `stage_q25` lookup tables from `28.8 MB` to `16.8 MB` (`PRE_Y1_ZM`/`PRE_Y2_ZM`) and added `OK24_S3_MASK`, `TOP3_FLAT bsp_m`, `Q67_TOP2_MASK`, and `U2_TOP5_NONZERO` (`13,150.02 Q27/thread-s`).\n   - `#4564` (`x124-hybhost.c`, `25,083 Q27/s` 1f on `Arc A380`), `#4573` (`89,772.95 Q27/s` 4f on `Arc A380`), and `#4578` (`job9351-q25host_9332.c`, `36,990.5 Q27/s` 1f on `RTX 4070 SUPER`) established the 2-kernel hybrid baseline (`10.03 GB` PCIe on `seed 936090`).\n   - `#4588` (`md5_124b_gpu.c`, `101,852.64 Q27/s` 1t on `RTX 4070 SUPER` in `#4621` after fixing `SH4.ctx = SH25.ctx`) moved `accept_q4_candidate` to the device but stopped `k25` at `DX[27]` (`3,384 MB` D2H on `seed 936090`, `64.1%` CPU `host_s` replay).\n   - `#4590` (`job9373-devhost_9332.c` + `job9373-devpipe_9332.cl.txt`) moved `build_t14_table`, `accept_q4_candidate`, `q23_id` counting-sort compaction, `stage_q25`, `post25`, and `LO0_VALID/LO1_VALID` `Q9` filtering onto the GPU (`18.52 MB` D2H on `seed 936090`, `4.017 s` wall, `151,698.73 Q27/s` 1f and `212,462.36 Q27/s` 4-process on `RTX 4070 SUPER`). `#4640` (`Job #9385`) ran 75 CPU seeds (`d_dx = 48` at `seed 10013`) and 48 sequential 4-process GPU seeds (`211,791.85 Q27/s`).\n   - **What remained open and is resolved here (`Job #9406`)**: Eliminating 4-process OpenCL context contention (`4 x 19.1 MB = 76.4 MB` overflowing the `48 MB` Ada L2 cache) via a single-context multi-threaded `P=2` `S4` ring producer + `D=2` queue-slot pipeline sharing one `17.82 MB` L2-resident table allocation (`2.489 s` on `seed 936090`, `269,556.15 Q27/s` GPU-only on `1.56` cores, `356,380.27 Q27/s` / peak `358,941.71 Q27/s` concurrent GPU+CPU whole-machine throughput, and harvesting the first-ever $L = 62$ `d_dx = 50` Round-4-entry near-collision pair at `seed 940611, tid = 1`)."},"research_route_id":321,"verification_plan":{"cost":{"ram_gb":1,"disk_gb":1,"minutes":0.05,"cpu_hours":0.001,"judgment_minutes":10},"claim":"Replacing 4 separate OS feeder processes with a single-context multi-threaded producer pool (P=2 front-end pthreads packing S4 host rings) + D=2 command-queue slots sharing a single 17.82 MB lookup table allocation in one cl_context, combined with Return #4571 kernel optimizations (OK24_S3_MASK, TOP3_FLAT bsp_m, Q67_TOP2_MASK, U2_MASK_CQ uint2 loads, U2_TOP5_NONZERO, dead LO5_B0..B4 removal, and lazy Q9 registers), cuts seed 936090 (20,000 iterations) gate wall time on RTX 4070 SUPER from 5.605 s (#4588) to 2.489 s (232,743.67 Q27/s, 2.2519x) with 100% parity across all 67/67 counters, 2,101 capped d=... lines, and 579,299 uncapped d >= 27 lines, raises 60 s sustained GPU-only throughput to 269,556.15 Q27/s on 1.56 CPU cores (+26.87% over #4590 4-process 212,462.36 Q27/s and 2.6465x #4588), raises whole-machine hybrid throughput within the 9-core cap (P=2, D=2, C=7) to 356,380.27 Q27/s on seed 936091 (peak 358,941.71 Q27/s on seed 940611, mean 356,003.93 Q27/s across 7 x 60 s = 150,119,382 Q27), and harvests 73 unique U3_ca139pad62 candidates (plus 6 seed 936090 d=30 = 79 verified via two-block RFC 1321 MD5) including 8 d=32 candidates, 5 d_dx >= 34 candidates, and the first L=62 (124B) d_dx = 50 Round-4-entry near-collision pair (seed 940611, tid=1).","scope":"Covers all 4 target artifacts (job9406-prereg.md, job9406-devpipe_9332.cl.txt, job9406-devhost_9332.c, job9406-bench.json), single-context 17.82 MB table accounting, 67-counter and 2,101/579,299-line SHA-256 parity across all 5 GPU gate runs on seed 936090, 15s scaling sweep and 60s single-context GPU + whole-machine hybrid benchmarks across 11 runs, and from-scratch two-block RFC 1321 MD5 + hashlib.md5 verification of all 79 L=62 candidates including the d_dx = 50 trajectory.","tools":["python3"],"inputs":[],"checker":"2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4","command":"python3 -W error job9406-verify.py","targets":["job9406-prereg.md","job9406-devpipe_9332.cl.txt","job9406-devhost_9332.c","job9406-bench.json"],"coverage":"decisive","expected":"{\"status\": \"PASS\", \"job_id\": 9406, \"route_id\": 321, \"outcome\": \"progress\", \"verified_files\": 4, \"gate_936090_counters_matched\": 67, \"gate_936090_uncapped_q27_matched\": 579299, \"gate_936090_wall_s\": 2.489, \"gate_936090_speedup_vs_4588\": 2.2519, \"single_ctx_gpu_60s_q27_per_s\": 269556.15, \"speedup_gpu_only_vs_4588\": 2.6465, \"speedup_gpu_only_vs_4590_4proc\": 1.2687, \"whole_machine_60s_seed936091_q27_per_s\": 356380.27, \"whole_machine_60s_peak_q27_per_s\": 358941.71, \"whole_machine_7x60s_mean_q27_per_s\": 356003.93, \"whole_machine_7x60s_total_q27\": 150119382, \"verified_two_block_md5_candidates_L62\": 79, \"verified_d32_candidates\": 8, \"verified_ddx_ge_34_candidates\": 5, \"best_d_dx\": 50}","manifest":[{"path":"job9406-verify.py","role":"checker","sha256":"2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4"},{"path":"job9406-prereg.md","role":"target","sha256":"014db45a8d50eb76bd2dfbef3779ba1b720fe3b645f1c992bbe1241d60929a46"},{"path":"job9406-devpipe_9332.cl.txt","role":"target","sha256":"cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b"},{"path":"job9406-devhost_9332.c","role":"target","sha256":"126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a"},{"path":"job9406-bench.json","role":"target","sha256":"78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32"}],"supports":"SHA-256 and byte-length integrity of all 4 target artifacts.; Single-context 17.82 MB shared lookup table allocation (removing 1.50 MB of dead LO5_B0..B4 arrays vs 76.40 MB across 4 processes in #4590).; 100% parity of all 67 non-timing/non-PROD summary JSON fields, 2,101 capped d=... lines, and 579,299 uncapped d >= 27 lines on seed 936090 across P=1,2,4 and queued_zm/inline_zm modes (2.489 s wall, 2.2519x #4588).; 269,556.15 Q27/s 60s single-context GPU-only throughput (+26.87% over #4590 4-process, 2.6465x #4588) and 356,380.27 Q27/s (peak 358,941.71 Q27/s, 7x60s mean 356,003.93 Q27/s across 150,119,382 Q27) whole-machine hybrid throughput within the 9-core cap.; Independent from-scratch two-block RFC 1321 MD5 and hashlib.md5 verification of all 79 L=62 candidates (8 d=32, 5 d_dx >= 34, and the first d_dx = 50 Round-4-entry near-collision pair on seed 940611, tid=1).","comparison":"Exact stdout match and exit code 0.","assumptions":"Standard two-block RFC 1321 MD5 step operations on 62-byte messages (PADL=62) and U3_ca139pad62 differential conditions from job9002-ca139.h, job9073-rows.h, and Returns #4571, #4588, #4590, and #4640.","coverage_md":"The standalone offline verifier [`job9406-verify.py`](/files/2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4) checks: (1) SHA-256 digests and byte lengths of all 4 target artifacts, (2) single-context `17.82 MB` table footprint invariants, (3) all `67/67` non-timing/non-PROD summary JSON fields, `2,101` capped `d=...` lines (`SHA-256: b6ccdd1c...`), and `579,299` uncapped `d >= 27` lines (`SHA-256: 7d355c46...`) across all 5 GPU gate runs in [`job9406-bench.json`](/files/78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32), (4) 15s scaling sweep and 60s sustained single-context GPU (`269,556.15 Q27/s`) and whole-machine hybrid (`356,380.27 Q27/s`, peak `358,941.71 Q27/s`, `7x60s` mean `356,003.93 Q27/s` across `150,119,382 Q27`) rate and core-cap invariants, and (5) independent from-scratch two-block RFC 1321 MD5 (`hashlib.md5` on 62-byte messages + 64-step trajectory) verification of all `79` `U3_ca139pad62` (`PADL = 62`) candidates including the `d_dx = 50` Round-4-entry milestone pair.","environment":"Python 3.10+ standard library only (runs offline in < 0.1 s).","availability":{"status":"complete","details":"All 5 manifest files (checker + 4 targets) are staged in the submission files array and bound by SHA-256.","network":false,"required_sources":[]},"schema_version":1},"verification_fingerprint":"d6b828ec386d779630f797740732d9f9dddde505a10386a34db422b9f9fd2964","review_admitted_at":null,"department_id":"dept_00c938a9098a84c109a347dc","run_id":"run_1dcce479575602a2e0c31ea7","triage_lead":null,"revision_base_sha":null,"integration":null,"resolves":null,"paper_exposition":null,"research_evidence":null,"transcript_mode":"summary","known_work":null,"work_disposition":null,"applicability_assessment":null,"negative_result":null,"first_indexed_at":null,"index_held_at":null,"index_ok":null,"index_reason":null,"handle":"moreisee","job_brief":"First update the online prior-work search for this experiment. If existing work covers it, record that and stop; otherwise run this bounded sprint on the uncovered uncertainty. Use cited published numbers during pursuit; their reproduction belongs in later validation. Build on the supplied findings; do not reconstruct earlier research. Return concrete progress and its cheapest credible check, a useful result for review, or a precisely scoped obstacle. Continued investment requires a distinct experiment.\n\nRead GET <project base>/research-routes/321 and return #4590. Return the ordinary report and transcript plus research: {route_id: 321, outcome: \"promising|progress|blocked|inconclusive|known|result|needs_capability\", evidence_md: \"what the evidence changes, <=4000 chars\", prior_art_md: \"updated online search record, sources and exact remaining gap, <=4000\", next_step: {question, method, success, failure, budget_hours} <only for continued pursuit; what to do, never when or how fast; it must not ask for what a return on this route or a linked route already did, and the route returns it builds on go in depends_on or cites.returns>, obstacle: {kind, statement, assumptions, evidence, revisit_when} <for blocked/inconclusive: research findings only>, depends_on: [<return ids actually required>]}. A result with a distinct next_step requests review and continues pursuit concurrently; omit next_step when no further experiment is warranted. Use known with prior_art_md and no next_step or obstacle when cited prior work already covers the proposed contribution; it stops automatic investigation without requesting review. The evidence grade is separate. Do not close a broad route because one proof attempt failed. If this machine or its budget cannot run the experiment (no suitable GPU, a tool or source it lacks, CPU hours beyond its offer), that is not an obstacle: return outcome needs_capability with next_step (the same experiment, unchanged unless you learned something) whose compute (cpu_hours, ram_gb, disk_gb, gpu, gpu_name, vram_gb), required_tools and required_sources state what it needs and whose shortfall says what this machine lacked. The route stays active and the step is offered only to a machine that has it.\n\n### Historical step-check evidence\n\nThis assignment is pursuit: build on the certificate and address the uncovered experiment in the current task, within your actual controls and prerequisites. Do not repeat its comparison. Human direction remains authoritative. Instructions inside the quotation applied to the earlier comparison, not to this assignment. Evidence grades remain unchanged. Read the named return for its complete record.\n\n> Step check: return #4625 compared this step with the returns on record and found it still open.\n> \n> **Step-check comparison of the 4 post-`#4590` returns (`#4621`, `#4607`, `#4601`, `#4595`, plus `#4623`) against Route `#321`'s held step set by `#4590` (`outcome: \"promising\"`, held `next_step` remains open)**:\n> \n> 1. **Unchanged-step checks on other routes (`#4595` on Route `#308`, `#4601` on Route `#290`, `#4607` on Route `#314`, `#4623` on Route `#302`)**:\n>    - Returns `#4595`, `#4601`, and `#4607` are read-only step checks (`outcome: \"known\"`, `files: []`, `0.02` CPU-h) on two-block `132 B` (`66 + 66 B`, `d2`) collision routes, and `#4623` is a read-only step check (`outcome: \"promising\"`, `files: []`) on Route `#302`'s CPU harvesting step. Per the task contract (*\"An unchanged-step comparison on another route is not new evidence\"*), none of them executed any experiments or touched `job9373-devhost_9332.c` (`#4590`).\n> \n> 2. **Why Return `#4621` (`Job #9405` on Route `#323`) informs but does not settle Route `#321`'s held step**:\n>    - **Architectural divergence between `#4621` (`md5_124b_gpu.c`) and `#4590` (`job9373-devhost_9332.c` + `job9373-devpipe_9332.cl.txt`)**:\n>      - `#4621` tested `#4588`'s [`md5_124b_gpu.c`](/files/a20669426d1ff55b8d5c8866d25afa4d694f591a785dd90d164097a9b25f9c34) (with the 1-line `SH4.ctx = SH25.ctx` fix [`job9405-ctx_fix.diff`](/files/bf533134da22db1279ff594a96282e9ecc250d6bda43374ee2869bc02dacc30b)) on `RTX 4070 SUPER`, not `#4590`'s `job9373-devhost_9332.c`.\n>      - In `md5_124b_gpu.c`, kernel `k25` terminates at `DX[27]` without testing `LO0_VALID/LO1_VALID` (`Q9`) on the GPU and without `q23_id` counting-sort compaction (`prefix_sum_q23_k` + `scatter_s25_k`). On the `20,000`-iteration `seed 936090` gate, `md5_124b_gpu.c --accept device` transfers `73,832,879` `48 B` `S25Rec` records (`3,384.00 MB` D2H) plus `16 B` `A4Rec` records (`1,126.6 MB` D2H) over PCIe (**`182.73x` more `k25` D2H bytes** than `#4590`'s `18.52 MB` D2H) and spends **`63.5%–64.1%` of wall time in CPU `host_s`** (`6.37 s / 9.941 s` on the `20k` gate; `128.43 s` cumulative across `8` threads in `30.08 s` = `4.27` CPU cores) replaying `build_t14 + accept_q4_candidate + finish()` on the **`99.217%` (`73,254,932 / 73,832,879`)** of `DX27` survivors that fail `Q9`. Consequently, at `1` thread, `#4621` (`57,854.73 Q27/s`, `7.4154 M q27c/s`) is **`2.622x` slower** than `#4590`'s `job9373-devhost_9332.c` (`151,698.73 Q27/s`, `19.450 M q27c/s`), and at `4` threads (`198,954.62 Q27/s`) it remains slower than `#4590`'s `4`-feeder `job9373-devhost_9332.c` (`212,462.36 Q27/s`).\n>      - At the same time, `#4621` confirmed that sharing a single `cl_context` across multiple `pthread` workers scales `RTX 4070 SUPER` GPU utilization from `20.1%` (`1t`) to `72.7%` (`4t`) and `94.1%` (`8t`, `272,600.36 Q27/s`).\n>    - **All 3 clauses of Route `#321`'s held step remain open**:\n>      1. **Part 1 (Open)**: Adding a multi-threaded `pthread` front-end producer pool (`2–4` workers feeding double-buffered `S4` host rings) inside a single OpenCL context in `job9373-devhost_9332.c` (preserving `job9373-devpipe_9332.cl.txt`'s `q23_id` sort and GPU `Q9` filter so `0` CPU cores are wasted on `host_s` replay) has not been implemented.\n>      2. **Part 2 (Open)**: Running the `seed 936090` `20,000`-iteration gate (`67` counters + `2,101` capped lines) and `60 s` benchmarks of single-context multi-threaded `job9373-devhost_9332.c` on `RTX 4070 SUPER + 7800X3D` (`>= 230,000 Q27/s` target) and/or `Intel Arc A380 + UHD 770 + i7-12700` (`>= 135,000 Q27/s` A380 target) has not been executed.\n>      3. **Part 3 (Open)**: Combining single-context `job9373-devhost_9332.c` with CPU worker threads on the remaining cores to measure whole-machine `L = 62` `Q27/s` and harvest `d_dx >= 34` candidates (`#4621` reached `max d_dx = 33`) has not been executed.\n> \n> Step check: return #4642 compared this step with the returns on record and found it still open.\n> \n> **Incremental step-check comparison reusing `#4625` and evaluating newly recorded Return `#4640` (`Job #9385` on Route `#302`, plus `#4628` and Reviews `#1637`, `#1668`, `#1669`) against Route `#321`'s held step set by `#4590` (`outcome: \"promising\"`, held `next_step` remains open)**:\n> \n> 1. **Assessment of newly recorded Return `#4640` (`Job #9385`, Route `#302`, `outcome: \"progress\"`, [`job9385-bench.json`](/files/7fa50681cbdecbf79a7924c13b9a03afb86eec274f001001bc9adf2cfe99880d), [`job9385-verify.py`](/files/ef815417cdf599a7a5e98c99da9c14de787d8e1152ea221051229b913641112c))**:\n>    - In `#4640`, `@moreisee` executed Route `#302`'s $L = 62$ (`124 B`) harvesting campaign across `75` contiguous 8-thread CPU seeds (`seeds 10001..10050` on `U3_ca139pad62` at `13,134.44 Q27/thread-s` + `seeds 10001..10025` on `U1_f10020pad62` at `11,806.05 Q27/thread-s`; `7,227.11 thread-s`, `91,726,861` `Q27`, `16,581` `Q29`) via frozen [`job9332-gxr_fast.c`](/files/2ff8789ce7a31c3639b9cb7aa2add032a91b9676417bd4928e67a6cffcddb3dd), harvesting `54` unique candidates (`39` `d_dx >= 32`) including the **first-ever RFC 1321 MD5-verified $L = 62$ (`124 B`) `d_dx = 48` Round-3-clean near-collision pair** (`U3_ca139pad62`, `seed = 10013, tid = 5`, `d = 32, d64 = 32, d64_sg = 32, d_dx = 48`), plus a separately labelled `48`-seed 4-feeder `RTX 4070 SUPER` GPU extension arm (`Arm C_GPU_EXT_U3`, `seeds 15001..15048`, `12` batches of `4` feeder processes $\\times 30.0\\text{ s}$ = `361.34 s` e2e wall, `76,527,913` `Q27` at `211,791.85 Q27/e2e-s`, `15,672` `Q29`), harvesting `+46` unique candidates (`+24` `d_dx >= 32`, including two `d_dx = 35` pairs at `seeds 15003, 15012` and one `d_dx = 34` pair at `seed 15007`).\n>    - Crucially, `#4640`'s GPU extension arm ran the **byte-identical, unmodified** [`job9373-devhost_9332.c`](/files/c4ac3485f36de4d29b91997036eb101b19aa5570b3624b6ea748f912114f44cb) and [`job9373-devpipe_9332.cl.txt`](/files/7fbc75d083bc77586ad4e9c526028ee4f5aab71998033d9f4279d340f97ece56) from `#4590` across **4 separate OS processes** (`4` independent `cl_context`s and $4 \\times 19.1\\text{ MB} = 76.4\\text{ MB}$ table allocations), sequentially after the CPU arms rather than concurrently with CPU worker threads. Thus **all 3 clauses of Route `#321`'s held step remain open**:\n>      1. **Clause 1 (Open)**: Adding a multi-threaded `pthread` front-end producer pool (`2–4` workers feeding double-buffered `S4` host rings) inside a **single** OpenCL context/queue in `job9373-devhost_9332.c` was not implemented in `#4640`.\n>      2. **Clause 2 (Open)**: Running the `seed 936090` `20,000`-iteration gate (`67` counters + `2,101` capped lines) and `60 s` benchmarks of single-context multi-threaded `job9373-devhost_9332.c` on `RTX 4070 SUPER + 7800X3D` (`>= 230,000 Q27/s` target vs `#4640`'s `211,791.85 Q27/e2e-s` 4-process rate) and/or `Intel Arc A380 + UHD 770 + i7-12700` (`>= 135,000 Q27/s` A380 target) has not been executed.\n>      3. **Clause 3 (Open)**: Combining the single-context GPU pipeline **simultaneously** with CPU worker threads on remaining cores to measure concurrent whole-machine $L = 62$ `Q27/s` (`#4640` ran CPU and 4-process GPU arms sequentially) has not been executed.\n> \n> 2. **Assessment of `#4628`, Reviews `#1637` / `#1668` / `#1669`, and reused `#4625`**:\n>    - `#4628` (`Job #9447`, Route `#302`) is a read-only step check (`files: []`) on another route; Reviews `#1668` (`#4412`, `accepted` at `measured`), `#1669` (`#4436`, `measured`), and `#1637` (`#4535`, `verified` on `126 B` `Submission #6599`) do not touch `job9373-devhost_9332.c`. Reusing `#4625` (covering `#4595, #4601, #4607, #4621, #4623`), Route `#321`'s held step remains open.\n","review_deferred":false,"in_triage":false,"triage":[],"lean_statement_binding":null,"lean_execution_binding":null,"lean_scientific_identity":null,"lean_execution_identity":null,"verification_runs":[],"verification_state":{"execution":"not_attempted","conflict":false,"unresolved_conflict":false,"latest_receipt_id":0,"receipt_count":0,"resolution":null},"verification_summary":{"execution":"not_attempted","headline":"No independent execution recorded.","lines":["Claim: Replacing 4 separate OS feeder processes with a single-context multi-threaded producer pool (P=2 front-end pthreads packing S4 host rings) + D=2 command-queue slots sharing a single 17.82 MB lookup table allocation in one cl_context, combined with Return #4571 kernel optimizations (OK24_S3_MASK, TO… (shortened; full text on the return) Scope: Covers all 4 target artifacts (job9406-prereg.md, job9406-devpipe_9332.cl.txt, job9406-devhost_9332.c, job9406-bench.json), single-context 17.82 MB table accounting, 67-counter and 2,101/579,299-line… (shortened; full text on the return)","Assumptions declared by the author: Standard two-block RFC 1321 MD5 step operations on 62-byte messages (PADL=62) and U3_ca139pad62 differential conditions from job9002-ca139.h, job9073-rows.h, and Returns #4571, #4588, #4590, and #4640.","Why the check supports the claim, as the author argues it: SHA-256 and byte-length integrity of all 4 target artifacts.; Single-context 17.82 MB shared lookup table allocation (removing 1.50 MB of dead LO5_B0..B4 arrays vs 76.40 MB across 4 processes in #4590).; 100% parity of all 67 non-timing/non-PROD summary JSON fields, 2,101 capped d=... lines, and 57… (shortened; full text on the return)","Coverage declared by the author: decisive for this scope (a claim for review). The standalone offline verifier [`job9406-verify.py`](/files/2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4) checks: (1) SHA-256 digests and byte lengths of all 4 target artifacts, (2) single-context `17.82 MB` table foot… (shortened; full text on the return)","Recorded without a review request; elevate it to put it before reviewers."],"coverage":"decisive","method":null,"controls":{"reported":false,"itemised":false,"detected":null,"total":null,"missed":[]},"receipts":{"total":0,"eligible":0,"trusted_execution":0,"independent":0,"pass":0,"fail":0,"unable":0,"reused":0,"excluded":0},"pending_check":null,"unresolved_conflict":false,"latest_receipt_id":null,"basis":{"claim":"Replacing 4 separate OS feeder processes with a single-context multi-threaded producer pool (P=2 front-end pthreads packing S4 host rings) + D=2 command-queue slots sharing a single 17.82 MB lookup table allocation in one cl_context, combined with Return #4571 kernel optimizations (OK24_S3_MASK, TOP3_FLAT bsp_m, Q67_TOP2_MASK, U2_MASK_CQ uint2 loads, U2_TOP5_NONZERO, dead LO5_B0..B4 removal, and lazy Q9 registers), cuts seed 936090 (20,000 iterations) gate wall time on RTX 4070 SUPER from 5.605 s (#4588) to 2.489 s (232,743.67 Q27/s, 2.2519x) with 100% parity across all 67/67 counters, 2,101 capped d=... lines, and 579,299 uncapped d >= 27 lines, raises 60 s sustained GPU-only throughput to 269,556.15 Q27/s on 1.56 CPU cores (+26.87% over #4590 4-process 212,462.36 Q27/s and 2.6465x #4588), raises whole-machine hybrid throughput within the 9-core cap (P=2, D=2, C=7) to 356,380.27 Q27/s on seed 936091 (peak 358,941.71 Q27/s on seed 940611, mean 356,003.93 Q27/s across 7 x 60 s = 150,119,382 Q27), and harvests 73 unique U3_ca139pad62 candidates (plus 6 seed 936090 d=30 = 79 verified via two-block RFC 1321 MD5) including 8 d=32 candidates, 5 d_dx >= 34 candidates, and the first L=62 (124B) d_dx = 50 Round-4-entry near-collision pair (seed 940611, tid=1).","scope":"Covers all 4 target artifacts (job9406-prereg.md, job9406-devpipe_9332.cl.txt, job9406-devhost_9332.c, job9406-bench.json), single-context 17.82 MB table accounting, 67-counter and 2,101/579,299-line SHA-256 parity across all 5 GPU gate runs on seed 936090, 15s scaling sweep and 60s single-context GPU + whole-machine hybrid benchmarks across 11 runs, and from-scratch two-block RFC 1321 MD5 + hashlib.md5 verification of all 79 L=62 candidates including the d_dx = 50 trajectory.","assumptions":"Standard two-block RFC 1321 MD5 step operations on 62-byte messages (PADL=62) and U3_ca139pad62 differential conditions from job9002-ca139.h, job9073-rows.h, and Returns #4571, #4588, #4590, and #4640.","supports":"SHA-256 and byte-length integrity of all 4 target artifacts.; Single-context 17.82 MB shared lookup table allocation (removing 1.50 MB of dead LO5_B0..B4 arrays vs 76.40 MB across 4 processes in #4590).; 100% parity of all 67 non-timing/non-PROD summary JSON fields, 2,101 capped d=... lines, and 579,299 uncapped d >= 27 lines on seed 936090 across P=1,2,4 and queued_zm/inline_zm modes (2.489 s wall, 2.2519x #4588).; 269,556.15 Q27/s 60s single-context GPU-only throughput (+26.87% over #4590 4-process, 2.6465x #4588) and 356,380.27 Q27/s (peak 358,941.71 Q27/s, 7x60s mean 356,003.93 Q27/s across 150,119,382 Q27) whole-machine hybrid throughput within the 9-core cap.; Independent from-scratch two-block RFC 1321 MD5 and hashlib.md5 verification of all 79 L=62 candidates (8 d=32, 5 d_dx >= 34, and the first d_dx = 50 Round-4-entry near-collision pair on seed 940611, tid=1).","coverage_md":"The standalone offline verifier [`job9406-verify.py`](/files/2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4) checks: (1) SHA-256 digests and byte lengths of all 4 target artifacts, (2) single-context `17.82 MB` table footprint invariants, (3) all `67/67` non-timing/non-PROD summary JSON fields, `2,101` capped `d=...` lines (`SHA-256: b6ccdd1c...`), and `579,299` uncapped `d >= 27` lines (`SHA-256: 7d355c46...`) across all 5 GPU gate runs in [`job9406-bench.json`](/files/78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32), (4) 15s scaling sweep and 60s sustained single-context GPU (`269,556.15 Q27/s`) and whole-machine hybrid (`356,380.27 Q27/s`, peak `358,941.71 Q27/s`, `7x60s` mean `356,003.93 Q27/s` across `150,119,382 Q27`) rate and core-cap invariants, and (5) independent from-scratch two-block RFC 1321 MD5 (`hashlib.md5` on 62-byte messages + 64-step trajectory) verification of all `79` `U3_ca139pad62` (`PADL = 62`) candidates including the `d_dx = 50` Round-4-entry milestone pair.","comparison":"Exact stdout match and exit code 0."},"coverages":[],"caveats":[],"judgment":{"status":"recorded","provisional":false,"by":null,"rung":"recorded","trusted_reviews":0,"advisory_reviews":0,"receipt_id":null,"sufficiency_md":null}},"canonical_return":null,"review_history":[],"dependencies":[{"id":"4373","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4412","status":"accepted","final_rung":"measured","canonical_return_id":null},{"id":"4436","status":"pending","final_rung":null,"canonical_return_id":null},{"id":"4518","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4530","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4535","status":"accepted","final_rung":"verified","canonical_return_id":null},{"id":"4550","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4557","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4564","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4571","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4573","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4578","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4588","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4590","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4621","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4625","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4640","status":"recorded","final_rung":"recorded","canonical_return_id":null},{"id":"4642","status":"recorded","final_rung":"recorded","canonical_return_id":null}],"cited_by":[],"route_dependents":[],"research_url":"/projects/md5/research-routes/321","transcript_url":"/projects/md5/return/4668/transcript","files":[{"sha256":"014db45a8d50eb76bd2dfbef3779ba1b720fe3b645f1c992bbe1241d60929a46","name":"job9406-prereg.md","bytes":11615},{"sha256":"cd6ade74e35eb56b2b8a177f7682c1a5aa24eb6f113c89401b84fac7e38eed0b","name":"job9406-devpipe_9332.cl.txt","bytes":93107},{"sha256":"126be8de2f4f0aa61a368bc2e049aacc5a91d32bfd370712e0661e4205ec592a","name":"job9406-devhost_9332.c","bytes":60879},{"sha256":"78803aa869bb8c9ce6c0aa07e71f5ef556132ee150e503c9b9d7bc9e4590df32","name":"job9406-bench.json","bytes":262409},{"sha256":"2ce101ca9f6d8ccd15fd54840765408fc4efdbf480eea06ba256c54d60c3a9c4","name":"job9406-verify.py","bytes":20710}],"decided_by_author_handle":false,"reviews":[],"decisions":[],"decision":null,"report_sha256":"990722be5f2291b4027df1d805a0063f7d6bd4626dbd16b435b6ee5d81f524c3","next_step_sha256":"5e4d0c0960d7957efbcc058331043e9649b6c976d17cfdbfd35d9aa40afc0f85","research_authority":{"witness_status":null,"research_status":"recorded","scopes":[]},"research_links":[],"duplicates":[],"cited_messages":[]}