[chaoming0625]

CELL-20. event_backend="brainevent" had been unusable on every brainevent release since 0.1.0 because it called brainevent.coomv, which 0.1.0 removed. The tests that cover this path skipped when coomv was missing, so CI never ran it.

Decision

brainevent has a replacement with the same meaning: brainevent.csrmv(data, indices, indptr, v, *, shape, transpose, backend). Like coomv, it is a float sparse matrix-vector product, not the binary-gated variant. It has the same signature and the same jax_raw kernel (registered for cpu, gpu and tpu) in every release from 0.0.7 (the current floor) through 0.4.1. The brainevent path now uses it, and the brainevent>=0.0.7 floor stays as it is.

• Each delay-grouped block is stored as a CSR matrix with one row per presynaptic member. Contacts are sorted stably by pre_index, and repeated destinations are kept. Weights are permuted when the op is called, so trainable weights still work. A weight with one element is passed as homogeneous.
• The exact bilinear custom JVP built on scatter is unchanged.
• "auto" now always resolves to "scatter". On brainevent >= 0.1 it already did. With the default brainevent_backend="jax_raw", the brainevent operator lowers to the same gather and scatter-add, so preferring it would change the default path for every user with no benefit.
• If the installed brainevent has no csrmv, "brainevent" raises a RuntimeError during run/prepare_run setup. The message names event_backend='scatter'. make_delivery_op no longer falls back to scatter silently.

Changes

• braincell/network/delivery.py: resolve_event_backend and make_delivery_op as described above.
• delivery_test.py: the brainevent derivative test no longer skips. It spies on brainevent.csrmv to check that the path is really taken, and it uses unsorted pre_index, repeated destinations and a presynaptic member with no contacts. It compares primal, forward-mode, vmap-batched JVP, reverse-mode and bool-event results with scatter (rtol=1e-6). Backend-selection tests cover "auto"/"scatter"/"brainevent", and a brainevent without csrmv must raise an error that mentions scatter.
• engine_test.py: the brainevent gradient test no longer skips. Two tests are new. The first runs scatter, brainevent and auto end to end; brainevent must match scatter with rtol=1e-6, and auto must equal scatter exactly. The second checks that prepare_run(event_backend="brainevent") is rejected with a "scatter" message when csrmv is missing.
• CI.yml: the slow_tests deselect of bidirectional_test.py::...test_all_parameter_gradients_across_losses_delays_and_backends and its comment are removed.
• changelog.md: a Bug Fixes entry.

Tests

• Before the fix, the new and unskipped tests fail: 5 tests, 9 failures counting subtests. After the fix, pytest braincell/network/delivery_test.py braincell/network/engine_test.py gives 21 passed, 4 subtests passed. The same result holds with JAX_ENABLE_X64=1, and with brainevent 0.0.7, 0.1.0, 0.1.2, 0.2.1, 0.3.0, 0.4.0 and 0.4.1.
• pytest "validation/optim/gradient_correctness/bidirectional_test.py::BidirectionalTest::test_all_parameter_gradients_across_losses_delays_and_backends" gives 1 passed, 4 subtests passed. pytest validation/optim -m slow gives 17 passed.
• pytest braincell/ gives 2739 passed and 4 failed. The 4 failures are vis topology tests that need networkx, which is not installed in the local env. They are unrelated to this change.
• pre-commit run --all-files passes.

Edge cases

• Unsorted presynaptic indices, duplicate (pre, target) destinations and presynaptic rows with no contacts are covered by the delivery test.
• Homogeneous weights with one element are covered by the delivery test.
• Unitless and unit-carrying weights are covered by the delivery test.
• Bool events are covered by the delivery test.
• float64 was run locally with JAX_ENABLE_X64=1.
• A missing csrmv is covered by the delivery and engine tests.
• A second commit covers the missing-brainevent ImportError branch, which the changed-lines coverage gate had flagged (test_event_backend_selection_without_brainevent).
• A manual Daily CI dispatch on this branch passed, slow_tests included (run 42541898789), so the bidirectional test now passes in CI without the deselect.
• The branch was rebased onto main after fix(cell): convert integer-valued V_init and V_th to the default float dtype #200, QUAD-6: install braintools from PyPI again now that 0.5.0 is released #202 and MECH-16: assert NEURON voltage agreement for the channel_no_conc configurations the sweep singled out #203 and keeps main's install steps. ci-ok passes on 187f2f4.