[aien-spark]

Evidence on head 8aff3ff (omega-mv final d9461717721856602626c3202751d7653e1aa5f7)

• tools/extract/drift.sh --self-test <omega at d9461717> d9461717721856602626c3202751d7653e1aa5f7: DRIFT-GATE: PASS (784 GPU paths, omega d94617177218, none moved away from CUDA++ since the baseline), SELF-TEST: PASS (both controls red, real check green).
• omega #986 synced: tools/sm121dis/check.sh decoder exit status (delta applied), tools/sm121dis/check_selftest.sh (byte copy), test-sm121dis-check-controls in mk/sm121dis.mk; baseline rows record omega's d9461717 blobs. make test-sm121dis-check-controls: PASS: decoder exit controls (serial and parallel). Negative control: the same selftest against the previous check.sh prints FAIL: nonzero decoder exit accepted or not diagnosed.
• crumb verify .: OK: 33 crumbs current.

Earlier heads:

70782f0 (aarch64 host, no GPU)

• omega-mv main then moved to 613765d2b75383212c50520ff7a7295bff01f1bb (#985), which drifted one more path. tools/extract/drift.sh --self-test <omega at 613765d> 613765d2b75383212c50520ff7a7295bff01f1bb on this head: DRIFT-GATE: PASS (783 GPU paths, omega 613765d2b753, none moved away from CUDA++ since the baseline), SELF-TEST: PASS (both controls red, real check green).

On af55527 (the 6d7b883 sync commit):

• tools/extract/drift.sh --self-test <omega-mv at 6d7b883844a0f17a7a3ceb427108be7e01e4c677> 6d7b883844a0f17a7a3ceb427108be7e01e4c677: control 1 red (Makefile), control 2 red (scratch Omega edit), DRIFT-GATE: PASS (783 GPU paths, omega 6d7b883844a0, none moved away from CUDA++ since the baseline), SELF-TEST: PASS (both controls red, real check green).
• Before this branch, on d2c83a1 (the Mirror omega b5ab3fe GPU drift and record the rest #203 head, merged to main as 769997d) against the same omega: DRIFT-GATE: FAIL (10 of 783 GPU paths changed in omega 6d7b883844a0 and not in CUDA++).
• make test-bw-hazard test-bw-hazard-nvdisasm test-sm121dis-corpus-controls: disassembler cross-check (sm121dis, in process): 2360 kernels, 1028712 instructions, 0 mismatched, bw_hazard_test: 2621 checks, 0 failed: PASS, PASS: missing and failing corpus producers are refused (includes the new partial-dump control).
• crumb verify .: OK: 33 crumbs current.
• test-sm121dis, test-zero-nvdisasm (need the nvdisasm oracle / strace setup) and GB10 runs: NOT_RUN.

Change

omega-mv main moved to 6d7b883 (#984, C6: hazard cross-check through Omega's sm_121 disassembler; four fail-open paths closed), which turned CUDApp's drift gate red on 10 GPU paths.

• tests/bw_hazard_test.c: byte copy of omega 6d7b883 (--disasm replaces --nvdisasm).
• Omega's delta applied to the CUDA++ copies of Makefile, mk/sm121dis.mk, src/omega_blackwell_codegen.c (a short kernel dump is removed and reported), src/omega_sm121_disasm.c, tools/sm121dis/{README.md,check.sh,corpus.sh,corpus_selftest.sh,zero_nvdisasm.sh}. These keep their SPDX lines and CUDA++ corpus scope; the build/test_omega_reduce_cpu prerequisite of test-zero-nvdisasm is not taken because the numeric reduce checks stay in omega-mv.
• src/omega_blackwell_codegen_tern.h: omega 613765d's comment fix applied (QMD shared cap is 32 KiB; this tree's omega_blackwell_qmd.h defines OMEGA_BW_QMD_MAX_SHARED_BYTES 32768u). Comment only.
• tools/extract/drift-baseline.tsv records omega's 613765d blob for that header, and omega's 6d7b883 blobs for those 9 paths.
• tools/extract/.crumb regenerated by crumb compile . (generated fields only).

No .github/workflows, tests/gates or other gated path is touched.

Grokbot