On PR #1127 head b63f9de, macos-latest CI reports four matmul contract failures. This is retained CI evidence, not a local reproduction or an established GC regression.
macOS CI job, raw log lines 1525–1530 at 2026-09-12T05:13:23Z:
| Fixture |
Expected source contract |
Actual rc/output |
| SM49a default list |
rc0, [0] 1 (invalid flag) |
rc0, [1e+308] 0 |
| SM49b default buffer |
rc0, null 0 (pinned existing behavior) |
rc0, 1e+308 0 |
| SM54 strict list |
rc1, matmul: result is not a number |
rc0, [1e+308] |
| SM55 strict buffer |
rc1, same diagnostic |
rc0, 1e+308 |
The suite summary says five failures because its fifth printed FAIL is the child-exit ledger (test_strict_math.sh exited 1 without completing), not a fifth numerical fixture. The Intel macOS job reports no corresponding strict-math failures before reaching later unrelated failures. That difference suggests a platform/toolchain-sensitive investigation; it does not establish a cause such as contraction or reassociation.
Existing source fixtures in tests/test_strict_math.sh:223–259 are the bounded repro recipe. The list expression is matmul of [[[1e200, 1e200]], [[1e200], [0 - 1e200]]]; buffer fixtures use corresponding 2-element inputs. Capture actual platform/architecture/compiler version and Makefile flags, then run the existing full strict-math child with a bounded owner on clean base 6fa1cb0 and this head. Preserve exact stdout/stderr/rc; do not change the expected invalidity/default buffer result or numerical policy to make the test pass. No such new execution has yet occurred.
Read-only issue search 2026-09-12 (matmul strict, macOS math) found related closed numerical-reform issues #971/#975 and no matching platform defect; search is not an exhaustive duplicate guarantee. Track separately from #1126's shell portability/budget defects. The four above are not repaired by the pending Bash3/sed/wc harness patch.
On PR #1127 head b63f9de, macos-latest CI reports four matmul contract failures. This is retained CI evidence, not a local reproduction or an established GC regression.
macOS CI job, raw log lines 1525–1530 at 2026-09-12T05:13:23Z:
[0] 1(invalid flag)[1e+308] 0null 0(pinned existing behavior)1e+308 0matmul: result is not a number[1e+308]1e+308The suite summary says five failures because its fifth printed FAIL is the child-exit ledger (
test_strict_math.sh exited 1 without completing), not a fifth numerical fixture. The Intel macOS job reports no corresponding strict-math failures before reaching later unrelated failures. That difference suggests a platform/toolchain-sensitive investigation; it does not establish a cause such as contraction or reassociation.Existing source fixtures in tests/test_strict_math.sh:223–259 are the bounded repro recipe. The list expression is
matmul of [[[1e200, 1e200]], [[1e200], [0 - 1e200]]]; buffer fixtures use corresponding 2-element inputs. Capture actual platform/architecture/compiler version and Makefile flags, then run the existing full strict-math child with a bounded owner on clean base 6fa1cb0 and this head. Preserve exact stdout/stderr/rc; do not change the expected invalidity/default buffer result or numerical policy to make the test pass. No such new execution has yet occurred.Read-only issue search 2026-09-12 (
matmul strict,macOS math) found related closed numerical-reform issues #971/#975 and no matching platform defect; search is not an exhaustive duplicate guarantee. Track separately from #1126's shell portability/budget defects. The four above are not repaired by the pending Bash3/sed/wc harness patch.