[aien-spark]

Purpose and scope

Ports the GPU-path edits from omega-mv #995 (train r4, merge f9a9a40b5989) and #996 (train r5, merge 19b8340dbae7) into the paths tools/extract/drift.sh compares. The drift script and .github/workflows/drift.yml are unchanged.

The sm_121 opcodes are written in om/encoder/bw_enc_sm121.om. om/units.sh regenerates every portable binding from the pinned omc (omega-mv e5b32567, sha256 2c82480596bbdc50cda9626866209ef662d5834f278f839ff4ce0250c495029f) and omega-bridge 3dc4a00e (om/omega-bridge.lock).

Training R5 deleted omega_tc_patch_code. The Ampere corpus records an SM86 refusal for opcodes past MUFU.LG2 the same way it records an FP8 marker. APPEND prefill, f32 GDN, and the f32 TC2 epilogue still encode.

ATOMG.E.ADD.F32.FTZ.RN.STRONG.SYS, ATOMG.E.ADD.F32.FTZ.RN.STRONG.GPU, and destination presence slot 15 (put o 15 pres d) stay. They are required for drift parity with omega-mv 19b8340d, or for the host golden-word check that parity depends on. Details are in the receipt. They are NOT QUALIFIED. GB10 chip runs are NOT RUN. BLOCKED_OPERATOR. No authorized GPU window.

Evidence on the exact commit

• Pushed head: f7ea2c21e00421a8bb73f2a7dbf98f5ad9c3202d. That commit only adds evidence/host-gates-87d795b-20261102/.
• Head that was tested: 87d795b4446a762a0b12e05721f41b8abbde0c36.
• Machine: Linux 6.12.94+ x86_64, gcc 13.3.0. No GPU opened.
• Fresh-clone receipt: evidence/host-gates-87d795b-20261102/RECEIPT.md.
• CI cancellation note (not edited into an older receipt): evidence/host-gates-84553c1-20261102/RECEIPT.md.
• evidence/ampere-r5-sm86-refusal-20261102/RECEIPT.md matches its 1c3d33f bytes. evidence/drift-r4-r5-20261102/RECEIPT.md was not edited.

Fresh clone of 87d795b:

• make test-native-artifact: exit 0. test_gpu_artifact_export: 47 checks, 0 failed.
• make zero-cuda-gate: exit 0. ZERO_CUDA SELF-TEST PASS. ZERO_CUDA PASS for build/hello_gpu and build/gpu_matmul_api_test.
• build/hello_gpu --sim: exit 0. PASS. Host simulator, not a GB10 result.
• make doctor: exit 2. Printed NOT READY: 5 problems above. Expected on this machine. No device opened.
• GB10 hardware tests: NOT RUN. BLOCKED_OPERATOR.

make test without OMEGA_OMC reports 80 targets. The native-x86-host job exports OMEGA_OMC and reports 83. Both counts match the manifest for that command. The new receipt explains the difference. Nothing was corrected by changing a gate.

GB10 commands, when a window is authorized

make doctor first. Do not open /dev/nvidia* unless the maintainer has given a GPU window.

make -j8
cc -std=gnu11 -Wall -Wextra -Werror -D_GNU_SOURCE -O2 -ffp-contract=off \
  -Isrc -Im16 -Invrm \
  -Ithird_party/nvidia-open-580.173.02/src/common/sdk/nvidia/inc \
  -Ithird_party/nvidia-open-580.173.02/kernel-open/common/inc \
  -Ithird_party/nvidia-open-580.173.02/kernel-open/nvidia-uvm \
  -Ithird_party/nvidia-open-580.173.02/src/nvidia/arch/nvalloc/unix/include \
  -Itests \
  -o build/gpu_atomf32_probe_test \
  tests/gpu_atomf32_probe_test.c build/libomega_gpu.a -lpthread -lm
./build/gpu_atomf32_probe_test --host-only
./build/gpu_atomf32_probe_test

The last command is the chip run. Case 5 is ATOMG.E.ADD.F32.STRONG.GPU with no destination (RZ). That is the presence-slot check: dst_vreg < 0 must encode RZ and the add must still land. --cases 5 runs that case alone. The verdict line is GPU_ATOMF32_PROBE: PASS or FAIL. Save the output in a new evidence folder. Do not edit an existing receipt.

Correctness, compatibility and provenance

• Drift self-test against omega 19b8340d passed on the code this head adds evidence to. The drift script was not changed.
• Generated bindings come from om/units.sh. No hand edit of src/cudapp_om_*.c.
• New files carry SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception. No vendor binary or NVIDIA-tool listing.
• Older receipts were not rewritten. The 1c3d33f Ampere receipt matches that commit byte for byte.

Sweep re-check (2026-11-04, head f7ea2c2)

• EXERCISED: make -j8 test with OMEGA_OMC = omega-mv e5b32567 seed omc and OMEGA_BRIDGE = omega-bridge 3dc4a00e (the om/omega-bridge.lock pin, built with tools/bootstrap.sh && tools/build.sh), nice -n 19 taskset -c 0-9,15-19, linux aarch64: CUDA++ HOST TESTS PASS (83 targets), rc 0.
• bash tools/extract/drift.sh <omega-mv> d3d1b1bc7 (omega-mv main): this head DRIFT-GATE: PASS (920 GPU paths …). Negative control, CUDApp main 183fa4f: DRIFT-GATE: FAIL (45 of 920 GPU paths changed …).
• CI on f7ea2c2: all 20 checks pass (https://github.com/aien-dev/CUDApp/actions/runs/41479256224 drift, https://github.com/aien-dev/CUDApp/actions/runs/41479256466 host-tests).
• GB10: NOT_RUN (no GPU window in this sweep).

Open PRs #251 #262 #263 #264 #265 (and other branches off main) fail drift only because main fails it; they clear when this PR merges.

Known limitations

• BLOCKED_OPERATOR: GB10 runs of the f32 atomic-add probe, including the RZ presence-slot case, the coalesced KT1 kernel, and the training GEMM chip gate.
• EXP and LOG whole-program kernels stay in omega-only src/omega_numeric_divsqrt_gb10.c.
• This pull request has host evidence only. No GB10 chip run is claimed.

  

[cursor]

Risk: medium. Approved. Cursor Security Agent completed with no findings that need human review; Cursor Bugbot was not running, so that signal was skipped. This is a bounded Omega train r4/r5 GPU-path sync with host and drift evidence, so it stays within the medium approval threshold.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[aien-spark]

verdict: fix first

port itself checks out. i re-ran tools/extract/drift.sh --self-test from a box clone at 89ba4e8 against omega-mv main 19b8340d: DRIFT-GATE: PASS (920 GPU paths, omega 19b8340dbae7 ...), SELF-TEST: PASS. 20/20 checks green at head, native-x86-host finished in ~43 min so the 60 min cap is justified. no en/em dashes or banned phrasing in title, body, commits or added lines. new source files carry the apache spdx. sm121 ops 60..63 have golden words (omega_blackwell_verify_codegen_fixtures_bfly_f2fp, _atomf32). the process gaps below still block a merge.

findings:

1. evidence/ampere-r5-sm86-refusal-20261102/RECEIPT.md:45-48: 89ba4e8 edits a receipt committed in 1c3d33f (adds "CI on 1c3d33f"). AGENTS.md says never edit an existing receipt, and the pr body says "existing receipts were not edited". put that note in a new evidence folder and leave the 1c3d33f receipt as it was.
2. receipts are on 1c3d33f and d7c2e25, not on head 89ba4e8. REVIEWING.md 1.2/1.3 needs the gates re-run from a fresh clone at the exact head, with the receipt path posted. the code is byte-identical to 1c3d33f (head only adds the workflow timeout and the receipt text), but the rule is the exact head. make test-native-artifact, zero-cuda-gate, hello_gpu --sim and doctor are NOT RUN locally (ci host-gates is not a receipt).
3. om/encoder/bw_enc_sm121.om:486-493, src/omega_blackwell_codegen.c:561-570: ATOMG.E.ADD.F32.STRONG.{SYS,GPU} are new encodings with no chip run. tests/gpu_atomf32_probe_test.c is NOT RUN (BLOCKED_OPERATOR). AGENTS.md: every encoding needs golden words and a chip run. either get a gb10 receipt or mark these not qualified in docs/status. BFLY/F2FP keep their earlier patched bytes, so i treat those as lower risk [INFERENCE].
4. om/encoder/bw_enc_sm121.om:77: put o 15 pres d changes the presence slot for an existing op family. units regen and the sm86 oracle are green, but no sm121 chip run covers it.

fyi, not blocking:

• docs/review/CODEOWNERS.proposed is still inactive and uses placeholder teams. this pr touches core paths (omega_blackwell_codegen*, omega_bw_*), so section 2 is by convention only. the #264 concern still applies and this pr doesn't make it worse.
• the only approval is cursor[bot] at e140c1b. it's stale (not at head) and it's not a human. a non-author approval at head is still needed. merge with gh pr merge --match-head-commit 89ba4e8436b0ab7ad0020e23df14935f34845c0c.
• receipts report 80 targets, ci x86 logs 83 targets. explain the count difference in the new receipt.
• the pr is not a draft (REVIEWING 1.1).
• squash: drop Co-authored-by: aien-spark <aien@aienos.com>. it's not on the co-author list in docs/AUDIT.md section 2 or docs/LICENSE_PROVENANCE.md. no listed model row matches "Cursor Agent", so add no model trailer unless someone verifies which model did the work.
• no perf claims, so benchmarks-mv doesn't apply. none of the touched workflows use dtolnay/rust-toolchain, and the diff adds no aien-spark/ refs.

[cursor]

Fresh-clone host gates for the code commit 84553c1a429fec094d719641b2fb5315b4190ec4.

Receipt: evidence/host-gates-84553c1-20261102/RECEIPT.md.

That receipt tested 84553c1. The branch head 87d795b4446a762a0b12e05721f41b8abbde0c36 only adds the evidence files in that folder. It does not change code, docs, or workflows.

From the fresh clone of 84553c1, on Linux 6.12.94+ x86_64, gcc 13.3.0, no GPU opened:

• make test-native-artifact exit 0
• make zero-cuda-gate exit 0
• build/hello_gpu --sim exit 0 (PASS, host simulator)
• make doctor exit 2 (NOT READY: 5 problems above, expected here)
• bash tools/extract/drift.sh --self-test against omega 19b8340dbae70888c49fcc98299be928831bfd26 exit 0 (DRIFT-GATE: PASS, SELF-TEST: PASS)

GB10 chip runs remain NOT RUN. ATOMG.E.ADD.F32.FTZ.RN.STRONG.SYS, ATOMG.E.ADD.F32.FTZ.RN.STRONG.GPU, and destination presence slot 15 are NOT QUALIFIED (pending a GB10 receipt), listed in docs/status.md.

evidence/ampere-r5-sm86-refusal-20261102/RECEIPT.md is back to its 1c3d33f bytes. The CI cancellation note that had been appended there is in the new receipt instead.

[cursor]

Fresh-clone host gates for 87d795b4446a762a0b12e05721f41b8abbde0c36.

Receipt: evidence/host-gates-87d795b-20261102/RECEIPT.md.

That receipt tested 87d795b. The branch head f7ea2c21e00421a8bb73f2a7dbf98f5ad9c3202d only adds that evidence folder.

From the fresh clone, Linux 6.12.94+ x86_64, gcc 13.3.0, no GPU opened:

• make test-native-artifact exit 0
• make zero-cuda-gate exit 0
• build/hello_gpu --sim exit 0 (PASS)
• make doctor exit 2 (NOT READY: 5 problems above, expected here)

evidence/ampere-r5-sm86-refusal-20261102/RECEIPT.md still matches 1c3d33f byte for byte. The CI note is in evidence/host-gates-84553c1-20261102/RECEIPT.md.

80 and 83 host targets are both the HOST_TESTS count. 80 is make test on x86-64 with OMEGA_OMC unset. 83 is the same suite after ampere-x86.yml exports OMEGA_OMC (the three omc targets stay, test-omega-target-only is absent, test-q38-head-bounds is added).

GB10 chip runs remain NOT RUN. BLOCKED_OPERATOR. The f32 atomic encodings and the presence slot stay, because dropping the opcode declarations in src/omega_blackwell_codegen.h would fail drift against omega 19b8340d, and dropping put o 15 pres d would fail the hosted RZ golden-word check. The Spark commands are in the pull request body.