Why does layering matter specifically for a constrained IoT device?

Two reasons: it lets you reuse a proven stack while controlling cost, and it makes the overhead tradeoffs explicit. On the reuse side, an MCU runs a lightweight TCP/IP stack like lwIP that implements the Link, Internet, and Transport layers behind a socket-like API, so you only write the Application layer (your MQTT client or CoAP handler), you get standards-compliant networking without reimplementing IP and TCP, and you can run the same application logic over Ethernet or WiFi by swapping the link driver. On the overhead side, encapsulation means every layer adds header bytes, and on a constrained link with a small MTU and tight RAM, the TCP+IP+link headers can dwarf a small sensor payload and the per-connection state (buffers, windows) costs precious memory, which is exactly why IoT favors compact, UDP-based application protocols like CoAP over heavyweight TCP+HTTP, and why embedded stacks are trimmed (fewer/smaller buffers, reduced window sizes). So the layered model both hands you a portable, maintainable stack and frames the memory/bandwidth decisions (which transport, how many connections, how big the buffers) that dominate embedded networking design.