fix: a handshake's return wires are the ones laid backwards

[endrix]

Why

A handshake answers backwards: ready leaves an instance on the west, beside the valid it answers, and enters the queue on the east. ELK sees a cycle for every channel and has to lay one of its wires against the flow. Its greedy cycle breaker counts wires, and a queue drives five (data, valid, size, ready, space) while it takes three, so every queue became a source and every actor a sink. Core.vhd of the chips workflow, ten queues and six actors, came out as a column of queues beside a column of actors.

What

WorkflowElkLayoutEngine now tells ELK which wires go with the flow. In a diagram that has a wire leaving a west port or entering an east port, every other wire gets elk.layered.priority.direction 10 and those wires 0, so the return wires are the ones reversed. A real feedback loop still reverses a data wire, as it must.

Measured standalone on the graphs the engine hands to ELK:

graph	before	after
Core.vhd, 30 nodes, 94 edges	2 layers, 20 data wires backwards	8 layers, 12 backwards (the CPU's feedback)
Load_ld_ld_config.vhd, 175 nodes, 189 edges	21 layers, 16 data wires backwards	20 layers, 1 backwards

Priority 2 and below changes nothing on Core.vhd; 5 and up gives the eight layers.

A diagram without return wires gets no priority at all, so wfpy and streamblocks layouts are byte for byte what they were. The new test pins both: a producer → queue → consumer handshake chain comes out in that order, and a plain graph hands ELK no priority.

Checks

• diagram-server: 72 files, 356 tests pass
• tsc --noEmit clean

https://claude.ai/code/session_01UtmwstZRcWJeg3C69s1TnS