[noelEOS]

The fusion rounding probe failed on this VM because it assumed a sequential float64 dot product, while the reference convolution uses PyTorch/MKL matrix multiplication with a CPU-dependent accumulation order. The replacement probe keeps the intended portable distinction: formulation A (rounded partials) differs from B (double partials) and from the monolithic convolution. The unchanged historical fixture also checks A and B through the real grouped code. A separate scalar calculation checks only the historical monolithic result, whose accumulation-order dependence is documented.

Diagnosis

On unchanged main at base a988d79101a401f55ea77bbf24795253b67bf5ff:

• Process-local MKL_ENABLE_INSTRUCTIONS=AVX2 returns monolithic [1, 1]; SSE4_2 returns [1, 0] for the original probe.
• Permuting sequential additions of [2^53, 1, 1, -2^53] produces {0, 1, 2}. The two unit contributions can be absorbed at 2^53.
• The original test fails with OMP/OpenBLAS thread counts 1, 2, and unset (4.41, 4.39, and 4.42 seconds). NumPy OpenBLAS does not execute this reference convolution's dot product: it is torch float64 matmul using MKL.
• VM: Python 3.12.3, torch 2.14.0+cpu, MKL 2024.2 build 20240605, NumPy 2.5.3, OpenBLAS 0.3.34.106.0. CPU: Intel Xeon at 2.20 GHz under KVM, AVX2 and FMA3 present, AVX512 absent; torch selects AVX2. threadpoolctl is unavailable.
• Hosted baseline run 41588916414 passed the same base revision. Hosted log retrieval returned empty output, so its package versions and CPU path could not be verified.
• Unverified hypothesis: the VM's preemption/restart may have changed its host CPU and SIMD path since the earlier passing local baseline. No pre-restart CPU record establishes that change; package metadata timestamps likewise do not establish a dependency change since that baseline.

Change and scope

Test layer only: test_upernet_fusion_optimized.py and its helpers.py reference helper.

The portable convolution fixture uses integer-generated dyadic products of magnitude at most 2^40, with absolute sums at most 2^41 + 2. Every double subtotal is exact in any order: monolithic and B return [1, 2], while A loses the units when its partials are rounded and returns [0, 0].

The original historical fixture is restored unchanged except for its name, grouped_probe_inputs_historical (verified by AST comparison). It checks A [0, 0] and B [1, 1] through the real MU.grouped_rounded and MU.grouped_double implementations. Each group dot has at most two nonzero terms; its result does not depend on the BLAS reduction order. Only the historical monolithic [1, 0] result uses explicit scalar sequential addition of products derived from that fixture. Its docstring explains the 2^53 cancellation caveat and points to the real grouped checks. The helpers docstring distinguishes the unchanged historical fixture from the new portable fixture. There is no library, golden graph, tolerance, skip, xfail, or shared-environment change.

Validation

Final clean revision: 08b907a748a0e4f94912e07b3e35ed3f6ee4ab28. Fix round 1 retains the portable fixture and all three assertions unchanged, restores real-code historical grouped coverage, and limits the explicit sequential test to monolithic. Main bb76ab958f7f4e0a0135edf7671a4451b45137b9 was merged without conflicts in c6c3b3b7b77abd24ddfd04e7b02072aa5451e4b1; no rebase was used. Only the two test files differ from main. Local Python commands used PYTHONPATH=. and two OMP/MKL/OpenBLAS threads; heavy checks ran serially under the shared lock.

• python tools/graph_identity.py --check docs/evidence/l00-golden-graphs.json: passed before and after; 36 fixture graphs equal, 9 trained graphs not compared because checkpoints are unavailable. No --resources used.
• python tools/graph_identity.py --check-ext docs/evidence/x00-g1ext-golden-graphs.json: passed before and after, 19/19 graphs equal. Asset-backed SegFormer raw-input drift was reported informationally; graphs and weight-shard plans match.
• python tools/block_catalogue.py --check: passed, catalogue in sync.
• python -m pytest -q tests/test_upernet_fusion_optimized.py: 9 passed in 11.47 seconds.
• Same changed file with optional SMP and timm imports blocked using temporary sitecustomize.py: 1 expected module skip in 0.14 seconds, no failure.
• New probe tests with process-local MKL_ENABLE_INSTRUCTIONS=AVX2 and SSE4_2: 3 passed, 6 deselected each (4.53 and 4.47 seconds), including the real historical grouped A/B assertions.
• python tools/live_fixtures.py --offline window/group_product/g4-w7 --output <temporary-output>: passed, zero cloud calls, largest absolute mirror-versus-PyTorch error 8.34e-7, zero values over 1e-4.
• python tools/live_fixtures.py --plan window/group_product/g4-w7: within budget, 1 request, 0 assets, 0 tasks; 14 by 14 grid, 2,744 values, 208,749 request bytes.
• git diff --check: passed.

The full suite is left to hosted CI. No live Earth Engine calls were made. This change makes an offline test portable; it adds no live numerical or model-accuracy claim. Hosted CI results are recorded below.

Hosted CI result before fix round 1

Run 41613302939 passed at 5fe7bfe07968b8c803b7c1e2860d08b4c40ca23d:

• core (Python 3.11): passed, 8m35s.
• core (Python 3.12): passed, 9m54s.
• optional-models: passed, 19m1s.

All three jobs completed successfully on the earlier head. Hosted CI for fix round 1 passed; the completed result is recorded below. This PR remains a draft pending independent re-review.

Hosted CI result for fix round 1

Run 41617177380 passed at 08b907a748a0e4f94912e07b3e35ed3f6ee4ab28:

• optional-models: passed, 21m22s.
• core (3.11): passed, 5m37s.
• core (3.12): passed, 8m42s.

All jobs completed successfully. This PR remains a draft for independent re-review and the maintainer’s decision.