[drakestapleton]

Evidence on head d2c83a1

• tools/extract/drift.sh --self-test <omega-mv at 4376c2112d7469b6908c6403720c9001de4622b5> 4376c2112d7469b6908c6403720c9001de4622b5, rerun on this head (aarch64 host):
  • control 1: red as required (Makefile)
  • control 2: red as required (src/omega_gpu_attention_api.c edited in a scratch Omega commit)
  • DRIFT-GATE: PASS (783 GPU paths, omega 4376c2112d74, none moved away from CUDA++ since the baseline)
  • SELF-TEST: PASS (both controls red, real check green)
• make test-gpu-decode-graph test-gpu-prefill-host check-own30b-decode-profile test-gpu-fp8-slots test-gpu-fp8gemm passed at 02538d3 (host and simulator only). d2c83a1 adds one baseline row and crumbs, no code.
• crumb verify .: OK: 33 crumbs current on this head.
• GB10 hardware tests. NOT RUN. No GPU window; /dev/nvidia* was not opened.

What this PR does now

Main already holds the earlier omega b5ab3fe sync (#207). This branch merged main, took main's side for every file of the earlier b5ab3fe commits, and now carries only the sync to omega-mv 4376c2112d7469b6908c6403720c9001de4622b5 (train r3 omega-mv#979, omega#456).

Before this branch the drift gate failed on main with 4 of 782 GPU paths changed in omega 4376c2112d74 and not in CUDA++: Makefile, src/omega_gpu_decode_graph.c, src/omega_gpu_fp8_api.c, tests/gpu_fp8gemm_test.c. Carried from omega#456:

• the decode graph per-token timeline never writes past stamp[]; new host test tests/test_omega_gpu_decode_graph_timeline.c (fake platform, no device) and test-gpu-decode-graph-timeline with 2 mutants, run by test-gpu-decode-graph;
• fp8 tensor upload fills and syncs under the session lock, and omega_gpu_fp8_test_set_no_fold writes under the lock;
• gpu_fp8gemm_test emu_row uses a static buffer bounded by k;
• test-gpu-prefill-host (8 prefill tests, host model and simulator) and check-own30b-decode-profile with tools/own30b_decode_profile.c.

CUDA++ code that was ahead stays: the fp8 upload owner generation and the test_out_open 0600 receipt in gpu_fp8gemm_test. tools/extract/drift-baseline.tsv records the 4376c21 blobs and names those paths. Not carried: the require_physics_pin.sh line in omega's chip runners (no such runners here).

Drift row: d2c83a1 adds omega's blob for tests/test_omega_gpu_decode_graph_timeline.c to the baseline, so the new mirrored test is tracked by the gate.

Crumbs: crumb compile . with crumb tool d0f0b95 also rewrote om/moe/.crumb and tools/extract/.crumb, which were STALE on main; only generated fields changed.

Removed from the earlier head 063a0c2

• A stricter dagsim mutant check (the assertion name had to appear on a FAIL line). That is a CUDApp gate change without a Drake decision, so it is not in this PR.
• The omega_runtime_api.internal abort_token row: main declares that function in the manifest as EXPERIMENTAL.

Known limitations

NOT_RUN on GB10: decode graph timeline and fp8 upload locking on the chip; all claims above are host and simulator only.

Grokbot

[drakestapleton]

Review at b3f3d5e (the head is unchanged; not a draft; base main b5746b9).

Classification: needs-changes. The PR is small and correct where it mirrors. One recorded divergence hides a real gap and should be mirrored now (tests/gpu_decode_attn_test.c, see below). Three CI jobs were also still running.

CI at this head (I didn't rely on the VM numbers)

• drift (124335613116): success. The log says DRIFT-GATE: PASS (772 GPU paths, omega b5ab3fe94304, none moved away from CUDA++ since the baseline) and SELF-TEST: PASS (both controls red, real check green). CI's own omega clone resolved to b5ab3fe94304.
• Passing: runtime-api, library, audit-scanner, sim, build (124335614026), build (124335612635), om-units, sbom, hello-sim, fresh-checkout, figures, Analyze (actions), Analyze (python).
• Neutral: CodeQL.
• Still running: host-tests (124335614249), x86-host-tests (124335613853), Analyze (c-cpp) (124335609474). Until host-tests finishes, the author's HOST TESTS PASS (58 targets) is a VM receipt only.
• Fake-green: none found. No workflow file and no tools/extract/drift.sh change. The self-test still requires both controls red. All 375 data rows in drift-baseline.tsv carry omega's blob at b5ab3fe exactly (0 mismatches), so a recorded path goes red again as soon as omega changes it. Nothing is permanently allowlisted.

Exact mirrors: 18/18 byte-identical

I compared git blob SHAs with omega-mv b5ab3fe943048ac896b368a687a38fdc1dea4077. All 18 listed paths match, and so do the 7 companions: mk/omega-mufu-ftz.mk, tests/test_omega_gpu_decode_graph_abort.c and the 5 tests/dagsim/fixtures/*.oev. src/omega_cubin.c (frozen, ADR 0044) is untouched, and no backend// path is touched.

Recorded divergences: 31 of the 49, plus 2 extras

The count adds up: 18 mirrored + 31 recorded = 49. The two extras (tests/dagsim/mutant_lifetime.sh, tests/dagsim/mutant_predict.sh) are recorded at drift-baseline.tsv:345 and :347. Each recorded path has a # Divergence: reason on the line above its row, and CUDApp's blob does differ from omega's for every one.

My verdicts, by drift-baseline.tsv row:

Justified: CUDA++ build, docs, ABI or licence

• :7 Makefile
• :9 docs/runtime/API.md, :11 PORTABILITY.md, :13 omega_runtime_api.manifest, :15 portability_table.tsv (0.6.0 vs omega 0.1)
• :158 mk/om-encoder.mk, :161 mk/omega-gpu-engine.mk, :164 mk/runtime-r1.mk, :166 mk/runtime-r3.mk (SPDX only)
• :316 src/omega_runtime_api.h (0.6.0)
• :406 tests/test_omega_codegen_freestanding.c (licence header only)

Justified: not imported (persistent device omega #575 / hazard checker / FP8 slot pool), with ADR 0043 Decision 0 relevance

• Persistent device: :210 src/omega_blackwell_engine.c, :212 src/omega_blackwell_engine.h, :404 tests/test_omega_blackwell_engine.c, :351 tests/fake_m16_native.c, :353 tests/fake_m16_native.h. Not importing persistence is consistent with the Decision 0 gate. Making it permanent needs an Architecture ruling.
• Hazard checker: :226 src/omega_bw_hazard.c, :228 src/omega_bw_hazard.h, :327 tests/bw_hazard_test.c. Its fixtures quote an nvcc listing, which is a provenance concern.
• FP8 slot pool: :254 src/omega_gpu_fp8_api.c, :257 src/omega_gpu_fp8_internal.h, :260 src/omega_gpu_fp8_slots.c, :262 src/omega_gpu_fp8_slots.h, :367 tests/gpu_fp8_slots_test.c. It depends on src/omega_fp8_slot_table.c, which is outside the GPU map.

Justified but needs an Architecture ruling

• :169 src/omega_blackwell_codegen.c. Besides the header and cudapp_native.h differences, I confirmed a real encoding fork: CUDApp emits LDG.E.MMIO.GPU with w[2]=0x0c1f1900, and omega now emits LDG.E.EF with 0x0c0e1900. Keeping CUDApp's word is right without chip evidence, because both are NOT_RUN on GB10. It is an instruction-encoding decision, though, and should go to Architecture. It isn't omega_cubin.c and isn't under backend//.

Acceptable deferral (comment-only, or a test-reference style difference)

• :303 src/omega_gpu_session.h (omega's change is comments only)
• :355 tests/gpu_attention_test.c, :384 tests/gpu_pfgemm_test.c, :388 tests/gpu_prefill_attn_test.c, :391 tests/gpu_qgemm_test.c (MUFU helper routing). These are reasonable to defer now that tests/bw_mufu_ftz.h is mirrored, but they should be follow-ups.

Deferred mirror: should be mirrored now

• :361 tests/gpu_decode_attn_test.c. This PR mirrors src/omega_gpu_decode_attn_api.h, which raises MAX_POS to 131072, but keeps the test sampling % 65536u. The host has no coverage for the doubled cap. Omega's host-only extension costs nothing, so please mirror it (or add equivalent host checks up to 131072). Chip points above 65536 stay NOT_RUN.

Extras (justified): :345 and :347. They link src/sha256.c because this tree has no omega_om_units.c.

Makefile (hand-edited)

• The dagsim unit link adds -Wl,--wrap=rx_cm_predict, matching omega.
• test-dagsim runs both mutant scripts.
• test-mufu-ftz is on HOST_TESTS.
• test-gpu-decode-graph now depends on the abort test plus 5 sed mutants. Each mutant fails if its pattern isn't found, if it survives, or if it runs with zero FAIL lines. grep -c … || true only feeds that count, so it isn't fake-green.

Companion scripts

mutant_lifetime.sh and mutant_predict.sh both use set -eu and check that the mutation applied. They require a nonzero exit plus the named assertion in the output.

Nit: they only grep -q the assertion name, so a crash that happened to print that name in a PASS line would count as a kill. Matching a FAIL-prefixed line would be stricter.

API surface

omega_gpu_decode_graph_abort_token is added to docs/runtime/omega_runtime_api.internal:130 as INTERNAL. src/omega_runtime_api.h is not changed by this PR, so it has no version bump and the public surface stays 0.6.0. The runtime-api job passed.

Chip claims

I found no GB10 or hardware claim without evidence. The PR body marks IMAD.HI even-Rc pairs, MUFU flush, decode-graph abort, contexts above 65536, and the doctor result as NOT_RUN or host-only. The Makefile comment at the #492 note says chip probe NOT_RUN.

— Grokbot

[drakestapleton]

Re-review at 063a0c2. The head matches what was requested and the PR isn't a draft. Compared with my previous review head b3f3d5e, this push is one commit touching 4 files.

Classification: ready-to-merge on content, once host-tests goes green. host-tests (124339970655) was still running at 5:20 PM CT. Separately, the failing CodeQL check is the 2 expected alerts in the mirrored test (below), which still need to be dismissed. The must-fix from my last review is resolved.

1. tests/gpu_decode_attn_test.c

• Now a byte copy of omega. Its blob is ff35e2a039efa661a056ab6a289e47a9b43cc73b, the same blob the baseline recorded for omega-mv b5ab3fe.
• Recorded row removed. The # Divergence comment and the row (formerly drift-baseline.tsv:360-361) are both gone.
• The 131072 cap is now covered on the host:
  • fixed positions now include 65536, 99999 and 131071;
  • the random samples are taken % OMEGA_GPU_DATTN_MAX_POS;
  • a new check refuses OMEGA_GPU_DATTN_MAX_POS + 256 before any device work.
• Chip points 65536, 100000 and 131072 were added to the default chip mode. These are NOT_RUN: no receipt is claimed.

2. LDG word divergence in src/omega_blackwell_codegen.c (drift-baseline.tsv:168)

The comment now records the encoder word as a CUDApp edit: CUDApp replaced omega's 0x0c0e1900 with LDG.E.MMIO.GPU 0x0c1f1900 in 3c16328, after the shared import.

• 3c16328 checked: it exists in this repo and edits this file: "Honor the LDG MMIO GPU IR contract…", 2026-10-28, 1:04 PM CT.
• No NAK citation.
• No GB10 claim. The comment says neither word has a NAK citation or a GB10 test, so neither wins under ADR 0043 provenance. That's correct, and it keeps the question open for an Architecture ruling.
• The file and its baseline blob are unchanged. No frozen-stack path (omega_cubin.c, backend//) is touched.

3. Mutant scripts

• tests/dagsim/mutant_lifetime.sh:33 now matches ^ FAIL .*inverted_lifetime_masks_overflow.
• tests/dagsim/mutant_predict.sh:35 and :40 match ^ FAIL .*predict_unsupported_despite_known and ^ FAIL .*predict_refused_despite_known.
• The regex fits the real output. dagsim_unit.c's CHECK writes " FAIL %s:%d " and then the message, on one line to stderr. The messages start with those assertion names (dagsim_unit.c:458, 465, 575), and the scripts capture 2>&1.
• No fake-green. The scripts still need a nonzero exit, still use set -eu, and still check that the mutation applied. A PASS or summary line can no longer count as a kill.

4. What changed since b3f3d5e (nothing else is new)

• tests/gpu_decode_attn_test.c (+27/−14): the full omega mirror described above.
• tools/extract/drift-baseline.tsv (+1/−3): the reworded LDG comment, and the removed decode-attention divergence.
• tests/dagsim/mutant_lifetime.sh (+1/−1) and tests/dagsim/mutant_predict.sh (+2/−2): the FAIL-line anchors.

There are no Makefile, workflow, drift.sh, src/ or API changes.

CodeQL (124340161204): failure, 2 new high-severity alerts. These are the expected pair in the mirrored test, where fopen creates a file with default 0666 permissions:

• line 337, fopen(out_path, "w") in chip_run, visible in the diff;
• line 435.

They come with the byte mirror and are being dismissed as "used in tests", so I'm not counting them as must-fix. The check reports exactly 2 annotations and no others.

5. Drift baseline

19 mirrored + 30 recorded = 49, and the 2 mutant-script rows are still recorded separately. drift (124339970752) passed at 5:11 PM CT.

6. Checks at this head (CT)

• Running: host-tests 124339970655 (started 5:11 PM).
• Failure: CodeQL 124340161204, the 2 alerts above.
• Passing: x86-host-tests 124339970723 (finished 5:16 PM), drift 124339970752, build 124339971023, build 124339970583, runtime-api 124339970895, library 124339970449, sim 124339970386, om-units 124339970927, sbom 124339970943, hello-sim 124339971047, fresh-checkout 124339970759, figures 124339971011, audit-scanner 124339970836, Analyze (c-cpp) 124339959499, Analyze (python) 124339959400, Analyze (actions) 124339959221.

— Grokbot