Summary

Fixes #1483 — tests/st/a2a3/host_build_graph/bgemm intermittently failed its golden compare with max_diff of 1e25–1e36, but only in full-suite sim runs, never in isolation.

Root cause

C is declared D.OUT in the orchestration signature, but the task graph read-modify-writes it: p_add.add_inout(c_view) feeds kernel_tile_add, whose first k iteration TLOADs C before any task has written it.

The runtime stages only IN/INOUT tensors host→device (src/a2a3/runtime/host_build_graph/host/runtime_maker.cpp — "pure write-only OUTPUT buffers are never read by the kernel"), and since #1365 (affb4612) it no longer zero-fills OUT buffers either. So C accumulated onto whatever the device allocation happened to hold. On sim that allocation is a plain malloc block on a process-lifetime worker (memory_allocator.cpp): fresh zeros when the case runs alone, a previous case's bytes under the full suite. Same class as #1449.

Both sibling bgemm variants already tag their accumulator correctly — examples/a5/tensormap_and_ringbuffer/bgemm and examples/a2a3/tensormap_and_ringbuffer/benchmark_bgemm use D.INOUT. The hbg ST was the only one mislabelling it.

Fix

• signature: [D.IN, D.IN, D.OUT] → [D.IN, D.IN, D.INOUT].
• Seed C with a non-zero base (0.25) and drop C[:] = 0.0 from compute_golden, so golden is C_init + A @ B. This makes the staging load-bearing: a zeroed or unstaged device buffer can no longer match, so a future regression fails deterministically instead of hiding behind a happens-to-be-zero allocation.
• Drop the pytest.mark.skip that quarantined the case — the defect it tracked is fixed, so the case runs in CI again.
• Orchestration header comment and README updated — the comment previously asserted "C arrives zero-initialized (pure OUTPUT), so the accumulation is exact", which is no longer true.

Verification

• Failing repro first: with the non-zero base applied but the signature still D.OUT, the case fails on the first attempt with max_diff=8.0e+35 — the reported signature, now deterministic rather than one run in two or three.
• With the fix: passes in isolation on a2a3sim and onboard on a2a3 under task-submit.
• pytest examples tests/st --platform a2a3sim --device 0-15 --pto-session-timeout 600 --require-pto-isa — green on 5/5 consecutive runs, 3 before and 2 after the rebase onto 23de96e0 (issue reported ~1-in-2-to-3 failures).
• pre-commit clean on all three changed files.

Scope

I surveyed every other add_inout orchestration under tests/st/a2a3/host_build_graph/: available_aicore_counts and predicated_dispatch already declare their tensors INOUT, and dump_args / matmul / vector_example cover their OUT tensor with an add_output before any read. bgemm was the only offender, so no other test is touched.

🤖 Generated with Claude Code