llama.cpp enables attn rotate by default

This is one of those upstream changes that matters more than the headline suggests: the feature is no longer just "merged," it is becoming the default behavior people will actually hit in production.

• –The release notes point to `kv-cache : support attention rotation for heterogeneous iSWA`, which is the concrete signal that SWA-style workloads are now part of the supported path.
• –The Reddit reply says it is enabled by default, and that there is no CLI flag to turn it on manually, which matches the "quiet default" pattern common in llama.cpp.
• –For developers running long-context or hybrid-attention models, this shifts the debugging question from "how do I enable it?" to "which model/runtime combinations still need extra tuning?"
• –The confusion in the thread is itself a signal: llama.cpp keeps adding low-level optimizations faster than its surface-level flags and docs fully catch up.