Fix cross-process EXEC_* execution over the M162 socket transport

[Copilot]

M166's exec hook resolves tensor operands via raw host pointers valid only in the process that issued them. Forwarding those pointers to a socket-backed att1-aimu-endpoint daemon's separate address space and dereferencing them there was undefined behavior (reproducibly crashed the daemon), so test_two_tile_decode_socket only ever exercised fabric/barrier traffic, standing in precomputed cpu-f32 values for real tile compute.

Client: forward real bytes, not pointers

• src/aimu_endpoint_client.c: client_cmd_submit() copies the actual operand bytes for LOAD_TENSOR_TILE (f32), EXEC_MATMUL, EXEC_RMSNORM, EXEC_ROPE, EXEC_FFN into the wire request's existing payload field instead of forwarding input_buf_addr verbatim.
• A small pending-transfer table (keyed by command_id, since dispatch happens asynchronously relative to submit) remembers where each result belongs; client_cmd_poll_completion() copies the daemon's real result bytes back into the caller's output buffer.
• q8/q4 LOAD_TENSOR_TILE is rejected client-side with ATT1_ERR_UNSUPPORTED: those weight tensors are referenced by a struct pointer with its own nested owned buffers (and, for q4, a group_size the frozen v1.0 packet has nowhere to carry), so safe reconstruction needs wire-protocol changes beyond this fix.

Daemon: own the memory, rewrite the command

• tools/att1-aimu-endpoint.c: new dispatch_cmd_submit() copies the payload into daemon-owned buffers and rewrites cmd.input_buf_addr/output_buf_addr to point at them before submitting, so the exec hook only ever dereferences memory valid in the daemon's own process. Client-visible addresses are restored in the echoed response so no daemon pointer leaks back.
• dispatch_cmd_poll_completion() looks up the daemon-owned output buffer by command_id and returns the real result bytes in the response payload, then frees the buffers.
• Same-address/mismatched-byte-count commands and result-size mismatches are rejected explicitly (ATT1_ERR_INVALID_ARG/ATT1_ERR_IO) rather than silently truncated.

Tests

• tests/test_aimu_cluster_decode.c: test_two_tile_decode_socket now binds a real att1_backend_pcie to each socket-backed endpoint, mirroring the in-process test, and asserts the combined logits still match the cpu-f32 reference exactly — proving genuine cross-process EXEC_* execution rather than just fabric/barrier traffic.

Docs

• docs/PHASE2_PLAN.md (M167 entry) and docs/OPERATION_LOG.md updated to record the fix and note the backend-comparison report's pcie column is separate, not-yet-implemented follow-up work (no pcie wiring in att1-bench today).