Summary

The three a5 examples with no a2a3 counterpart to read instead.

bgemm — one source, two cores, no GM round trip

It keeps its kernel in kernels/mix/ rather than kernels/aic/ + kernels/aiv/, and that directory name is the example. The same source is registered as two incores with different core_types:

Both see the same three-tensor args[]. The matmul product moves between the halves through VEC_FIFO and never reaches GM — only the accumulator C round-trips, once per GRID_K iteration.

a2a3's benchmark_bgemm computes the same thing with separate sources and a GM round trip, and exists to sweep shapes. The two are not ports of each other, which the READMEs now say on both sides.

urma_ / sdma_deferred_completion_demo — same protocol, two transports

kernel_consumer.cpp is byte-identical between the two directories, so the transport is the only variable — the useful property when choosing between them. What differs: the transfer kernel, its completion header, and the build flag.

The load-bearing constraint is in src/a5/platform/onboard/host/CMakeLists.txt:

option(SIMPLER_ENABLE_PTO_SDMA_WORKSPACE "..." OFF)
option(SIMPLER_ENABLE_PTO_URMA_WORKSPACE "..." OFF)
if(SIMPLER_ENABLE_PTO_SDMA_WORKSPACE AND SIMPLER_ENABLE_PTO_URMA_WORKSPACE)
    message(FATAL_ERROR "... mutually exclusive because CommContext exposes a
                         single workSpace/workSpaceSize pair")
endif()

Both default OFF and one build can carry one overlay, not both. So a stock build skips both tests even on a5 silicon — a green CI run is not coverage of either path, which seemed worth stating plainly. Each README gives the two-step invocation, since the env var is read twice for different reasons: simpler_setup/runtime_builder.py:381-386 forwards it to CMake so the overlay compiles in, and the test's skipif reads it from the environment.

Index corrections

The a5 index carried the same wrong one-liner the a2a3 one did before #1571:

• paged_attention_unroll_manual_scope was "Manual scope plus loop unrolling". The kernels are rewritten — aiv_softmax_prepare.cpp changes 269 lines against a 149-line original — so it is a second implementation.
• paged_attention_manual_scope now records that its four kernels are byte-identical to the baseline's and only the orchestration differs, 52 lines of 288.

Testing

• markdownlint-cli2 --config tests/lint/.markdownlint.yaml — 0 errors on all four files
• Every relative link resolves
• Mix-kernel mechanics from kernels/mix/kernel_bgemm.cpp and the two incores entries in test_bgemm.py; "mix" confirmed as a live sentinel in deps_viewer.py, l0_swimlane.py, sched_overhead_analysis.py
• Overlay exclusivity and defaults read from src/a5/platform/onboard/host/CMakeLists.txt:47-53; env forwarding from simpler_setup/runtime_builder.py:381-386; kernel_consumer.cpp identity via diff -q
• Per-file diff counts for the index corrections reproduced with diff | grep -c '^[<>]'
• Examples not executed — no built project venv in this worktree; documentation only

Docs only; no code, test, or build files touched.